Nathaniel Whittemore
👤 PersonPodcast Appearances
Today on the AI Daily Brief, a very launch-rich episode.
In the main episode, we discuss ChatGPT's new browser, Atlas, while in the headlines, we talk about Google's new AI Studio with one-click AI integration.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
Send us a note at sponsors at AIdailybrief.ai to learn about the opportunities.
Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes.
It is a trope at this point for when some new announcement happens, basically anything from Google or OpenAI or Anthropic, and frankly now moving down a farther and farther line of companies, inevitably some hyped up thread will say, this is a game changer, this changes everything.
And honestly, when I saw that Google was announcing a new vibe coding experience, my default is to assume that it would be cool, competent, good, useful, valuable because it's integrated with the Google ecosystem, but ultimately not all that different.
Imagine my surprise when there was a feature that stands out as actually so clearly game changing that it deserves that ridiculously overused moniker.
On Tuesday, Google AI Studio product lead Logan Kilpatrick tweeted, Introducing the new AI-first vibe coding experience in Google AI Studio, built to take you from prompt to production with Gemini and optimized for AI app creation.
Start building AI apps for free.
Okay, so one part of this that's cool is that prompt to production part.
This is the big trend that we've seen in vibe coding apps, moving you away from just prototypes to things you can fully deploy.
It's why lovable cloud has been so valuable in their ability to integrate the full end-to-end experience.
That's been transformative just in terms of how far you could get with something.
But what makes this unique is that second part that he said, the idea that this is optimized for building AI apps.
When you go into the vibe coding tool in Google AI Studio, you have your standard describe your idea, see it come to life kind of prompt input.
But then you also have this ability to quote unquote supercharge your apps with AI.
So with a single click, you can add photo editing with nano banana.
You can add conversational voice agents to the app.
You can animate images with VO.
You can integrate Google search data, Google maps data.
You can add a chat bot to your app.
I had been messing around with a prototype for an enterprise ROI tracker.
And the difference that this made is that I was able to add voice agent integration even into this prototype so that in addition to just regular surveying, I could add a voice experience where the agent asked for further information.
This is something that we have spent months building into super intelligent.
And while of course the off the shelf voice agent that just plugs into your vibe app isn't anywhere near the tuned tweaked thing that we've built for these sort of enterprise grade, large scale discovery process we do with super intelligent.
The fact that that's a single click now as part of a vibe coding experience is a total actual, I'm sorry to keep using the word, but game changer.
Now, I had to mention this is the lead in the headlines because it just came out, but I'm planning on some content that goes much deeper on this, maybe provide some examples, maybe does even a live demo of how I'm using this either this week or next.
So I'll cut it there, but suffice it to say, go check out Google AI Studio, add a chatbot to your website.
Go nuts.
Of course, the vibe coding competitors are not sitting on their laurels, and Lovable has just announced what is an extremely hyper-useful integration as well.
Lovable users will now be able to build an entire online store thanks to a partnership with Shopify.
The new process of creating an online store is now as simple as a vibe code prompt.
The example they give, create a Shopify store for a minimalist coffee brand selling beans and brewing products.
Lovable, they say, then instantly designs and builds the storefront, complete with product pages, checkout, and navigation.
However, because it's lovable, you also have much more fine-tuned controls than you would if you were just, for example, using Shopify's templates.
Now, what's interesting about this is that people are a little anesthetized to announcements at this point, and yet people instantly saw that this one was a big deal.
Sumia wrote, This is a proper use case for the mask, not some AI slop pseudo-coding time-waste stuff.
This clearly says lovable is going to remain mainstream for many more months.
Aditya writes, Wait, this is actually huge.
Imagine just how low the bar has become to start an online store, basically non-existent.
And while one could argue that just using Shopify already with their templates was good enough for most people, as someone who has built websites with both the template builders and WordPress and all these sort of things, and now maintains my website with Lovable, it is an incomparable experience in terms of how much control, how much creativity you can exert.
These things are not shifts in scale, they are shifts in kind.
And I actually think that this integration is going to be a huge deal.
Next up, one that has gotten a lot of chatter on Twitter.
A new investigative piece from the Wall Street Journal took readers behind the scene of OpenAI's multi-hundred billion dollar deals.
The article discusses how deals with SoftBank, Oracle, AMD, and Broadcom came together over the past year.
Of particular interest was a discussion of the NVIDIA deal.
The article claimed that Jensen Huang, Nvidia's CEO, of course, was jealous of the White House reveal of Project Stargate in January.
Basically, he wanted to be the one, standing alongside Altman as the president announced half a trillion dollars worth of AI investment.
Nvidia later pitched OpenAI on a similar project, effectively looking to sideline SoftBank and help to raise the required funds themselves.
However, by the summer, progress had stalled and the two companies had put negotiations on hold.
Then in June, the Information reported that OpenAI had begun renting Google's TPU chips to supplement their compute.
That report apparently caused a ton of stir inside NVIDIA HQ, and Huang quickly called Altman to get negotiations back on track.
Now, OpenAI never did consummate a deal to use Google's chips, and the result of this convoluted process was the $100 billion strategic partnership announced in September.
Under the agreement, NVIDIA will lease up to 5 million chips to OpenAI worth some $350 billion, and NVIDIA also has the right to invest up to $100 billion in OpenAI in order to help them pay for the deal.
New from this reporting is that NVIDIA is also discussing a guarantee of OpenAI's debt financing for the new data center bills, basically meaning that NVIDIA's free cash flow is also backstopping things.
Now, Sam Altman already has a reputation as perhaps the most gifted dealmaker of his generation.
And this only reinforced that point.
Amit is investing summed it up.
Remember when the information reported that OpenAI was thinking of using Google's TPUs?
A few days later, NVIDIA's X account posts a screenshot of Reuters article denying that OpenAI was going to be using Google.
All of this seemed calculated from Sam to get Jensen to the table to make that $100 billion investment in them and further intertwine OpenAI's success to NVIDIA's success.
After the deal was made, Sam went on to get deals with Broadcom and AMD, really making sure he could diversify as much as possible while bringing more companies and their progress into OpenAI's trajectory.
The punchline being, Altman is ruthlessly trying to make sure that if AI ends up being the transformational technology we all think it can become, he is going to be at the center of it.
Essentially, he's trying to make OpenAI too big to fail by making sure that every other important AI company fails if OpenAI does.
absolutely crazy time to witness this happening in front of us.
And staying on the theme of these massive deals, Anthropic is in talks to sign a multi-billion dollar cloud deal with Google.
Bloomberg sources said the deal was still in its early stages, but would be valued in the high tens of billions of dollars.
Now, while some are tempted to think that maybe this represents some sort of wobble in the relationship between Anthropic and AWS, to me, it's pretty clear that the name of the game when it comes to AI compute is polyamory.
And no one is being mad at anyone else for doing deals with everyone and everywhere they can get compute.
Still, the market seems to think that it's bad news for Amazon with their stock falling 2% after the news broke and Alphabet's stock up a point and a half.
Wild times out there in AI land, but we got a big product announcement from OpenAI as well that we have to get into.
So with that, let's close the headlines and move on to the main episode.
Small, nimble teams beat bloated consulting every time.
Robots & Pencils partners with organizations on intelligent, cloud-native systems powered by AI.
They cover human needs, design AI solutions, and cut through complexity to deliver meaningful impact without the layers of bureaucracy.
As an AWS-certified partner, Robots & Pencils combines the reach of a large firm with the focus of a trusted partner.
With teams across the U.S., Canada, Europe, and Latin America, clients gain local expertise and global scale.
As AI evolves, they ensure you keep peace with change.
And that means faster results, measurable outcomes, and a partnership built to last.
The right partner makes progress inevitable.
Partner with Robots and Pencils at robotsandpencils.com slash AI Daily Brief.
What if AI wasn't just a buzzword, but a business imperative?
On You Can With AI, we take you inside the boardrooms and strategy sessions of the world's most forward-thinking enterprises.
Hosted by me, Nathaniel Whittemore, and powered by KPMG, this seven-part series delivers real-world insights from leaders who are scaling AI with purpose.
from aligning culture and leadership to building trust, data readiness, and deploying AI agents.
Whether you're a C-suite executive, strategist, or innovator, this podcast is your front row seat to the future of enterprise AI.
So go check it out at www.kpmg.us slash AI podcasts, or search You Can With AI on Spotify, Apple Podcasts, or wherever you get your podcasts.
Today's episode is brought to you by my company, Super Intelligent.
You've got a hundred what-if ideas, but which one becomes an agent?
Super Intelligent maps every AI use case across your company and helps you create an agent plan that you can actually execute.
We match opportunities to your tech stack, your data profile, and your team.
No more guesswork, just a clear path from pilot to production.
If you want agents that deliver business outcomes, start with planning.
Go to besuper.ai and sign up for a demo.
This episode is brought to you by Blitze, the enterprise autonomous software development platform with infinite code context.
Blitze uses thousands of specialized AI agents that think for hours to understand enterprise scale code bases with millions of lines of code.
Enterprise engineering leaders start every development sprint with the Blitze platform bringing in their development requirements.
The Blitze platform provides a plan then generates and pre-compiles code for each task.
Blitzy delivers 80% plus of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint.
Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org.
Blitzy is providing a limited time, 30-day free proof of concept for qualifying enterprises.
The team will provide a 5x velocity increase on a real development project in your org.
Visit Blitzy.com and press book demo to learn how Blitzy transforms your STLC from AI-assisted to AI-native.
That's B-L-I-T-Z-Y dot com.
Today we're going to talk about the announcement, the first reactions both good and bad, and my humble take on where currently to think about how to get value out of AI browsers.
Let's talk first about the announcement.
At Tuesday's launch, Sam Altman presented OpenAI's new browser as a fundamental revolution for computing.
Here's how he put it.
We think AI represents a rare, once-in-a-decade opportunity to rethink what a browser can be.
In the same way that for the previous way people used the internet, the URL bar, and the search box were a great analog, what we're starting to see is that the chat experience and the web browser can be a quick analog.
which is a little bit mealy-mouthed because it came out of a live presentation, but basically he's saying new types of behavior inspire new types of interfaces, and an AI-native browser is going to look different in some way than a pre-AI browser.
The announcement post struck a similar tone.
In it, they wrote, AI gives us a rare moment to rethink what it means to use the web.
Last year, we added search and chat GPT so you could instantly find timely information from across the internet, and it quickly became one of our most used features.
But your browser is where all of your work, tools, and context come together.
Editors note, big surprise, that context word is going to be the important one here.
Continuing from their blog post, they write, a browser built with ChatGPT takes us closer to a true super assistant that understands your world and helps you achieve your goals.
Now, when it comes to features, Atlas largely matches competitors so far.
Maybe the key aspect of the product is having a ChatGPT assistant embedded in a sidebar.
being able to draw context from what's going on in the browser window to answer questions as well as carry out actions in agent mode.
And it is these two pieces which I think ultimately will be key to the AI browser experience, understanding the context of what you're actually interacting with on the web on the one hand and taking actions on your behalf on the other via agents.
With agent mode, they write, in Atlas, you can now ask ChatGPT to take action and do things for you right in your browser.
This is basically a native home for the ChatGPT agent that was introduced earlier this year.
As always with agents right now, they gave a food-related example.
Imagine you're planning a dinner party, they write, and you have a recipe in mind.
You can give the recipe to ChatGPT and ask it to find a grocery store, add all the ingredients to a cart, and order them to your house.
The work example they give is ChatGPT opening and reading through past team documents, performing new competitive research, and compiling all those insights into a team brief.
One of the areas where they're trying to stand out is around memory.
Atlas has the same memory design as normal ChatGPT, so can learn your preferences and recall previous chat sessions.
However, this gets turbocharged by giving Atlas the ability to draw from the browser history as another source of memory.
OpenAI said you'll be able to ask Atlas things like, find all the job postings I was looking at last week and create a summary of industry trends so I can prepare for interviews.
Now, of course, none of these features are completely novel.
We already have Perplexity's Comet browser, which I've been using for the last couple of months.
And there's also the browser company's Dia.
But OpenAI entering with a ChachiBT native version of the experience obviously ratchets up the competition to another level.
The first thing that people discussed was what this means for OpenAI.
Hater at Slow Developer on Twitter writes, "'OpenAI is clearly going for a full consumer strategy if you realize how big the ChatGPT Atlas browser update is.
Unlike perplexity, OpenAI can train its models to work natively with the browser.'"
Because it controls the full stack, it likely delivers stronger agent capabilities than wrappers.
Hold aside any claims that its performance is going to be better in the short term.
I think what Hater is getting at is that OpenAI, with the run-of-the-way most popular consumer chatbot, has the ability to more closely integrate that into this experience, which could be a differentiator when it comes to these new AI browser wars.
Noah Epstein thinks that it's all about targeting Google search dominance.
He wrote, Over 50% of Alphabet's $237 billion in annual revenue comes from search advertising.
Chrome to Google search to behavioral data to targeted ads equals their entire empire.
Atlas threatens every single link in the chain.
He then goes on to point out how all of OpenAI's moves start to add up to something that starts to not only command more and more of people's attention, but is able to collect context from them and then turn that context into both advertising, should that be the route they go choose, as well as commerce.
They point out the recent checkout features as significant with this.
Another set of first response takes was that some of the people who tried this really just liked the experience.
iRuleTheWorld, which is an OpenAI leaker account, says using the new OpenAI browser, it's actually insane how smooth it is.
Feels like the future of the internet.
Pat Walls from Starter Story writes, OpenAI is so good at product.
ChatGPT Atlas is amazing.
Immediately switched from Chrome and I've used that for 10 years.
Everything they create is so, so good.
This is why they'll win.
But holding aside just fawning tweets about the quality of the product, others tried to figure out what it was actually useful for.
Liam Ballin gave his first impressions.
He found the new agent mode to work pretty well for him around things like ordering coffee or filling out a TSA pre-check form.
He also pointed out that because it's native to macOS, it's integrated with things like auto-filling iMessage two-factor authentication codes, and he really liked the UI.
He did find that there was some variance in certain sites being blocked when it comes to summarizing content, creating kind of a balkanized experience when it came to the news, but overall favorable first impressions.
Jackie Chow at Indexy said that he's currently using Atlas as a CRO, asking ChatGPT for landing page changes, using it for meta ads, asking how to optimize campaigns, using it to improve YouTube thumbnails, drafting cold email responses, and getting other suggestions for UI UX improvements.
Reza Martin, who helped create NotebookLM and is now working on a new startup, did a head-to-head test between OpenAI Atlas and Perplexity Common on a very specific real-world use case.
She writes, "...I have a very real, very tedious use case, which is a manual task that I do every day.
One, I go to the school website to look at each of my daughter's classes.
Two, I look at her grades.
Three, I look at her assignments, quizzes, and all classwork that are due.
Four, I make a table that keeps track of all of this, which helps keep both of us accountable."
Since there's so much manual clicking, scrolling, reading, and data, she asked both Common and Atlas to help with that, and then gave them a score on context, i.e.
how well they understood the task, speed, how long it took, and then how well it did completing it correctly.
Atlas pretty much smoked Common in that test, especially the overall completion, where Common got just a 1, but the Atlas agent got a 5.
She concluded, overall, I still don't have a ton of agentic web browsing use cases, but it's great to be able to automate this one particular task.
And when it comes to finding use cases, I think that some people are viewing this as the opening salvo for an environment in which people will discover new use cases rather than there being preset use cases just poured over.
Ada McLaughlin, who is a research scientist at OpenAI, writes, "...my quick two cents on the browser.
I didn't use Codex that much when it was cloud only, but when it came to my CLI, it became super useful."
I didn't use agent that much when it was cloud only, but now that it's come to my browser, dot, dot, dot.
With the implication being that having the integrated browser in that context opens up opportunities to use it.
Some folks, of course, were zooming out ahead to the possibilities.
Greg Eisenberg writes, my takeaway from today's open AI browser launch is that the internet just got hands.
The average person won't Google, click, compare, or fill out forms within the next 24 months.
They'll just say book my trip, find me a job, launch my store, and the agent will do 20 steps behind the scenes.
That means whole industries, travel, e-commerce, real estate, insurance, education, are about to get rebuilt around outcomes instead of pages.
If you're a founder, this is the moment to think in verbs.
You won't go to Expedia, you'll just get the trip.
The web is shifting from human browsing to agent doing.
Now, this is the bet that a lot of people are making is the trajectory of the internet.
Less human browsing, more agent browsing, and things are slowly or quickly, depending on your perspective, being redesigned around agents doing things on people's behalf.
And yet, not everyone had so rosy a take on the browser.
Ben Heilack initially tweeted, I'm not sure it'll be successful, but the level of polish on ChatGPT Atlas is just way higher than anything they've shipped in recent memory.
Which he followed up a few hours later with, I take this back.
Everything just feels really janky, lots of little missing features.
I can't get used to the address bar search box thing, the ChatGPT sidebar being the thing that opens when I click the sidebar is just insane to me.
It's a confused app.
Now maybe a more discreet line of critique had to do with privacy.
Tiffany Fong wrote, If you liked OpenAI downloading the internet, you'll love OpenAI downloading your personal data.
Aidan Bai had the same thought, saying, Pretty sus way to collect massive amounts of data to train computer use.
Definitely not spyware, trust me bro.
Now, it should be noted that hold aside any sort of tinfoil hat conspiracy theories.
One of the next frontiers for the model companies is absolutely collecting data that comes from real human usage patterns.
It's why companies like Cursor are valuable, not just because of how many people are paying to use them, but because of the data exhaust that comes from that usage, which is unique and totally discreet from the existing training sets that are available to companies like OpenAI and Anthropic.
Meaning that while there may be wildly different interpretations of what is and isn't reasonable to collect and what sort of guardrail should be around that, it is certainly the case that at least part of the strategy here is to have access to this very discrete level of consumer behavioral data.
For some, they think that's just going to be too insurmountable a challenge.
Simon Willison tweeted, Wrote up my first impressions of ChatGPT Atlas, OpenAI's new browser.
I remain unconvinced by the entire category of browser agents.
The security and privacy challenges still feel insurmountable to me.
And for him, it's less about OpenAI training on his data and more about the wide world of security concerns that come out of this.
For example, he writes, I'd like to see a deep explanation of the steps Atlas takes to avoid prompt injection attacks.
Right now, it looks like the main defense is expecting the user to carefully watch what agent mode is doing at all times.
Okay, so you've got some questions around the experience, you've got concerns around privacy and security, and then there are also censorship concerns.
Jason Botterill gave it a test saying, asking the browser to look up videos of Hitler, to which the Atlas browser said, I can't browser display videos of Hitler, since footage of him and Nazi propaganda are tightly restricted for ethical and legal reasons.
However, if what you're after is historical context, I can point you towards legitimate archives and documentaries that use such footage responsibly.
It also refused to translate a Hitler video on YouTube.
And for some, this is just a level of nanny state sort of behavior that they're never going to be comfortable with.
I think that it brings up the concern in general of how much power the chatbot companies are going to have to shape reality based on what they do and don't allow people to have access to.
Now, Hamza did point out that this censorship is a choice, and it's one that while OpenAI is making, Perplexity isn't.
They put in the same prompt into Comet and found a set of videos from Getty Images, Shutterstock, and YouTube.
Hardly radical sources.
And yet for all of this, I think the most damning critique so far is just the underwhelmingness of the agent.
Back to that same post from Simon Willison, he wrote, I also find these products pretty unexciting to use.
I tried out agent mode and it was like watching a first-time computer user painstakingly learn to use a mouse for the first time.
I have yet to find my own use cases for when this kind of interaction feels useful to me, though he says I'm not ruling that out.
And this I feel like was a lot of people's experience.
Yuchen Jin wrote, tried the OpenAI browser for 20 minutes, quit and went back to Chrome.
Agent mode is slop.
Most of the time I just want to yell, stop thinking and click that effing button.
The models are not there yet.
We've got a whole decade of AI agents ahead.
Of course, referring to the Andre Carpathy dustup that we've covered extensively.
Yuchen also pointed out, it's also scary to give an incompetent agent access to all my passwords and data.
What if it goes crazy?
If the OpenAI browser agent leaks my bank and Robinhood passwords, causing me to lose all my money, who's responsible?
John Rush writes, Nothing is more disappointing than the browser agents.
They work so slow and get stuck in infinite loops.
I think it'd be smarter for AI to watch the network inspector learn the APIs, reverse engineer them, and make direct calls.
E.g.
I'd use the website while the agent is watching the API calls to reproduce them again later.
And to be clear, these are not AI critics that are having this negative experience with this agent.
The other big question is how does this all come together?
Chubby writes, the question that arises for me is what is OpenAI's next goal?
We have ChatGPT, arguably OpenAI's most important asset with 800 million weekly users and growing.
Codex, a coding agent specifically for software engineering.
With ChatGPT Atlas, we now have a browser that combines many functions such as ChatGPT Agent as computer use agent.
And next year, the standalone device together with Johnny Ive.
But where is OpenAI headed?
What is the next big goal?
And I'm not talking about AGI.
Rather, where do the threads come together?
Is ChatGPT supposed to become the all-in-one app of the future, replacing smartphones with devices, searching the internet with browsers, and then perhaps even its own operating system?
Is that what it's all leading up to?
Currently, there are numerous useful applications running alongside each other.
I wonder where they will ultimately converge.
So in my estimation and my experience so far, both using Common for the last couple of months and playing around with Atlas over the last day, is that at core, there are two big features of the AI browser experience.
The first is agents.
And frankly, in my estimation, they are just not there yet.
I think that the juice isn't worth the squeeze.
I think the things that they can do are not sufficiently difficult to justify all of the new complications they introduce.
I think that there may be discrete, very person-to-person specific use cases like raises that we heard about.
And to the extent that there are ways to incrementally grab 15 minutes or 20 minutes of time back a day or even a week, I'm all for it.
I also think that there's no way to learn what generalist agents are useful for without people trying, so I'm encouraging of it.
But for me, I just don't see it replacing any meaningful behavior in any short order.
I've also, as I said, think I'm going to be pretty far back on the adoption curve when it comes to having agents do things like shopping or ordering food or plane tickets or whatever for me.
So take that with that grain of salt.
Still, I think that agentic use is one of the two big potential long-term value propositions of these browsers, and the one that is ultimately the most different in some ways from your existing browsing experience.
The second, frankly, is less about browsing and more just a better way to use your LLM.
And that might sound dismissive, but frankly, these LLMs are so powerful.
that a better way to use them could be enough of a reason to switch browsers.
And of course, what matters here is that the native integration into the browser means that ChatGPT has all sorts of context without you having to port over the context.
As Swix put it, this is the single biggest step up for OpenAI in collecting your full context and giving fully personalizable AGI.
Context is the limiting factor, and as Marc Andreessen said, the browser is the new operating system.
The only move bigger than this for collecting context is shipping consumer hardware.
So what does that actually mean?
Let's take a use case of creating social media content, in this case, trying to write a tweet that has the potential to go viral.
The way that you would do this before is, of course, you would go to ChatGPT, you would copy-paste in the thing that you were thinking about, and then when you found a version that you liked, you would bring it back to Twitter or whatever social network you're using, copy it in there, and that's the experience.
Now, I don't mean to pretend that this is some wildly burdensome process, but here's the version of it integrated into this AI browser.
So here we have X slash Twitter pulled up now.
I've drafted my tweet.
Computer use and agents are great and all that, but my definition of AGI is what percentage of banger tweets were written by ChatGPT.
I pull up the ChatGPT browser and I say, make this tweet better.
I don't have to say anything else because it knows what I'm referring to because of the context from the other side of the browser window.
The integrated ChatGPT sees what I'm doing and gives me a set of inline suggestions for how to improve it.
Now, again, I am not pretending that it's so wildly challenging to draft something in ChatGPT and then moving it over, but context relevance without context switching is actually a valuable reduction in your cognitive load and is going to be not only time-saving, but also mental process-saving in a world where every incremental mental process is valuable.
Let's take an example, however, where the context switching would be a lot more difficult.
Here we are in YouTube Studio, where I've got a bunch of recent videos.
I could say, how should I be thinking of thumbnails?
Now this context would be a lot harder.
I would have to port in all of these thumbnails alongside the associated data, whereas with the integrated browser, it can pull on the context that is right there.
So for example, when it says number two, every thumbnail should communicate one emotion and one idea, it uses examples from my actual thumbnails.
It can also then go deeper in the analysis.
It asks if you want, I can make a thumbnail strategy matrix for your channel.
Now, like I said, this is a whole different kettle of fish.
Porting in this context to the normal ChatGPT window would be an enormously difficult and time-consuming process.
And this is where I think that right now, in the moment, the Atlas browser can be super valuable, especially if you are already using ChatGPT.
And for me, candidly,
Although I don't anticipate shifting all of my behavior over into Atlas, this set of context-relevant use cases for ChatGPT is probably enough to have me spend some amount of time there.
Now this brings up one of the other interesting points, which is how this is going to compete against Google.
As Ryan Carson says, you gotta assume Chrome will also relaunch as a fully agentic browser soon.
And yet he says, I think I'll probably switch to Atlas because I already use ChatGPT for all my personal stuff.
The most important moat in AI is your personal context.
And that's why there's so much emphasis on all these places where it can get more of it.
Now, there are a lot of other interesting thoughts that this is bringing up.
Behance founder Scott Belsky writes, One thing I believe will be very clear in retrospect as browsers evolve, we will have a consumer personal browser optimized with our personal memory, our credit card, our social graph, buying history, and the many agents and apps of daily life.
And we will have a work browser optimized for teamwork via graph of who we work with and permissioning, working across apps, tapping enterprise memories, agents as colleagues, etc.
And browser will probably become an antiquated term as this interface becomes the OS itself.
And so what that means is that while I think you can safely dismiss all of the AI hype boy threads that say this changes everything, at least right now, it is certainly worth some time to go experiment and play around with this, if for no other reason than to get a glimpse of the direction that we're headed.
For now, that's going to do it for today's AI Daily Brief.
Appreciate you listening or watching as always.
And until next time, peace.
Today on the AI Daily Brief, a new definition of AGI that suggests that GPT-5 is 58% of the way there.
Before that in the headlines, Claude Code comes to the web.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, Super Intelligent, KPMG, and Robots and Pencils.
To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can sign up on Apple Podcasts.
One note about Apple Podcasts, because this has been coming up a bit.
Apple Podcasts is set up on a very different system to Patreon.
With Patreon, I upload a distinct file.
I can schedule it the same way as I can with my normal episodes.
And so it's always set to come out at the same time as the main ad version.
With Apple Podcasts, it's a little bit different.
I have to wait for the episode to post to Apple Podcasts, meaning that there's a short delay after I publish in general, and then I have to go in and manually replace the file.
What this means is that if for any reason I can't immediately replace that file, you will still see the normal ad version on your feed.
It only gets replaced when I add that manual file.
This is normally fine.
It just means about 15 minutes of waiting around after I press publish on the normal episode, but that lag creates more possibilities for problems.
For example, yesterday, Apple's podcast connect system was down for about 12 hours, which meant that pretty much overnight, even subscribers on Apple still saw only the ad version.
I promise you that I will always try to get the ad-free version up on Apple as fast as I can, but sometimes it's going to be out of my control.
I apologize.
I really wish it was a better system, but that is just the way that it is.
Lastly, of course, for any information about the show, sponsorship, speaking, job opportunities, go to ai-dailybrief.ai.
And with that all out of the way, let's dive in.
Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes.
First up today, Anthropic is making Claude code available through a web app and within the Claude iOS app.
Previously, the feature was only available through terminals and IDEs, and the big unlock is being able to spin up background agents.
With Cloud Code running in the cloud, you can now run multiple tasks in parallel across different repositories from a single interface and ship faster with automatic PR creation and clear change summaries.
This asynchronous workflow is quickly becoming a powerful tool for AI-enhanced coders.
Cloud Code product manager Kat Wu said, As we look forward, one of our key focuses is making sure the command-line interface product is the most intelligent and customizable way for you to use coding agents.
But we're continuing to put Cloud Code everywhere, helping it meet developers where they are.
Web and mobile is a big step in this direction.
Certainly there is a lot of excitement about this.
Josh JDJ Kelly on Twitter wrote, I can work with Claude Code while out on a walk.
Speaking of agentic coding, Replit is projecting massive growth to reach a billion dollars in revenue by the end of next year.
Speaking with Business Insider, CEO Amjad Massad said that the AI coding startup has reached 240 million in ARR and expects that to quadruple next year.
The company's growth this year has been absolutely skyrocketing, gaining more than 10x from their 16 million in ARR at the end of 2024.
The company now has over 150,000 paying customers and over 40 million free users.
And while at this stage, all those free users mean that the consumer segment is unprofitable, Massad boasted that enterprise margins are close to 80%.
This follows the same profit model that other AI companies are currently pursuing.
The consumer segment is a loss leader due to large volumes of free users, but building familiarity with consumers means they demand access to the same tools at work.
Massad said that the surging revenue was largely due to adoption in mid-sized companies including Duolingo and Zillow.
He said, Replit is kind of replacing a lot of the no-code, low-code tools which never really work very well.
They get initial productivity boosts, but a lot of times that ended up actually slowing down a lot of companies.
Whatever the case, they are seeing enough growth that they are pushing forward their expectations.
This article came about after Business Insider saw a leaked investor memo that gave the billion-dollar projection for 2027.
Speaking of growth, after a spike in downloads this month, could Meta's AI app actually be gaining traction?
According to similar web data, Meta's standalone AI app now has over 300,000 downloads per day, up from around 100,000 in mid-September.
In addition, the app now has 2.7 million daily active users, up from 775,000 last month.
And while SimilarWeb said they hadn't seen any meaningful collation with either search or advertising volume, however, they noted that Meta could be promoting the platform on Facebook or Instagram, which aren't included in SimilarWeb's data.
The other possible explanation is that Meta's new vibes feed has been more of a success than people gave it credit for.
The AI generated image and video feed was released September 25th and decried by many as the introduction of infinite feeds of AI slop.
However, the spike in downloads and daily active users both do line up with the introduction of that feed.
OpenAI's launch of the Sora app a week later could also be boosting Meta's platform as an alternative.
Sora still requires an invite code while Meta's platform is freely available.
Now obviously these numbers in aggregate are still quite low relative to the billions of users that mainstream social apps have, but the growth is notable nonetheless.
Next up, some fundraising news.
OpenEvidence, the AI assistant for doctors, has raised $200 million at a $6 billion valuation.
This is the second large fundraising round for the company this year.
They raised $210 million at a $3.5 billion valuation back in July.
And with the level of growth they've displayed recently, it's not hard to see why the valuation has almost doubled.
Open Evidence now supports around 15 million clinical consultations a month, up from 8.5 million in July.
The product is free to use for registered medical professionals and monetized through advertising rather than subscription.
That unconventional approach for a professional tool has allowed Open Evidence to expand into 10,000 medical centers.
OpenEvidence only began commercializing their app three months ago and is already halfway to their target of $100 million in advertising revenue for next year.
The assistant is trained on leading medical journals like the New England Journal of Medicine and is designed to help doctors quickly access the literature for diagnosis and treatment options.
The system is also designed to reject low-confidence outputs, reducing hallucination risk.
Alongside medical journals, the model is also being fine-tuned on the 100 million clinical consultations assisted by the tool.
Co-founder Daniel Nadler said that this is one of the company's largest moats, adding, no one else in the world has that data.
Speaking to Adoption Among Doctors, Zangin Zeb of Google Ventures, the lead investor in the round, said it's reaching verb-like status.
Now, this data type of moat, where companies in verticals have access to actual real-world data based on the usage of their tool, is one of the most interesting themes and questions.
So far in the history of LLMs, we've seen that the bitter lesson applies.
In other words, that mass access to data beats out specialized data when it comes to pre-training.
However, where a lot of people are looking in the future is that the data that's left that the foundation model labs don't have is the data exhaust that comes from real-world usage, and that could in and of itself be extremely valuable.
That's certainly the argument that Open Evidence is making, and we'll have to see how it plays out.
Staying on fundraising, music-gen startup Suno is said to be in talks to raise $100 million at a $2 billion valuation.
Sources speaking with Bloomberg said the deal would quadruple the company's valuation since their last raise.
That last round closed in May of last year and brought in $125 million, although the valuation was not disclosed at the time.
Importantly, the startup is now generating $100 million in ARR, according to sources familiar with the numbers.
And what's more, Suno may be able to settle their legal disputes very shortly.
In June of last year, Universal and Warner Music filed a lawsuit for copyright infringement against Suno and competitor Unio.
But this June, Bloomberg reported that the labels are in talks to settle the litigation and establish a licensing framework for generated music.
The labels are also rumored to be looking to take an equity stake in both of those companies.
Reinforcing the idea of a truce between the music industry and AI startups, last week, Spotify announced plans to work with the record labels on AI-powered features.
Universal Music Group CEO Lucian Grange is boosting a pro-AI message internally.
Last week, he sent a memo to staff reemphasizing his interest in partnering on AI products as long as they respect artists' copyrights and likenesses.
Now, for anyone who has watched the history of the record labels all the way going back to Napster, this should be no surprise at all.
There is no industry, frankly, more adept at figuring out how to monetize the new thing.
Lastly today, the latest company to make some big AI pronouncement is Starbucks.
Starbucks CEO Brian Nicol said that they're all in on AI.
Appearing on a Yahoo Finance podcast recorded at the Dreamforce conference last week, Nicol discussed a wide range of AI deployments at the company.
A major scaled use case is an in-store knowledge assistant referred to as the Green Dot.
It helps store leaders manage daily operations, including troubleshooting equipment and providing drink recipes.
Nicol also said that Starbucks has pilots for inventory, supply chain forecasting, and scheduling...
although none of those use cases are at scale.
Speaking to ROI, he commented, We're still in the early days of this, but I believe there is definitely opportunity here to help us get things done faster and more efficiently, to what scale that is to be determined.
We're definitely already seeing a big impact in our technology area.
The ability to get code done so much faster is real.
One thing he did reject is the idea of robot baristas anytime soon, commenting we're not near that right now.
Some folks tried to dig into the specifics around what that would mean, while others just let it be vibes.
Sophie at NetCapGirl writes, Okay, yeah, whatever, F it.
Starbucks AI.
And that is going to do it for today's headlines.
Next up, the main episode.
Today's episode is brought to you by my company, Superintelligent.
Look guys, buying or building agents without a plan is how you end up in pilot purgatory.
Superintelligent is the agent planning platform that saves you from stalling out on AI.
We interview teams at scale, translate real work into prioritized agent opportunities, and deliver recommendations that you can execute on.
What to build, what success looks like, how fast you'll get results, and even what platforms and tools you should consider.
Visit besuper.ai and book your AI planning demo today.
AI isn't a one-off project.
It's a partnership that has to evolve as the technology does.
Robots and Pencils work side by side with clients to bring practical AI into every phase.
Automation, personalization, decision support, and optimization.
They prove what works through applied experimentation and build systems that amplify human potential.
As an AWS-certified partner with global delivery centers, Robots and Pencils combines reach with high-touch service.
Where others hand off, they stay engaged.
Because partnership isn't a project plan.
It's a commitment.
As AI advances, so will their solutions.
That's long-term value.
Progress starts with the right partner.
Start with robots and pencils at robotsandpencils.com slash AI Daily Brief.
What if AI wasn't just a buzzword, but a business imperative?
On You Can With AI, we take you inside the boardrooms and strategy sessions of the world's most forward-thinking enterprises.
Hosted by me, Nathaniel Whittemore, and powered by KPMG, this seven-part series delivers real-world insights from leaders who are scaling AI with purpose.
from aligning culture and leadership to building trust, data readiness, and deploying AI agents.
Whether you're a C-suite executive, strategist, or innovator, this podcast is your front row seat to the future of enterprise AI.
So go check it out at www.kpmg.us slash AI podcasts, or search UPenn with AI on Spotify, Apple Podcasts, or wherever you get your podcasts.
Welcome back to the AI Daily Brief.
One of the things that I have said frequently on this show, including being effectively the entire theme of yesterday's show, is that when it comes to the practical, lived, applied experience of AI inside a work setting, I don't think that AGI matters.
In fact, I think it is one of the more useless terms when it comes to how you think about applying AI in your daily life or your company thinks about applying it at work.
So why do definitions of AGI matter then?
And the short answer is, it's the exact same reason that we had that entire conversation on the show yesterday, which is that all of a sudden, progress towards AGI is going to be considered a meaningful factor when it comes to how markets should treat AI stocks.
Given how much AI stocks are at the core of the entire economy right now, these otherwise nebulous definitions start to take on a greater importance.
Now, of course, for those who haven't listened to yesterday's episode, AGI timelines are back in the news this week, specifically because OpenAI co-founder Andrej Karpathy said that he believes the technology is still a decade away, as opposed to estimates that have it more in a year or two.
Now, one critical point that came out of that conversation is that Andre actually has an extremely high bar for how he defines AGI.
He said, when people talk about AI in the original AGI and how we spoke about it when OpenAI started, AGI was a system that you could go to that could do any economically valuable task at human performance or better.
That was the definition.
He noted that since then, the definition has been watered down to just covering knowledge work, certainly nothing like physical work.
Now, knowledge work is certainly a huge part of the global economy, but at 10 to 20% of all the work in the world, at least as per his estimates, that leaves a lot off the table.
Now, this is far from the only definition floating around.
Way back in February of 2023, OpenAI laid out their framework for thinking about the approach of AGI.
They gave a very basic definition, AI systems that are generally smarter than humans.
Since then, Sam Altman has updated his thoughts.
He acknowledged in February of this year that AGI is a, quote, weekly defined term, but generally speaking, we mean it to be a system that can tackle increasingly complex problems at human level in many fields.
You might also hear Altman talking about AGI in reference to the five levels of AI framework.
Now, this built off of something that Google DeepMind scientists had introduced in a November 2023 paper, but then OpenAI expanded into these five stages.
Level one, chatbots, which were AI with conversational language.
Level two, which were reasoners with human-level problem solving.
Level three was agents with systems that can take actions.
Level four were innovators, AI that can aid in invention.
Level five, organizations, or AI that can do the work of an organization.
As we discuss a lot at this show, we are somewhere in the three to four range right now.
Beyond that, there are a range of other definitions you might come across.
Stahl, we're told, Gardner defines AGI as, quote, the intelligence of a machine that can accomplish any intellectual task that a human can perform.
Google leans into a different aspect, describing AGI as hypothetical intelligence of a machine that possesses the ability to understand or learn any intellectual task that a human being can.
Amazon has another distinct focus, describing AGI as software that is, quote, able to perform tasks that it is not necessarily trained or developed for.
Now, if these are one-off definitions for blog posts, one of the more prominent attempts to define and test AGI capabilities is, of course, the Arc AGI Prize.
On their website, they write, "...the consensus definition of AGI, a system that can automate the majority of economically valuable work, while a useful goal, is an incorrect measure of intelligence."
Measuring task-specific skills is not a good proxy for intelligence.
Skill is heavily influenced by prior knowledge and experience.
Unlimited priors and unlimited training data allow developers to buy levels of skill for a system.
This masks the system's own generalization power.
Intelligence lies in broader general purpose abilities.
It is marked by skill acquisition and generalization rather than skill itself.
So they propose a better definition for AGI, is AGI is a system that can efficiently acquire new skills outside of its training data.
The ArcAGI test then seeks to test two elements of AGI contained in the definition.
The ability to acquire new skills by ensuring the tests have internal logic that can be learned, and the ability to complete tasks outside of training data by ensuring the tasks are not generally available.
So these are all the things that are floating around, and you can see while they broadly get us in the right category, there are a lot of different definitions which lead to a lot of debates and a lot of AGI is in the eye of the beholder kind of conversations, which, as I said, I don't think really matters for our day to day, but does matter when it comes to whether giant funds are going to press the sell button because they think things are overbought because we're not making enough progress towards AGI, which means all these contracts aren't going to play out the way that they want to.
So this is the context into which a group of researchers working with the Center for AI Safety have attempted to nail down a common definition and a metric for assessing models as they progress.
The group has produced a paper called Definition of AGI, which you can find at agidefinition.ai.
In the abstract, they write, the lack of a concrete definition for artificial general intelligence obscures the gap between today's specialized AI and human-level cognition.
This paper introduces a quantifiable framework to address this, defining AGI as matching the cognitive versatility and proficiency of a well-educated adult.
This group, then, has grounded their analysis in Cattell-Horne-Carroll theory, one of the more well-accepted models of human cognition.
Applying the theory, the researchers split AI performance into 10 distinct categories.
Reading and writing, math, reasoning, working memory, memory storage, memory retrieval, visual, auditory, speech, and knowledge.
Now, you'll note that these categories cover some of the general performance categories, things like reading and writing or math, but it also addresses a model's ability to learn and apply its intelligence to topics outside of its training data.
Each of these categories has multiple subcategories that can be assessed individually.
In fact, assessment was one of the main focuses of this paper.
Researchers wrote, "...applications of this framework reveal a highly jagged cognitive profile in contemporary models."
While proficient in knowledge-intensive domains, current AI systems have critical deficits in foundational cognitive machinery, particularly long-term memory storage.
Each category was equally weighted and given a score out of 10, and researchers measured GPT-4 and GPT-5 to demonstrate the framework.
GPT-4 scored 27%, while GPT-5 achieved a 58%.
You can see from the two sets of results mapped out on a chart that while GPT-5 only made minor progress in knowledge, it made significantly more progress in reading and writing as well as math.
What's more, GPT-5 scored in multiple categories where GPT-4 was entirely deficient.
This included reasoning, working memory, memory retrieval, visual, and auditory.
And while those areas of intelligence are developing in the latest models, they're still very nascent compared to, for example, math.
Dan Hendricks, the director of the Center for AI Safety, commented, "...people who are bullish about AGI timelines rightly point to rapid advancements like math."
The skeptics are correct to point out that AIs have many basic cognitive flaws.
Hallucinations, limited inductive reasoning, limited world models, no continual learning.
There are many barriers to AGI, but they each seem tractable.
It seems like AGI won't arrive in a year, but it could easily arrive this decade.
Content creator Lewis Gleason wrote, What's powerful here is that this framework lets us track AGI like a scorecard.
For the first time, we have a framework that turns AGI from a buzzword into a measurable spectrum.
Instead of arguing are we close to AGI, we can now ask how much cognitive ground remains before parity.
Now, one of the interesting things about this framework is the focus on what's missing rather than highlighting a model's frontier abilities.
Over the summer, for example, GPT-5 and Gemini 2.5 Pro achieved gold medal performances in the International Mathematical Olympiad and the International Collegiate Programming Contest.
The leading models then are already at a human level, a very advanced human level when it comes to math or coding.
Importantly though, while achieving that level was a huge milestone on the path to AGI, based on this center's approach to an AGI definition, further progress in those areas isn't going to make a big difference.
In contrast, audio and visual understanding is still very nascent and needs to improve dramatically before AI models could be considered anywhere close to AGI.
Of course, those areas are arguably on the way.
Google has made incredible strides with their multimodal models over the past year, and visual understanding seems to be developing quickly.
The VO3 set of models and Sora 2 are also able to add appropriate audio to generated videos, implying strong auditory understanding.
The big area that is so clearly missing, the biggest hole, by a mile, is around memory.
The paper, in fact, describes this as perhaps the most significant bottleneck.
Now, of course, this is a huge area of focus for the labs.
Anthropic recently introduced their skills feature, which introduces a more efficient way of storing and accessing memory.
but we're yet to see a model that can intelligently store and retrieve information at anywhere close to a human level.
In fact, one of the things that you hear when people critique how far ahead the hype may have gotten in their estimation than where the capabilities of models are, it tends to come around to this part of cognition, where models don't have memory and they can't learn in the way that humans do.
Commenting on the study's exploration of memory from the paper, Rohan Paul noted, "...they show that today's systems often fake memory by stuffing huge context windows and fake precise recall by leaning on retrieval from external tools, which hides real gaps in storing new facts and recalling them without hallucinations."
They emphasize that both GPT-4 and GPT-5 fail to form lasting memories across sessions and still mix in wrong facts when retrieving, which limits dependable learning and personalization over days or weeks.
Anyone who has thought that they had locked in core knowledge and context about themselves with an LLM only to have it feed you back a response that has none of that understanding built in will understand what a big problem this actually is.
Now, what's valuable about this paper is, as Gleason put it, having a framework where there's an actual trackable numeric score that people can assess progress on.
For example, if all market actors accepted this framework, which of course won't happen, and then they went and looked and GPT-6 came out, instead of the inevitable endless debates about whether we had hit a wall again, theoretically, you could just look and see how much it had improved from GPT-4's 27% and GPT-5's 58%.
And yet at the same time, there is one highly problematic shortfall that could be very important.
Again, as Rohan Paul put it, the scope is cognitive ability, not motor control or economic output, so a high score does not guarantee business value.
In fact, increasingly other AGI definitions have fallen back on economic value as the most important proxy for intelligence.
Sometimes that's because more complex notions like continuous learning or performing tasks outside of the training set are too difficult to define.
One prominent example came from OpenAI's contract dispute with Microsoft.
Their agreement originally had Microsoft losing access to OpenAI's technology once AGI was achieved.
The problem was, of course, that the definition of AGI from OpenAI was pretty vague.
It defined AGI as, quote, highly autonomous systems that outperform humans at most economically valuable work.
The OpenAI board also had sole discretion to declare that AGI had been achieved.
This was viewed as an unfalsifiable claim that could cost Microsoft tens of billions of dollars.
The two companies ultimately settled on changing the definition of AGI to use a financial measurement as a proxy.
They decided that AGI would be deemed to have been achieved when OpenAI developed software that could generate $100 billion in profits.
Earlier this week, during the controversy around the Andre interview, Elon Musk revealed that he has a similar definition.
He posted on X that AGI is, quote, "...capable of doing anything a human with a computer can do, but not smarter than all humans and computers combined."
He said it's probably three to five years away.
He also put forward his belief that Grok 5 has a 10% chance to meet this definition and the odds are rising.
Now, I think there are of course merits to both economic and functional definitions of AGI.
The functional definition is laid out in the new paper, establishes the areas where current models are lacking and the new capabilities they will need to achieve AGI.
In some ways it functions almost as like a checklist, so we're all clear that incredibly intelligent models that forget everything at the end of the context window aren't really AGI.
But at the same time, an incredibly powerful model like Elon Musk is predicting Grok 5 will be, whether it's AGI or not, could have a profound impact on the economy.
In fact, as I've said numerous times, I think that these models are having and will have a profound impact on the economy exactly as they are right now.
Ultimately, I think this is an extremely useful contribution to the field.
I hope that more people dig in.
And if nothing else, it creates a useful heuristic for the future when inevitably, we rage and scream and kick with every new model release about how some big wall has been hit.
For now, that's going to do it for today's AI Daily Brief.
Appreciate you listening or watching as always.
And until next time, peace.
Today on the AI Daily Brief, some monster revenue numbers bring up a slew of questions on the business model for OpenAI and Anthropic, before that in the headlines, why Claude's new skills feature is potentially a really big deal.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
Firstly, thank you to today's sponsors, Gemini, Notion, Blitzy, Super Intelligent, and Robots and Pencils.
To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts.
And to find out about sponsorship opportunities, or pretty much anything else, speaking, job opportunities, etc., visit the show at ai-dailybrief.ai.
Welcome back to another AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes, and we have a jam-packed edition today.
We're kicking off today with a story that I think I will probably try to do a more operator's cut style episode as people dig in and figure out how to use these tools in the coming weeks.
But for now, it's rolled out a new feature called Skills, which provides agents with instructions, scripts, and resources to help them with specific tasks.
Users can fill a folder with these skills that might cover things like brand guidelines or instructions on carrying out a task in Excel.
The feature also allows users to provide executable code for situations where traditional programming is more reliable.
Cloud agents can then draw from these skills when they become relevant to the task at hand.
Essentially, skills are little barrels or buckets of context that Claude can draw on when it makes sense.
They're in a standard format that can be used across Claude apps, Claude code and the API, meaning you only have to build them once.
Said anthropic staffer Mahesh Murag, skills are based on our belief and vision that as model intelligence continues to improve, we'll continue moving towards general purpose agents that often have access to their own file system and computing environment.
The agent is initially made aware only of the names and descriptions of each available skill and can choose to load more information about a particular skill when relevant to the task at hand.
Now, part of the benefit here is that this makes the method token efficient.
Cloud agents can initially use very basic tools to figure out which skills they need for a given task and only spend significant tokens once it knows which ones to load.
They also function sort of like custom agentic scaffolding, but in a much more modular and user-friendly package.
A user doesn't need to know any programming language to create a custom skill that's fit for their purpose, which of course dramatically lowers the barrier to entry for advanced agent design.
You can also prompt Claude to design its own skills with the example that they give, saying, help me create an image editor skill.
Claude can also help them refine human design skills or monitor common failure points and then build skills to mitigate them.
Basically, Claude can be leveraged to collaborate on its own agentic design.
skills are also stackable.
So by way of example, Anthropic discussed an agentic workflow for building a quarterly investor deck where the agent would be able to tap into the company's brand guideline skill, a financial reporting skill, and a presentation formatting skill, coordinating all three without the need for manual intervention.
People very quickly picked up that this is sneakily a big deal.
Daniel Meisler wrote, Something I want to stress today.
MCP changed everything, but not because of a model improvement.
And today, skills change everything, but not because of a model improvement.
AI systems are the thing to watch, not just the intelligence of the models.
People are useful in jobs because they connect dots and can do many different things.
That's what Anthropic has been doing here, building a unified system that connects dots.
Simon Willison discussed in his blog how skills are awesome and maybe a bigger deal even than MCP, which is, of course, in bold words.
First, he described what skills are.
He said that conceptually, they're extremely simple.
They're a markdown file telling a model how to do something, optionally accompanied by extra documents and pre-written scripts that the model can use to help it.
However, he writes, there's one extra detail that makes this a feature, not just a bunch of files on disk.
At the start of a session, Claude's various harnesses can scan all available skill files and read a short explanation for each one.
This is very token efficient.
Each skill only takes up a few dozen extra tokens with the full details only loaded in should the user request a task that the skill can help solve.
He discussed how he would build a data journalism agent using the feature.
The skills that he would build were things like how to access and parse census data, how to load data from SQL and DuckDB with associated Python code provided, how to publish data online, and how to figure out an interesting story to tell based on the data.
Simon concluded, The core simplicity of the skills design is why I'm so excited about it.
Skills are marked down with a tiny bit of YAML metadata and some optional scripts in whatever you can make executable in the environment.
They feel a lot closer to the spirit of LLMs, throw in some text and let the model figure it out.
They outsource the hard parts to the LLM harness and the associated computer environment.
Given everything we have learned about LLM's ability to run tools over the last couple of years, I think that's a very sensible strategy.
Put differently, I think skills are in some ways a different user experience pattern for getting at agent creation.
That's really what this is about.
On the one hand, you've got the end-to-end style agent workflow builder, but this is basically where you articulate component parts and then can use natural language to help the LLM itself figure out which of those parts it needs to put together.
Again, in Simon's example, he's got a skill for parsing census data, a skill for loading data, a skill for publishing data online, and a skill for figuring out an interesting story.
Instead of having to put them together in a step-by-step workflow schematic,
He can just make sure that each individual piece has all of the context it needs, and then the LLM can figure out how to put it together, which strikes me if it works as a much more intuitive way to organize this than some of those other interfaces.
There's going to be a ton of examples, I think, coming up.
And like I said, I'm planning on maybe an operator's cut style episode where we get into some ways to actually use Claude's skills.
But for now, we've got a lot to cover.
So let's move on to Microsoft.
who wants to make every PC an AI PC with a big new update to Windows 11.
Now, on the surface, this is an update about bringing Copilot to all users and making it more central to the Windows experience.
Users can now summon the AI assistant by saying, hey, Copilot.
They're also rolling out Copilot Vision, which allows the AI model to see what's happening on the desktop and use it as input.
In addition, agentic features will be introduced as early previews.
Those agentic features will allow Copilot to tap into data from emails, calendars, and the Office suite, as well as taking actions in the file system itself.
The biggest difference from previous AI releases is that none of these features will be restricted to copilot plus hardware.
They will just be a default part of the Windows 11 experience for all users.
Executive VP and Consumer Chief Marketing Officer Yousef Mehdi explained, We think we're on the cusp of the next evolution, where AI happens not just in a chatbot and gets naturally integrated into the hundreds of millions of experiences that people use every day.
The vision that we have is, let's rewrite the entire operating system around AI and build essentially what becomes truly the AI PC.
The rewrite is built around two main features, agentics and voice.
Mehdi said, you should be able to talk to your PC, have it understand you, and then be able to have magic happen from that.
With your permission, we want people to be able to share with their AI on Windows what they're doing and what they're seeing.
The PC should be able to act on your behalf.
He added, In our minds, voice will now become the third input mechanism to use with your PC.
It doesn't replace the keyboard and mouse necessarily, but it's an added thing and it will be pretty profound and a new way to do it.
They also have a new feature coming called Copilot Actions, which is sort of similar to OpenAI's Operator or Google's new Gemini Enterprise.
The feature spins up a new window where the users can give directions and watch the agent complete tasks using local files.
They can either monitor the agent and take over at any time or click away to a different window and let the agent run in the background.
Now, as I've said before, despite what seem like some stumbles and missteps, Microsoft has an incredible distribution network and serious benefits when it comes to consumer installs.
And this is a great example of how owning the entire end-to-end experience could be a game changer when it comes to how much value you can unlock for a particular user.
Moving over to IP land, Spotify has reached a deal with the major music labels on how AI will intersect with the music industry.
Spotify will collaborate with Sony, Universal, Warner, Merlin, and Believe to, quote, develop responsible AI products that empower the artists and songwriters they represent and connect them with the fans who support them.
In a press release, Spotify wrote, Some voices in the tech industry believe copyright should be abolished.
We don't.
Musicians' rights matter.
Copyright is essential.
If the music industry doesn't lead in this moment, AI-powered innovation will happen elsewhere without rights, consent, or compensation.
Now, the announcement didn't discuss any specific products, but it did reinforce that the introduction of AI features would fundamentally be the choice of rights holders and artists.
They're also aiming to build AI products that create wholly new revenue streams and want to ensure that rights holders are properly compensated and credited appropriately.
A lot of the coverage was extremely skeptical.
Spotify has recently been criticized for allowing low-effort AI tracks to proliferate, leading to a purge of 75 million AI-generated tracks in September.
But others are willing to give the benefit of the doubt.
Ed Newton Rex, the CEO of copyright advocacy group Fairly Trained, posted,
In my opinion, it's a good thing that Spotify is working on AI music tools with the major labels.
Lots of the AI industry is exploitative.
AI built on people's work without permission served up to users who get no say in the matter.
This sounds like it will be different.
AI features built fairly with artist permission presented to fans as a voluntary add-on rather than an inescapable funnel of AI slop.
To be clear, Spotify's general stance on AI music is not good.
They seem happy to allow AI slop on the platform without labeling it, even featuring it in recommended playlists.
This needs to change.
But credit where it's due.
Licensing music for training instead of taking it without permission is very welcome.
The devil will be in the details, but in tone at least, this is a positive development.
Now, that might not sound all that positive to you, but given how absolutely contentious the relationship is between the AI and anti-AI crowd, that is not just a fig leaf, that's a fig tree.
Whether Oracle can actually pull off that line remains to be seen, but I think that they are in a good position to give it a shot.
A couple quick product updates.
General purpose agent Manus has been updated for version 1.5.
The team wrote faster, better quality results, unlimited context, build full stack web apps with real AI features, backends, user logins, custom domains, and analytics.
Run your biz, start your side hustle, or just have fun.
If you can dream it, Manus can build it.
Manus claimed that tasks are now four times faster on average, and the performance boost should allow tasks that once failed to be completed successfully.
Manus's internal benchmark showed a 15% improvement in task quality and a 6% increase in user satisfaction.
The big new feature is the ability to build web apps end-to-end without leaving Manus, making the agent more useful as a vibe coding platform.
Kind of seems like ultimately everything, even the general purpose agents, are a vibe coding platform in the end.
Another small but potentially powerful feature announcement.
Agent orchestration startup Lindy has announced an AI chief marketing officer for your company.
CEO Flo Crivello wrote, announcing Lindy AICMO, a team of agents running entire marketing workflows end-to-end, handling market research, analysis, and creative at scale, so you can start thousands of ad experiments in minutes.
As part of this, we're launching new integrations with Sora 2, VO 3.1, Nanobanana, and GPT ImageGen.
Over the last few years, agents have grown tremendously in their scope and autonomy, going from handling simple tasks to assuming entire roles.
This will be the first time that we see them take a chunk out of an entire org.
I'm considering trying this out as CMO of AIDB, given that, as I've mentioned, I'm hiring for some growth roles.
If that's an experiment that you would be interested in following along with, let me know and we'll see what can happen.
Lastly, I am continuing to do our ROI corner, ROI shoutouts.
ROI is going to be a major theme for next year.
And so when we see companies reporting actual ROI from AI, you are going to get it here.
The latest comes from Alibaba, who have announced that they've reached breakeven on AI in their e-commerce business.
In February of this year, the Chinese tech giant announced they would spend $53 billion on AI over the next three years.
They've since rolled out a string of features, including personalized AI search and virtual clothing try-ons.
Now, a company official has announced that preliminary testing of their AI features has shown consistent results, including a 12% increase in return on advertising spend.
VP Kaifu Zhang said it's very rare to see double-digit changes.
Zhang predicted that AI would have a, quote, very significant positive impact on sales during this year's Singles Day shopping period, which centers around November 11th.
Meanwhile, I would anticipate that in the US, we're going to see a bit of an AI-focused Black Friday as well.
But it's not even Halloween yet.
We still have a little time before that.
And so with that, we will wrap today's headlines.
Next up, the main episode.
Small, nimble teams beat bloated consulting every time.
Robots & Pencils partners with organizations on intelligent, cloud-native systems powered by AI.
They cover human needs, design AI solutions, and cut through complexity to deliver meaningful impact without the layers of bureaucracy.
As an AWS-certified partner, Robots & Pencils combines the reach of a large firm with the focus of a trusted partner.
With teams across the US, Canada, Europe, and Latin America, clients gain local expertise and global scale.
As AI evolves, they ensure you keep peace with change.
And that means faster results, measurable outcomes, and a partnership built to last.
The right partner makes progress inevitable.
Partner with Robots and Pencils at robotsandpencils.com slash AI Daily Brief.
Chatbots are great, but they can only take you so far.
I've recently been testing Notion's new AI agents, and they are a very different type of experience.
These are agents that actually complete entire workflows for you in your style.
And best of all, they work in a channel that you already know and love because they are purpose-built Notion superusers.
Notion's new AI agents completely expands the range of what Notion can do.
It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days and get them complete in minutes.
These agents don't just help with work, they finish it.
Getting started with building on Notion is easier than ever.
Notion agents are now your very own super user to help you onboard in minutes.
Your AI teammates are ready to work.
Try Notion AI for free at the link in our show notes.
What if AI wasn't just a buzzword, but a business imperative?
On You Can With AI, we take you inside the boardrooms and strategy sessions of the world's most forward-thinking enterprises.
Hosted by me, Nathaniel Whittemore, and powered by KPMG, this seven-part series delivers real-world insights from leaders who are scaling AI with purpose.
from aligning culture and leadership to building trust, data readiness, and deploying AI agents.
Whether you're a C-suite executive, strategist, or innovator, this podcast is your front row seat to the future of enterprise AI.
So go check it out at www.kpmg.us slash AI podcasts, or search You Can With AI on Spotify, Apple Podcasts, or wherever you get your podcasts.
Today's episode is brought to you by Superintelligent.
Now, for those of you who don't know, who are new here maybe, Superintelligent is actually my company.
We started it because every single company we talk to, all the enterprises out there, are trying to figure out what AI can do for them, but most of the advice is super generic, not specific to your company.
So what we do is we map your AI and agent opportunities by deploying voice agents to interview your teams about how work works now and how your people would like it to work in the future.
The result is an AI action map with high potential ROI use cases and specific change management needs, basically everything you need to go actually deliver AI value.
Go to bsuper.ai to learn more.
Welcome back to the AI Daily Brief.
At the core of all of the conversations around whether or not we are in an AI bubble is the question of the growth potential of the core model provider companies.
Basically, we have a situation where there's this set of deals that are perceived as increasingly circular
that are all based on getting more compute and more electricity to power that compute, more data centers to deliver that compute, et cetera.
But all of that is based on the assumption that that compute will turn into AI products that have enough demand to generate enough revenue to make all of that infrastructure build out make sense.
So for that reason, people are paying extremely close attention to indications of how these businesses are growing.
And the short answer seems to be like gangbusters.
So we got two interesting reports this week, both of which dealt to some extent with revenue from these companies.
The first was about Anthropic, and it showed just what an absolute tear this company has been on this year.
Now, Anthropic started the year at around $1 billion in annualized revenue.
That's where they were in the December-January timeframe.
Over the summer, they nudged up to $5 billion in annualized revenue.
And according to Reuters, they are now running at a $7 billion revenue run rate and are on track to meet a goal of $9 billion by the end of the year.
900x growth from a billion-dollar base in a single year is absolutely insane.
And while they don't think that they're going to continue that unbelievably breakneck pace into 2026, they still do think that they can reach between $20 billion and $26 billion in annualized revenue by the end of next year.
Okay, so that's the Anthropic story.
OpenAI, meanwhile, began the year at around $5.5 billion and have now grown apparently to around $13 billion.
Now, in any normal world, $0 to $13 billion in three years after the release of their core product would be seen as absolutely insane growth, and it is.
The challenge is that they are projecting $100 billion in annual revenue by 2028.
The Financial Times wrote a piece about the big obligations that OpenAI is putting on its back called OpenAI Makes Five-Year Business Plan to Meet $1 Trillion Spending Pledges.
This, of course, they're referring to all these deals with NVIDIA, Oracle, etc.,
Now, it's very clear that OpenAI isn't planning on being able to finance all of this just with revenue.
However, as I said, they are projecting $100 billion in three years, meaning they are going to need to see a big uptick in revenue from where they are.
And alongside the Financial Times article, we also got some reported information around the state of the business.
$13 billion in revenue includes about 70% of that from consumer subscriptions and around 30% of that from the API.
Of their 800 million users, about 5% are paying monthly or $40 million.
They're also at a $20 billion burn rate.
So let's talk about some of the interesting questions that I think this brings up that are more slanted towards the strategy of these companies going forward.
So one question is, does the radical breakout success of Anthropic push OpenAI to be more aggressive about enterprise?
In that Reuters piece, we also learned that 80% of Anthropic's revenue comes from the enterprise.
They have more than 300,000 business and enterprise customers.
They've recently announced big deals like the one with Deloitte, where they have hundreds of thousands of users coming from a single company.
And given their 9x growth...
The fact that they have gone from 1 20th or less of OpenAI's revenue to more than half of OpenAI's revenue in just 10 months has got to be pushing the question of whether there should be more emphasis on the enterprise from OpenAI.
Now, of course, it's not like OpenAI has no enterprise focus.
The company has even recently added a forward-deployed engineer team that will go in and actually help companies build out solutions if apparently they are a $10 million or more customer.
And interestingly, at Dev Day, Altman and Brockman did speak a little bit about enterprise.
Brockman talked about how there were a lot of really boring enterprise problems that were some of the juiciest things to solve.
And Altman talked about how they didn't just want FTEs to be bungling around, but now they feel like they really have a good idea of how to help enterprises, and so they want to scale that.
All of that suggests to me that we might see at least a bit more enterprise focus heading into 2026.
Now, of course, one of the interesting questions, which has a big stake in this strategy as well, is how much of this enterprise success is strictly anchored in Anthropic's coding prowess.
For most of this year, really up until GPT-5, Anthropic were the preferred coding models by far and away, with Gemini models a distant second and OpenAI models not really in the conversation.
We've covered a lot about how that shifted, and it was very clear that a huge goal of GPT-5, if not the primary goal of GPT-5, was to redress this balance.
And I wonder to what extent OpenAI sees coding as the Trojan horse for absolutely everything in enterprise and as the thing they need to focus on, even if their broader goal is more general enterprise revenue as well.
Still, I don't think it's a foregone conclusion that OpenAI's only play is to double down on enterprise.
The numbers that Anthropic is putting up certainly suggest how valuable that segment is, but at the same time, OpenAI really does have a unique position vis-a-vis consumer.
For a huge number of people, AI and ChatGPT are completely inextricable from one another.
In fact, ChatGPT is the term more than AI that they know.
Aside from a few dips and blips, they've been the number one app for a huge part of their life since ChatGPT was launched.
And every week that goes on, ChatGPT gets more and more deeply woven into the consumer experience.
Now, one of the questions for a lot of people is how much more there is to be gained from those 800 million users.
Max Sparrow, for example, tweeted, Is it just me or is 40 million paying ChatGPT users kind of low?
Spotify, for example, has 276 million paid subscribers.
Just kind of surprised that there are 760 million people who don't feel like they derive an additional $20 of value from a more capable model.
This to me, I think, is a really important question.
It certainly feels like there is a lot more to be gained from these users.
And my guess is that OpenAI is betting that the more valuable that these models get over time, the higher the percentage of users they'll be able to unlock and convert to a higher tier.
But it also seems like they're not taking that for granted.
And so another question that comes up is, does this make ads inevitable?
My read having watched all of these discussions is that OpenAI really doesn't want to have to do ads, at least not in the traditional way.
I think that they genuinely don't really want to be strictly speaking in the attention business in the way that Meta or TikTok or Instagram are.
But I also think that they feel like it is to some extent inevitable and a huge source of revenue.
And certainly, if you're trying to go from $13 billion where we are at the end of 2025 to $100 billion by the end of 2028, it seems like there are going to have to be new revenue streams unlocked that aren't just converting more paid subscribers.
Certainly, I think that that brings up the question of how much revenue is possible from ads and what they needed to do as part of the overall mix if they were to go into this area.
Now, one thing that hasn't been discussed as much is that it is possible that OpenAI finds some slight variation on the traditional ad unit experience that better befits the user experience patterns of interacting with ChatGPT and that make it less annoying and more palatable for users than, for example, big chunky display ads.
Indeed, it feels like that's kind of what they're thinking about with their checkout with ChatGPT business and in general, taking a cut of referrals from the part of the ChatGPT experience that's all about discovery.
This definitely feels like a thing that they're going to do and an area where a version of advertising feels less disruptive of the user experience and more potentially additive or at least integrated with the user experience.
And I wouldn't be surprised if that's the sort of template that we see when it comes to more traditional advertising style business models for open AI is trying to find things that are more subtle and more integrated as ways to capture additional incremental value from each average user.
I think one of the big questions for me is whether some amount of this solves itself as we just better understand what use cases actually drive the most value.
In other words, I think that probably a lot of the users of ChatGPT are radically under maximizing their use of it.
And the more time that goes on, the more people around them will have figured out how to use it for more value, and the more likely they are to uncover some use case that's worth 20 bucks a month or whatever the price is for them to become a paid user.
One of the interesting ways to look at the study that OpenAI did about how people use ChatGBT is through this lens of trying to understand which use cases actually drive the most value, again with an eye to are those possibly the use cases that could be better monetized.
There's also the interesting question, I think, of whether subscriptions in general are the right model and whether these subscription prices are the right prices for those subscriptions.
To some extent, we're kind of all just living in the environment that Altman and OpenAI plucked out of the air when it came to pricing.
They decided 20 bucks a month for consumer and 30 bucks a month for enterprise users sort of made sense.
And then when they added a $200 a month premium subscription, everyone anchored to that.
In other words, there hasn't been a ton of price discovery around whether these are exactly the right prices.
And I think that there's probably some room for experimentation there as well.
Now, on the enterprise side, one of the major questions, I think, is to what extent companies are going to prioritize just state-of-the-art models versus there's going to be more room for cheaper models for even production-grade use cases.
My instinct is that we are right on the cusp of a threshold where big chunks of high-value use cases are completely viable with not state-of-the-art models, but with the newer, more efficient models like Claude's Haiku 4.5.
I also think with agentic systems, they're going to start to get smart enough that they know which models they need to use for different tasks, and they're not always going to have to call on the most expensive models.
We even got a little glimpse of how this type of organization might work when Cloud Skills was launched, where one of the things that people are excited about for skills is that they don't need a super sophisticated model to initially see a user's request and figure out which skills to draw from, making the whole experience more efficient for the user.
Part of why I'm bullish on the enterprise use case, even for open AI that has such a distinct advantage in consumer, is that I just think that especially as the price equation gets figured out, we're going to see some amount of augmentation or automation across basically every core business workflow.
And that is just a ton of tokens that can be consumed for revenue.
Now, I should note it here at this point that I'm just talking about OpenAI and Anthropic because those are the ones we got revenue numbers from this week.
Obviously, Grok, Gemini, and Microsoft Copilot are all in this conversation as well.
And what they do could impact these questions.
For example, one of the good reasons for OpenAI not to focus on enterprise so much is if they think that space is just going to be too out-competed.
One sort of interesting question that's more about how much these companies become a black hole for revenue in the sector in general is whether the foundation model companies like Anthropic and OpenAI are likely to siphon off revenue from the vertical apps or the other way around.
Related to that, are we likely to see more verticalized products from OpenAI and Anthropic?
We've seen the first indications of them dancing with that.
Anthropic has released some industry-specific tools.
But I think this is one of the big important questions for how this all shakes out.
There are right now a lot of AI companies out there who have raised to nine figures of revenue coming up even on a billion.
And whether ultimately the foundation model companies are a great big sucking sound that pulls all that in, or whether there's room for those companies as well, could also shape how this market plays out.
I think related to that is the question of whether we're going to see these companies start to acquire their way to additional revenue.
The Information This Week also posted an article about how Anthropic is planning on doing more acquisitions.
They wrote, Anthropic told investment bankers in recent weeks that it's getting ready to move off the sidelines and do more acquisitions.
It is apparently considering M&A targets that could be good for its tech or for its end products.
Still, at least for now, it does seem like the acquisitions that they're interested in are a little bit more on the technical side and less on the product side, but there's no saying that that won't change.
Ultimately, the two big questions are these.
The first is, can any company grow this fast?
Epic AI research dug into this and pointed out that we are in extremely uncharted territory.
They wrote, OpenAI's revenue growth has been extremely impressive, from less than a billion to over 10 billion in only three years.
We found four other US companies in the past 50 years that have done this, but of those, they point out, only Google went on to reach above 100 billion in revenue.
Trying to find other companies that went from $10 billion to $100 billion in revenue in under a decade, they found seven.
Google, Walmart, Apple, Amazon, Tesla, Meta, and Nvidia, but none of them even did it in six years, let alone three.
Now, again, we are in uncharted territory, but it really would represent the fastest growth of anything that we've ever seen.
Which brings us to our second question, what is sufficient growth?
And by that, I mean sufficient growth to make the market still have confidence in all of these deals and the big infrastructure build-out.
This is, of course, a completely subjective question that could change wildly based on market conditions that have nothing to do with AI.
Sufficient in this case is a barometer of investor confidence in the future, not any sort of harder objective function.
And so to some extent, if you want to keep track of where we are in the bubble cycle, I think the key thing to look at is how the growth numbers that we're getting from these companies compared to what is the inevitably moving target of sufficient growth.
Look, ultimately, we are in completely uncharted territory, and the future is being written live as we watch.
It is going to be a hell of a couple years.
So buckle in, listen to AIDB, and get ready for a whirlwind.
That's going to do it for today's AI Daily Brief.
Appreciate you listening or watching, as always.
Until next time, peace.
Today on the AI Daily Brief, OpenAI opens the door to adult AI content.
Before that in the headlines, Citigroup's use of AI has freed up 100,000 hours for their developers each week.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's
Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes.
Since we have so many of these MIT-style studies claiming no positive results from AI floating around, we're going to start a new little corner on the show called the ROI Spotlight, where we highlight companies that are making announcements about how much actual tangible value they're getting from AI in the here and now.
Kicking off this segment, we have Citigroup, where the bank has claimed huge AI efficiency gains in their earnings report.
They said that their two main enterprise-wide AI tools combined for 7 million utilizations last quarter, which tripled the usage of the previous quarter.
In addition, AI coding tools have completed 1 million code reviews year-to-date, saving in their estimation around 100,000 hours per week across their developer teams.
Remember, this is not showing up in a press release.
This is showing up in an earnings report.
And as we move into a 2026 where ROI is one of, if not the dominant enterprise AI theme, we will continue to keep an eye out for reported results like this.
Now, speaking of enterprises and AI, yesterday, Walmart announced that they were partnering with OpenAI on AI shopping.
A couple of weeks ago, we heard about ChatGPT's new checkout features, and Walmart now joins as the biggest partner to take advantage of them.
ChatGPT users will now be able to shop Walmart products right in the app, complete with a buy button and integrated checkout.
Daniel Denker, Walmart's Executive VP of AI Product and Design, said, "...we view this as an opportunity to deliver convenience in a way that meets customers where they are."
Said Walmart President and CEO Doug McMillan, "...for many years, e-commerce shopping experiences have consisted of a search bar and a long list of item responses."
This is about to change.
There is a native AI experience coming that is multimodal, personalized, and contextual.
We're running towards that more enjoyable and convenient future with Sparky and through partnerships, including this important step with OpenAI.
Sparky, by the way, is part of Walmart's super agent strategy, where they reorganized dozens, if not hundreds of sub-agents into four main agents.
If you're interested in that, go check out the episode, Walmart Blast Past Agent Experimentation.
You can find it on YouTube or your podcast channels.
It came out at the end of July.
We talk frequently on this show about the difference between efficiency AI and opportunity AI.
Efficiency AI is, of course, using AI to do the things that you do now faster, better, cheaper.
Opportunity AI is about new experiences and opportunities that simply weren't possible before.
And while it may seem small, what's interesting to me about this partnership between Walmart and OpenAI is that it nudges more in that direction of opportunity AI and fundamentally new types of consumer experiences than just doing the old stuff cheaper.
Another big partnership with OpenAI yesterday came from Salesforce.
Salesforce CEO Mark Benioff tweeted, Salesforce and OpenAI partnership unleashed.
New AgentForce 360 apps now live in chat GPT.
Query CRM, build Tableau dashboards, analyze conversations, even close deals with AgentForce Commerce.
Trusted data and seamless workflows plus the world's best AI equals unstoppable enterprise power.
Now, we talked on Tuesday's show all about Salesforce's attempt to turn themselves into an AI context platform.
But what was interesting about this announcement yesterday is how much it did not move Wall Street, at least not in the direction that you would have expected.
Over the past month, we've seen Oracle, AMD, and Broadcom all see big pops from their announced OpenAI deals.
We also saw Coursera, HubSpot, Figma, and Bookings.com also all have huge pops when they were mentioned on stage at OpenAI Dev Day.
And yet following the announcement, Salesforce stock was down 3.6%, its worst day in over a month.
So does this mean that the OpenAI magic is fading and markets no longer care about big AI announcements?
Not necessarily.
Salesforce has had a number of tough years.
with earnings growth slowing way down across that period.
This year's growth forecast is less than 10%, which is below other leading tech companies and way down from the 25% plus that they maintained for over a decade until 2023.
As a smaller sub-story, Salesforce also announced on Tuesday that they would not be paying a ransomware demand that could impact customer data, so that could also be a part of what's driving the drawdown.
Still, you better believe as people look for signs of an AI bubble bursting, a stock tumbling on the announcement of an open AI deal is raising some eyebrows.
Back over in news that the market likes, we have the latest in the chip wars where Intel is preparing to get back into the AI chip market with the release of a new GPU next year.
One of the biggest knocks on Intel has been that they sat out the entire AI boom, but that's actually not technically true.
Intel produced various CPUs that were used in NVIDIA AI server racks.
They also designed a series of AI accelerator chips culminating in the Gaudi 3 chip released in April of last year.
That said, the Gaudi chips never captured any meaningful market share.
The new chip is codenamed Crescent Island and is expected to be ready for customer testing in the second half of 2026.
This will be the first release under a new plan to produce new GPUs on an annual basis rather than an irregular schedule.
During July's earnings call, CEO Lip Bhutan said that the company's focus would be on efficient AI chips for serving low-cost inference.
In a statement for this announcement, CTO Sachin Kati confirmed that plan, commenting, AI is shifting from static training to real-time everywhere inference driven by agentic AI.
Scaling these complex workloads requires heterogeneous systems that match the right silicon to the right task, powered by an open software stack.
He added that Intel's new chip will, quote, provide the efficient headroom customers need and more value as token volumes surge.
Remains to be seen whether Intel can actually accomplish this, but there is a lot of merit to their focus on cheap inference rather than training.
Then again, there's no shortage of companies designing their own GPUs, TPUs, and ASICs for efficient inference, so exactly where Intel fits into all of this remains to be seen.
Lastly today, Oracle announced plans to deploy 50,000 AMD GPUs as the Blockbuster AI infrastructure deals continue.
Oracle said on Tuesday that they expect the installation to begin in the second half of next year using AMD's new M1450 chips.
It's part of a longer-run commitment from Oracle to start making widespread use of AMD chips.
There's also a likely link to OpenAI's recent deal with AMD to buy 10 gigawatts worth of chip supply.
Kiran Bhatta, the senior VP of Oracle Cloud Infrastructure, suggested the move is about AMD somewhat catching up to Nvidia.
He said, "...we feel like customers are going to take up AMD very, very well, especially in the inferencing space.
I think AMD has done a really fantastic job, just like Nvidia, and I think both of them have their place."
Separately, Oracle's new co-CEOs have defended the company's aggressive AI build-out.
Last month, Mike Cecilia was promoted from President of Oracle Industries, and Clay McGorick was promoted from President of Cloud Infrastructure to take over leadership of the company.
Cecilia told the Wall Street Journal, "...we're really in a unique situation to deliver what we call applied AI."
Now, Oracle has been on a bit of a journey in the last month.
They saw a huge pop after announcing a $300 billion deal with OpenAI, but then the information later reported that Oracle has low margins on their AI product, driving the stock price down.
But now many analysts believe the reporting mischaracterized the accounting.
Derek Wood of TD Cowen said, you have to go build the infrastructure before you can turn on all the revenue meters.
But as the consumption meters start going on, you start to recoup a lot more of your capital expense and start to see gross margins significantly improve.
Still, the stock remains 0.8% down off its recent high, and in that context, the new co-CEOs will make their case to investors at an investor day on Thursday.
If there is anything interesting there, we will of course report on it, but for now, that's going to do it for today's AI Daily Brief Headlines edition.
Next up, the main episode.
Today's episode is brought to you by my company, Superintelligent.
Look guys, buying or building agents without a plan is how you end up in pilot purgatory.
Superintelligent is the agent planning platform that saves you from stalling out on AI.
We interview teams at scale, translate real work into prioritized agent opportunities, and deliver recommendations that you can execute on.
What to build, what success looks like, how fast you'll get results, and even what platforms and tools you should consider.
all customized for you.
Instead of shopping for hype, you get to deploy with confidence.
Visit besuper.ai and book your AI planning demo today.
What if AI wasn't just a buzzword, but a business imperative?
On You Can With AI, we take you inside the boardrooms and strategy sessions of the world's most forward-thinking enterprises.
Hosted by me, Nathaniel Whittemore, and powered by KPMG, this seven-part series delivers real-world insights from leaders who are scaling AI with purpose.
from aligning culture and leadership to building trust, data readiness, and deploying AI agents.
Whether you're a C-suite executive, strategist, or innovator, this podcast is your front row seat to the future of enterprise AI.
So go check it out at www.kpmg.us slash AI podcasts, or search You Can With AI on Spotify, Apple Podcasts, or wherever you get your podcasts.
Today's episode is brought to you by Robots & Pencils.
With AWS-certified teams across US, Canada, Europe, and Latin America, clients get local expertise and global scale.
And with a laser focus on real outcomes, their solutions help organizers work smarter and serve customers better.
They're your nimble, high-service alternative to big integrators.
Turn your AI vision into value fast.
Stay ahead with a partner built for progress.
Partner with Robots & Pencils at robotsandpencils.com slash AI Daily Brief.
Chatbots are great, but they can only take you so far.
I've recently been testing Notion's new AI agents, and they are a very different type of experience.
These are agents that actually complete entire workflows for you in your style.
And best of all, they work in a channel that you already know and love because they are purpose-built Notion superusers.
Notion's new AI agents completely expands the range of what Notion can do.
It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days and get them complete in minutes.
These agents don't just help with work, they finish it.
Getting started with building on Notion is easier than ever.
Notion agents are now your very own super user to help you onboard in minutes.
Your AI teammates are ready to work.
Try Notion AI for free at the link in our show notes.
To some, this news was completely inevitable.
To others, it feels like a major course shift.
To some, it's just a head scratcher.
And yet for others, in this latest announcement, they see the signs of deep cultural fracture that could create major problems for AI in the years ahead.
Mario Knopfel, X's aggregator-in-chief, writes, ChatGPT's adult mode is coming, literally.
Oh, Mario.
Starting December, OpenAI will let adult users access mature content on ChatGPT, but only if they verify their age.
CEO Sam Altman says it's part of treating adults like adults, which includes stuff like erotica and more customizable chatbot personalities.
OpenAI made the bot pretty strict before to protect people with mental health struggles, but now says it has the tools to ease those limits safely.
Basically, ChatGPT might soon flirt back, drop emojis, or talk like your bestie, but only if you want it to.
Alright, so as you might imagine, the erotica part of this has captured the most attention.
But let's try to get the whole picture before we dive to any conclusions.
On Tuesday, Sam Altman tweeted...
We made ChatGPT pretty restrictive to make sure we were being careful with mental health issues.
We realized this made it less usable and enjoyable to many users who had no mental health problems, but given the seriousness of the issue, we wanted to get this right.
Now that we have been able to mitigate the serious mental health issues and have new tools, we are going to be able to safely relax the restrictions in most cases.
In a few weeks, we plan to put out a new version of ChatGPT that allows people to have a personality that behaves more like what people liked about 4.0.
We hope it will be better.
If you want your ChatGPT to respond in a very human-like way, or use a ton of emoji, or act like a friend, ChatGPT should do it.
But only if you want it, not because we are usage-maxing.
In December, as we roll out age-gating more fully and as part of our Treat Adult Users Like Adults principle, we will even allow more, like, erotica for verified adults.
There is so much to unpack around all of this.
One interesting part of it has to do with the further fallout of the launch of GPT-5 and the deprecation of 4.0.
Back when that happened, there was, of course, a massive rebellion and outcry of users who had gotten used to the interaction patterns with 4.0 and who didn't like the cold and clinical feeling of GPT-5.
At the time, there were lots of questions around the sycophancy of models and what subtle but insidious problems that might create.
There were conversations around strategy and obligation to users and how people form relationships with models that the companies who released those models kind of had to accommodate for even if they didn't expect that to happen.
There was also just the straight up contingent of people who thought of 4.0 as something close to a romantic partner.
Now, interestingly, although the 4.0 rebellion worked and people got their beloved model back, at least for a time, as we later learned, it seemed like in part 4.0 had been deprecated because it didn't function properly with OpenAI's new content moderation system.
For example, shortly after Foro was brought back, OpenAI noted that conversations on certain sensitive topics would be filtered to GPT-5.
This was part of a blog post from mid-September called Teen Safety, Freedom, and Privacy.
As with most of OpenAI's most important blog posts, this one was authored not by the team, but by Altman himself.
In addition to noting that certain types of conversations were going to be auto-routed to GPT-5 because their content moderation system could better handle them, we also heard the first inklings of how OpenAI was going to approach this age-gating issue.
And while the thrust of the blog post in general was that OpenAI really wanted to prioritize freedom and privacy, and for adults over a sort of nannying safety, that when it came to teens, the reverse was true.
Altman wrote, We prioritize safety ahead of privacy and freedom for teens.
This is a new and powerful technology and we believe minors need significant protections.
Altman noted that they were building an age prediction system to estimate age based on how people use ChatGPT, but if there was any doubt, they'd play it safe to default to the under-18 experience.
So to the extent that this is just a continuation of what they've already been talking about, i.e.
a bifurcation of the experience into a non-adult and an adult experience, there's not all that much that's necessarily notable here.
Also, it should be noted that this is not the first time that Altman has talked about allowing adults to use ChatGPT for adult conversation and adult content.
This is something that's come up in previous interviews going back quite some time.
There are a couple things I think that make this conversation different right now, though.
The first is that in the weeks following GPT-5, Altman took a number of different potshots, presumably aimed at Elon Musk, who had allowed his Grok app to lean much more into the adult and erotic type content.
Writes Fortune, Altman did jab at companies developing Japanese anime sex bots, continuing, "...you will not see us do that.
We will continue to work hard at making a useful app, and we will try to let users use it the way that they want, but not so much that people who have really fragile mental states get exploited accidentally."
And while for Sam this may all feel consistent, to people who are just casually getting this, they're basically hearing one month Sam ripping on Elon for having what Altman characterizes as Japanese anime sex bots, and then a month later seeming to say, yeah, that's fine here too.
Even if obviously Sam and OpenAI would argue that there is a big difference between releasing a full adult mode and actually building these characters as Grok has done, versus just taking a more hands-off approach to whatever adults are going to do.
The other reason that this is hitting more skepticism, however, is that it comes on the heels of the Sora app.
We talked extensively about this when Sora was launched, about the consumer backlash to people feeling like OpenAI was shifting their focus from these big, lofty ambitions of changing the world and curing all diseases to instead just leeching our attention.
To some, it felt like the only reason for Sora the app, not the model, was to again lock our attention in in order to harvest us for data for further training or to make money in the form of ads that could go to support other things.
Now, as I argued, I think that that was not just a chat GPT critique.
I think that basically the challenge and the difference was that OpenAI was saying, hey, we're going to try to do social media and specifically short-form video responsibly.
And it just happened to come into a world where there is a large and growing contingent of adults who think that there is no way to do short-form attention-harvesting video responsibly.
That is a fundamental contradiction that has no room for compromise.
Around the time that Sora launched, former Stability founder Ahmad Mostak wrote, In an age of infinite content, human attention is one of the few finite things.
So they'll try to capture as much of it as possible.
Attention is all they need.
I've seen the echoes of many of those conversations that we had just a couple of weeks ago come back up now, as expressed in this tweet from Drew Harwell of the Washington Post.
Sam Altman went from AI will cure cancer to chat GPT porn in less than a month.
Nate Silver also wrote, OpenAI's recent actions don't seem to be consistent with a company that believes AGI is right around the corner.
If you think the singularity is happening in six to 24 months, you preserve brand prestige to draw a more sympathetic reaction from regulators and attract and retain the best talent, rather than getting into erotica for verified adults.
Instead, they're loosening guardrails in a way that will probably raise more revenues and might attract more capital and or justify current valuations.
They might still be an extremely valuable company as the new meta Google, etc., but feels more like AI as normal technology.
Now, I think that there is actually a really interesting question at the core of this that gets at one of the uncomfortable realities of this new world that we're living in.
ChatGPT currently has around one-tenth of the world's population logging into it each week.
When you operate at that scale, at what point do you stop seeing your role as trying to guide people to what you view as acceptable use cases versus effectively seeing yourself as a public utility where you have to basically allow people to do what they want?
I don't think that's as clean or clear-cut an answer as some people are making it out to be.
And my guess is that Altman and many at OpenAI see themselves to some extent
as Sherpas and stewards for allowing people to figure out what they want to do with AI without trying to be the arbiter of moral right.
I actually quite understand that point of view, especially as someone who has fairly socially libertarian views when it comes to adults being able to do whatever they want as long as they're not harming others.
At the same time, OpenAI still is a company.
In fact, it's fighting a lot of battles right now to become more of a company and less of a nonprofit.
And I do believe that as such, it gets to decide what it wants to care about and what it wants to stand for.
In other words, there are already ways.
for people to go have access to those adult experiences with AI should they so choose.
ChatGPT does not have to be one of the places they can do that.
It will not fundamentally hamper people's ability to get those experiences should they want them if ChatGPT said, we think that's fine for adults to do, but it's not for us.
I can't say for sure.
It's easy to be a backseat driver, but I think that if I'm in Sam's seat right now, that's the choice that I'm making.
Now, hold aside OpenAI itself.
There is a much larger societal conversation that this is digging up and that this is a part of.
The response from a number of corners has been loud and negative.
Mark Cuban wrote, this is going to backfire hard.
No parent is going to trust that their kids can't get through your age gating.
They will just push their kids to every other LLM.
Why take the risk?
Same for schools.
Why take the risk?
A few seniors in high school are 18 and decide it would be fun to show the hardcore erotica they created to the 14 year olds.
What could go wrong?
By the way, I can say as a parent, and I'm sure many of you have this experience, what you worry about isn't just what your kids do, it's exactly the scenario that Cuban is describing here, what they're going to learn on the bus or at a sleepover, because some other kid decided to go find some seedy corner of the internet.
Now, Cuban went on to clarify, this isn't about porn.
That's everywhere, including here.
This is about the connection that can happen and go into who knows what direction with some kid who used their older sibling's login.
And so what he's getting at is that this is not just about a new, different way to access pornography.
It's about this new relationship pattern, which we've seen indications, like in the 4-0 rebellion, that we don't really understand yet, and that could have some pretty negative consequences.
Cuban decided to clarify even further.
He writes, I'm not saying we should ban it.
I said it was a mistake, that it will hurt the business of OpenAI.
And I'll say it again, this is not about porn.
This is about kids developing relationships with an LLM that could take them in any number of very personal directions.
Parents today are afraid of books in libraries that kids don't read.
They ain't seen nothing yet.
I don't see how OpenAI can age-gate successfully enough.
I'm also not sure that it can't psychologically damage young adults.
We just don't know yet how addictive LLMs can be, which in my opinion means that parents in schools that would otherwise want to use ChatGPT because of its current ubiquity will decide not to use it.
It will be an ongoing battle for open AI.
I don't see the upside for them.
What's interesting about Cuban's take is that he's taking it both from the standpoint of parent and society, but also cost-benefit analysis as an entrepreneur.
Adding credence to his argument is this tweet from Vivek Ramaswamy, former Republican candidate for president and now candidate for governor of Ohio, who wrote...
the unnecessary overhumanization of AI is becoming troubling.
This new feature will do nothing to improve productivity or prosperity, but it will almost certainly increase addiction and loneliness.
I don't think government intervention will make it any better, but designing AI with the specific capability to sexually or emotionally manipulate humans warrants extreme caution.
Okay, so we have a conversation around what this means for ChatGPT and OpenAI from both a moral and a business perspective, and also a question about what it means for society and the nature of these relationships that we just don't understand yet.
But it is also so clearly a part of the growing contentious corpus of AI culture war battles that are rising to the mainstream.
This societal conversation about AI, loneliness, relationships is one of a number of cultural thought lines that we're seeing emerge that are going to shape the AI discourse in the years to come.
Some of these conversations are turning into political battles.
For example, this week, California Governor Newsom signed one bill adding guardrails to chatbot experiences, but vetoed another that had more restrictions on how kids could use them.
There's also the pop culture dimension of these battles around AI.
We haven't really covered it all that much here, but there's been a lot of scuttlebutt in Hollywood after the announcement of an AI actress named Tilly Norwood being signed by a talent agency.
In fact, this has been an ongoing conversation for like three weeks now.
Still, in my estimation, none of this pales in comparison to the core economic conversations, with the very obvious one that's emerging being focused on electricity.
As people see the full ambition of the AI industry revealed when it comes to data centers, they're drawing the natural conclusions that we're going to need a lot more power to power all those data centers.
And this is becoming a political conversation with much more resonance than the silly how much water does AI use per prompt conversation, which never really was able to pick up much resonance because of both how abstract and absurd it was.
This is a different kettle of fish entirely.
Take for example this tweet from Nick Huber.
AI is going to go down as a disaster of colossal scale.
My electricity bill in Athens, Georgia is up 60% since 2023.
Six increases in the last 24 months.
Just approved 20 plus data centers under construction in the region.
What gives?
Quality of life is dropping for 99% of people.
Now, we will probably do an electricity show at some point, and my guess is that you'll grok pretty intuitively that electricity being up in Athens, Georgia 60% since 2023 is not directly tied to AI.
Nick might not even be saying that it is either, but what he's recognizing is that the approval of data centers certainly seems like it's going to take that problem, whatever its cause is, and make it worse.
This backlash is happening on an increasingly national level.
You're starting to see protests around new data center construction and even plans being changed.
Now, it seems to me that there is a lot of room for answers from the companies here.
I'm actually gobsmacked at how little any data center construction firms or the hyperscalers themselves are engaging directly to design full-on community solutions that sit around what they're building to make these things not just something that people don't protest, but something that people see as actually additive to their community.
The number of dollars that we are talking about is so absolutely enormous that adding incremental costs on getting the buy-in of the local people who are going to potentially benefit from the positive impacts, but also definitely going to deal with the negative externalities just seems incredibly obvious to me.
There is some conversation starting to happen about this.
Chamath Palihapitiya writes, One simple solution for hyperscalers is as follows.
Option A, agree to a higher base rate with the utility so that you can guarantee people in the local geography won't see increased electricity rates.
Option B, agree to pay for residential solar and storage for local citizens so they won't see increased electricity rates.
Either way, if the hyperscalers don't use their gobs of free cash flow to cushion the inflation of electricity rates, you should expect to see a lot more pushback.
And so yes, of course, this is a totally separate conversation in some ways from the adult porn conversation.
I think it actually starts to blur in ways that could be very insidious for companies.
As High Yield Harry on Twitter put it, sure, my electricity bill is up 40%, but check out all these cool AI videos of SpongeBob evading arrest.
Anyways, guys, interesting times as always in AI land.
Appreciate you listening or watching as always.
And until next time, peace.
Today on the AI Daily Brief, these are the roles that people actually want AI to automate.
Before then in the headlines, Google is now processing 1.3 quadrillion tokens each month.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick notes before we dive in.
First of all, thank you to today's sponsors, Gemini, Robots and Pencils, Notion, InsightWise, and Superintelligent.
To get an ad-free version of the show, go to patreon.com slash AI Daily Brief, or you can subscribe on Apple Podcasts.
And if you are interested in sponsoring the show, send us a note at sponsors at AIDailyBrief.ai.
Right now, there are still some good slots open in Q1, but I have a feeling that that is going to turn very quickly.
So if you are at all interested, and even if you just want some information, again, shoot us a note at sponsors at AIDailyBrief.ai.
Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes.
We kick off today with an update from Google, where that company is now pumping out 1.3 quadrillion tokens a month to serve their AI products.
You might remember back earlier this year, when between May and July, we saw this massive inflection point
where Google went from processing 480 trillion tokens in May all the way up to 980 trillion towards the end of July.
That was monthly tokens, by the way.
So about 104% growth in just a couple of months.
My speculation was this was due in part to the expansion of actual deployment use cases, particularly around AI coding, that was just consuming a huge amount of additional tokens.
But whatever was driving it, it's clear that usage of AI is going up, up, up.
Now, Google DeepMind CEO Demis Hassabis recognized that at this scale, the numbers are frankly getting a little bit difficult to comprehend.
A quadrillion has 15 zeros in it.
And to reframe that 1.3 quadrillion number, he says that's 500 million tokens a second or 1.8 trillion tokens an hour.
Now, this is not the only indication that Gemini is undergoing some serious growth right now.
The latest edition of SimilarWeb's traffic report showed that Gemini was by far and away the big leader for AI platform growth in September.
The Gemini web app saw a 46% jump in traffic, which was more than triple the increase for Perplexity, which was in second place with a little over a 14% jump.
That report, by the way, also noted that DeepSeq notched its first month of growth since February.
And also that traffic to the Grok web app was the only one that had dropped, falling by 7.4%.
Now, similar web stats are a very light touch metric and not something that we should overly index on.
This looks only at traffic to web apps, doesn't really reflect usage at all on mobile apps.
And for something like Grok, it gets most of its use through the X platform.
So we actually don't know what the overall usage of Grok looked like last month versus in August.
But still, with all that said, what's undeniable from the report is that Gemini is growing at a tremendous rate.
Between that and it overcoming ChatGPT in the App Store for a time before Sora kicked it all back up, the race between ChatGPT and Gemini keeps getting tighter and tighter.
Plus, as the AI for Success account points out, I wonder what will happen when they release Gemini 3.0 Flash and Gemini 3.0 Pro in a few weeks.
Next up today, if you thought the risk of Mark Zuckerberg poaching your big talent was over, think again.
Meta has poached another very high-profile AI researcher to add to their superintelligence lab.
The Wall Street Journal reports that no less than a founding member of Thinking Machines Labs, Andrew Tulloch, has left to join Meta, informing co-workers of his decision on Friday.
Tulloch left OpenAI in 2024 to found TML with Meera Muradi and several other departing OpenAI leaders.
Prior to joining OpenAI in 2023, he had spent a decade at Meta as a machine learning engineer.
Confirming his resignation on Saturday, a spokesperson for TML said, Andrew has decided to pursue a different path for personal reasons.
And according to rumors, it sounds like it may have been well over a billion reasons.
Back in August, the Wall Street Journal reported that Tulloch had turned down a six-year $1.5 billion offer from Meta.
The story went viral, serving as the first solid reporting that Zuckerberg was personally recruiting AI researchers and offering 10-figure deals to top talent.
Tulloch was in fact the poster boy for the billion-dollar talent war that captured the narrative over the summer.
Now at the time, Meta said that the description of a billion-dollar offer was inaccurate and ridiculous, adding that any compensation package was predicated on Meta's stock rising, which frankly is a little bit of a non-denial regarding the maximum size of the comp package.
Overall, that article had been focused on an overall buyout offer to Thinking Machines Lab, which was turned down.
The tone emphasized that none of the leading researchers at the startup had accepted Meta's offer.
Reportedly, more than a dozen TML researchers were contacted by Zuckerberg over the summer.
Now there are a million different lines of speculation out there.
the rumor that is flying around X is that this was a $3.5 billion offer.
I don't know where that got started.
I've seen no evidence to support it.
The other, and to me, more compelling consideration that some are sharing is that this might reflect some amount of an assessment of the widening gap between the available resources in the sector.
Despite being one of the most well-funded early stage startups in the history of Silicon Valley, the resources available to TML in terms of compute and infrastructure are a tiny sliver of what is available to Meta.
Ultimately, we don't know if that's the reason or personal compensation or something else entirely is the reason that these moves have happened.
But as many people are pointing out, Llama 5 better deliver.
Moving back over to Elon's world for a minute, XAI is joining the race to develop world models.
The Financial Times reports that XAI hired a pair of researchers away from Nvidia over the summer to work on the technology.
NVIDIA, through its Omniverse platform, has been one of the leaders in practical world models used to train embodied AI in simulated environments.
Google and Fei-Fei Li's World Labs have also made significant progress, though their demos have been more focused on generating interactive video.
Through Tesla's cars and robots, XAI could have an opportunity to pair world models with actual embodied AI.
But that said, Elon Musk appears to be thinking about a different application as well.
Posting last week, the XAI Game Studio will release a great AI-generated game before the end of the year.
XAI is currently hiring technical staff for an Omni team, which quote, creates magical AI experiences beyond text, enabling understanding and generation of content across various modalities, including image, video, and audio.
Among the roles is a video game tutor who will teach Grok to produce video games.
The goal, it says, is to allow users to explore AI-assisted game design.
Some think this is a clever short-term play from Elon.
Phil Truby writes, Classic Elon strategy.
World models are proving to be needed for robots like Optimus, but Optimus revenue is years away.
However, world models can also be used sooner for AI-native video games.
Thus, Elon is creating near-term revenue for this otherwise long-term technology.
Now, one of the things always lurking behind people's minds is will there come a point where Elon decides that it makes sense to try to fold everything altogether under, for example, the banner of Tesla?
In that light, could this be a medium-term play to create a narrative that Tesla should buy out XAI?
Remains to be seen, but regardless, it is super interesting that XAI is jumping into the world model space as well.
Lastly today, escalation in the chip war as China cracks down on Nvidia imports and the Dutch government seizes a Chinese chipmaker.
If you're paying attention to the broader market at all, you will not need me to tell you that trade war tensions hit a fever pitch this weekend in the lead up to talks between the Trump administration and Beijing.
AI chips were just one front in the all-encompassing trade war.
On Friday, the Financial Times broke news that Chinese authorities had begun a crackdown on firms importing Nvidia chips.
They wrote that customs officers have been mobilized at major ports, searching for H20 and RTX Pro 6000D chips that are designed to meet U.S.
export controls.
One source told the FT that Chinese authorities were also looking for more advanced chips that were smuggled into the country in breach of U.S.
policy.
In the West, we had heard that Beijing had discouraged quote-unquote firms from importing Nvidia chips, but it seems that that was a little more than a suggestion.
Alongside cargo searches, officials are also poring over documentation to see if firms made false declarations about importing Nvidia chips in the past.
In a strange twist, Beijing now appears to be far more concerned about stopping the flow of advanced AI chips than even the biggest China hawks in Washington.
Then, breaking overnight on Sunday, the Dutch government has seized control of a Chinese-owned chipmaker.
Nexperia is a Dutch subsidiary of Wingtech Technology, which specializes in the production of high-volume, low-end chips for automotive and consumer electronics.
On Sunday evening local time, the Dutch Minister of Economic Affairs revealed that the Goods Availability Act had been invoked in September to seize the company, the first time that that 1952 law had ever been used.
He said that the move was made in order to, quote, "...prevent the situation in which the goods produced by Nexperia, finished and semi-finished products, would become unavailable in an emergency."
A government statement said that the highly exceptional decision had been made after the ministry observed "...recent and acute signals of serious governance shortcomings and actions."
WingTech responded in a now-deleted WeChat post, The Dutch government's decision to freeze Nexperia's global operations under the pretext of national security constitutes excessive intervention driven by geopolitical bias rather than a fact-based risk assessment.
Endgame Macro writes, What's happening with Nexperia goes way beyond a simple regulatory move by the Dutch government.
This is a frontline moment in the global tech power struggle between the West and China.
On paper, the Netherlands says it's stepping in because of administrative shortcomings and national security risks, but in reality, this is about cutting off one of China's quiet back doors into Western chip technology.
Nexperia may be based in Europe, but it's owned by China's wing tech, and the fear is that valuable know-how could end up back in Chinese hands.
Whatever the case, it is a major escalation, and just shows how many dimensions to this crazy AI story there are right now.
That, however, is going to do it for today's AI Daily Brief Headlines edition.
Next up, the main episode.
Small, nimble teams beat bloated consulting every time.
Robots & Pencils partners with organizations on intelligent, cloud-native systems powered by AI.
They cover human needs, design AI solutions, and cut through complexity to deliver meaningful impact without the layers of bureaucracy.
As an AWS-certified partner, Robots & Pencils combines the reach of a large firm with the focus of a trusted partner.
With teams across the U.S., Canada, Europe, and Latin America, clients gain local expertise and global scale.
As AI evolves, they ensure you keep peace with change.
And that means faster results, measurable outcomes, and a partnership built to last.
The right partner makes progress inevitable.
Partner with Robots and Pencils at robotsandpencils.com slash AI Daily Brief.
Chatbots are great, but they can only take you so far.
I've recently been testing Notion's new AI agents, and they are a very different type of experience.
These are agents that actually complete entire workflows for you in your style.
And best of all, they work in a channel that you already know and love because they are purpose-built Notion superusers.
Notion's new AI agents completely expands the range of what Notion can do.
It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days and get them complete in minutes.
These agents don't just help with work, they finish it.
Getting started with building on Notion is easier than ever.
Notion agents are now your very own super user to help you onboard in minutes.
Your AI teammates are ready to work.
Try Notion AI for free at the link in our show notes.
As a consultant, responding to proposals can often feel like playing tennis against a wall.
You're serving against yourself trying to guess what the client really wants.
That all changes with the InsightWise proposals platform.
Now you've got an AI coach that thinks just like your client.
It returns to the brief time and time again, identifying opportunities, showcasing your track record, and making recommendations to improve your pitch.
Suddenly you're on center court, but this time you've got a secret weapon.
InsightWise does away with all the time-consuming manual work so you can focus on winning more business more often.
Generate reports, pull insights from your own data, build competitive advantage, and go to sleep before 2 a.m.
When it comes to proposals, you only get one shot.
With InsightWise, make yours an ace.
Go to insightwise.ai to start your free trial today.
Today's episode is brought to you by Superintelligent.
Now, for those of you who don't know, who are new here maybe, Superintelligent is actually my company.
We started it because every single company we talk to, all the enterprises out there, are trying to figure out what AI can do for them, but most of the advice is super generic, not specific to your company.
So what we do is we map your AI and agent opportunities by deploying voice agents to interview your teams about how work works now and how your people would like it to work in the future.
The result is an AI action map with high potential ROI use cases and specific change management needs, basically everything you need to go actually deliver AI value.
Go to bsuper.ai to learn more.
Welcome back to the AI Daily Brief.
As time goes on and people get more acclimated to AI simply being a part of the society that we live in, the discourse has naturally shifted from high-level binaries and generics like, will AI take our jobs, or even silly little aphorisms like, your job won't be taken by AI but by someone using AI, into deeper analysis around where AI is actually useful, how it's evolving, and importantly, more recently, where people want AI to be involved.
You might remember this study that we covered earlier this year from Stanford that divided tasks into four different zones based on how good AI was at doing those things and how much workers wanted them to do those things.
There was a green light zone, which was things that AI was good at and that people were very excited for AI to do.
A red light zone, which was things that AI was good at but that workers didn't want AI to do.
A yellow zone, which was things that people wanted AI to do, but where capabilities were a little bit low.
And then a low priority zone, which was where AI couldn't do things or they were particularly hard and where people didn't really want AI to do those things.
Now, this was super interesting to me.
It's become a part of basically every keynote that I give.
Because again, it gets out of these binaries and starts to get into actual expressed human preferences from people on the ground who don't really have time for all these big philosophical debates.
They're just trying to figure out how AI is or isn't going to be useful in their own jobs and what it's going to mean for their careers moving forward.
Well, now we have something that almost forms an interesting companion study.
Earlier this month, researchers from Harvard Business School took a fresh look at AI job replacement.
Now, instead of coming from the angle of how capable AI was at performing certain tasks or trying to figure out how many jobs AI would replace, they instead asked people broadly how they feel about AI stepping in for humans in various occupations.
In the abstract, the researchers wrote, "...despite cultural anxiety about artificial intelligence displacing human workers, we find that Americans show surprising willingness to cede most occupations to machines.
Given current AI capabilities, the public already supports automating 30% of occupations.
When AI is described as outperforming humans at lower cost, support for automation nearly doubles to 58% of occupations."
In other words, where AI is competent and cheap, in many areas there isn't all that much moral objection to AI replacing humans.
However, the researchers continued that there are a narrow subset representing around 12% of occupations that include things like caregiving, therapy, and spiritual leadership that remain categorically off-limits because automation in those areas is seen as, in their words, morally repugnant.
They write, this shift reveals that for most occupations, resistance to AI is rooted in performance concerns that fade as AI capabilities improve, rather than principled objections about what work must remain human.
Now, the chart that got shared all over the internet was this one, another four quadrant chart.
On the x-axis, we have technical feasibility, i.e.
the percentage of tasks that are exposed to AI.
And on the y-axis, we have what they call moral repugnance towards AI.
The four quadrants then are, in this case, the green quadrant is in the lower right.
That's low moral repugnance towards AI and high capability.
This is their no friction quadrant.
Above that, in the capable but repugnant, which is sort of like their yellow light category, which they call moral friction.
That's where there's high AI capability, but also high moral repugnance towards AI.
Their red quadrant is where there is dual friction, low capability, and high repugnance.
And their blue quadrant, which is sort of like the opportunity quadrant of the other study that we just saw, is called technical friction, where there is moral permissibility, but low capability.
So let's talk about some of the types of jobs that are in each of these areas.
And let's do it in that order.
On the no friction side, where there is high capability and low repugnance, there are a lot of white collar jobs.
Search market strategists, financial quantitative analysts, economists, special effects artists.
All of those are in the green quadrant.
Now, again, I will remind you that while the other study is about what workers in those areas thought, this is about what the broader public thinks.
In other words, this is the broader public looking in on other people's jobs and saying whether they're fine with those jobs being automated.
It's not people self-assessing.
In the next quadrant, the moral friction quadrant, where AI is capable, but there is higher moral repugnance, some of the call-out examples include sociologists, history teachers, fraud examiners, OBGYNs, legislators, and school psychologists.
Basically, even if ChatGPT can theoretically give advice to students, that's not really something that people are super stoked on.
In the dual friction category, where there is both low capability from AI and high moral repugnance, they have nuclear technicians, oral surgeons, bailiffs, and nannies.
Apparently, even if we solve the problems of humanoid robots, people aren't willing to give their kids over to them just quite yet.
And then over in the blue area, which again is sort of an opportunity area of technical friction where there is moral permissibility but low capability, they have things like semiconductor technicians, cashiers, mail sorters, gambling dealers, and conveyor operators.
Of course, if you are a startup who is thinking about areas where you could vertically design AI solutions without people being mad at you, that blue area might be a place to look.
Now, Eric Brynjolfsson, one of the authors of the earlier task-based study, saw this new one and made the comparison directly, saying, "...it's interesting to compare their chart with what we found when we asked the workers themselves what they want."
Matt Bean from MIT Sloan actually ran the two studies through ChatGBT to compare, with Brynjolfsson of the Stanford Digital Economy Lab then synthesizing the analysis into a table.
This new four-quadrant chart had on the x-axis, worker automation desire, and on the y-axis, public moral acceptability.
So for this then green light, which was high moral acceptability and higher worker automation desire, that was things like scheduling and reminders, payroll error fixes, records upkeeping, standardized reporting, and database maintenance.
It will not surprise you at all, especially if you listen to yesterday's episode about where we're starting to see AI deployed.
These are in these areas where there's high value to getting it automated and people very much not protective of those areas.
The next quadrant, where there is public moral acceptability, they called augment carefully or co-pilot by default.
In other words, a human using an AI is probably the way to go versus handing it over to an agent.
That included things like assign and allocate stories, film editing and cuts, graphic layouts, and find a unique fact research.
I think this is a really revealing quadrant because if you take a step back, one of the things that these studies are telling us is that workers have a higher threshold for what they want in their job automated as opposed to people outside their job.
People outside of their job ask whether certain tasks within their job are okay to automate, basically are in many cases saying, sure, why not?
Even though the people who are doing those jobs say that there's something important or distinct about the human touch that they want to preserve.
I think that's why you're seeing things like film editing and cuts, where the people who are doing that understand what makes the difference between a really great version of that and an only okay version of that, whereas from the public at large who doesn't know about that craft, doesn't necessarily see it as craft, they just see it as something to get done, and version A done by a human craftsperson versus version B done by a robot doesn't particularly matter to them.
One of the real interesting challenges that we will face as a society is how to navigate the lines between what the people on the front lines who are doing a particular job think and where broader public sentiment is.
That's that quadrant that's going to see the most of that particular question.
Now, over on the other side, where worker automation desire is high but moral acceptability is low is kind of the inverse of that, where workers who are on the front lines think that there's more room to automate than people who are looking from outside who find it morally repugnant.
This quadrant they call assistive only and included by way of example care and therapy intake summaries.
Someone pointed out that this is actually one of the areas where the Harvard study shows its limitations.
Matt underscore amp on Twitter wrote, Eesh, caregiving is where automation can make the most difference if deployed appropriately.
No more elder neglect while warehoused in care homes administered by underpaid, overworked staff.
And to try to interpret this a little bit, the broader public is saying, absolutely not.
We should not have AI taking care of sick or elderly people.
That is a job that is for humans.
It is distinctly of humans.
We should have humans doing that.
That's the lens through which they're interpreting it.
However, what this combined chart is showing is that the people who are in that role understand that there are parts of this that are absolutely and incredibly valuable to automate and
As Matt points out, many of the facilities where this type of caregiving happens are plagued by the problems of, as he puts it, underpaid overworked staff.
To the extent that AI can take off big chunks of, for example, administrative work,
that allows them to just stay focused on the already emotionally taxing parts of human caregiving, there are probably really big benefits to be had from that.
And this is, of course, why it's going to be so important to not stay on the role level analysis, but actually get into task level analysis.
In many ways, I think that the best way to look at AI-related job displacement is from an additive task kind of level.
In other words, you take it from the task level, can AI and should AI automate a particular task?
And then from there, you look at what percentage of a particular role, as it's currently constituted, is tasks that can be automated.
So instead of saying, we're trying to automate role X,
You instead say, automation can do 70% of the work of that role.
And then we get to ask at what threshold that role needs to change.
Does the role stay the same, but there's just fewer of those people?
Is that role rolled into another role where the role's objectives are fundamentally changed based on the new capabilities that AI offers?
That's the sort of nuanced change that I think is going to happen much more than just job gone, see you later, which is of course the popular media kind of view, which we'll get into in just a minute.
The last area on the combined chart is where worker automation desire is low and public moral acceptability is low, which the AI summarization calls defer study or govern pilots, but which I think a lot of people are fine pretty much leaving off of the focus area for AI right now.
This is things like final hiring and firing, parole and probation risk calls, and ethics reviews and IRB-style oversight.
I think that both of these studies on their own are really valuable, but I think taken together, they represent something entirely different.
This is actually the beginning of a map of where society thinks AI can and should be valuable and where AI should be deployed to help move things forward.
Now, take that analysis as compared to another recent study that got a bunch of attention last week.
That was a Senate report that found 100 million U.S.
jobs could be replaced over the next decade.
The report was conducted by Democrat staffers on the Senate Health, Education, Labor and Pensions Committee.
Staff reviewed economic data, investor transcripts, and corporate financial filings to come up with this number.
However, the main source of data was ChatGPT itself.
The chatbot told staffers that AI and automation could replace nearly 100 million jobs over the next 10 years.
That includes displacing 89% of fast food and counter workers, 64% of accountants, and 47% of truck drivers.
Across the 20 workforces that ChatGPT said would be most affected, it said that 15 of them would see half of jobs displaced by AI and automation.
Now, 100 million jobs would obviously be a catastrophic number, well over half of the current 170 million strong US workforce.
Staffers did acknowledge that the methodology was a little questionable, writing, The reality is no one knows exactly what will happen.
There is tremendous uncertainty about the real capabilities of AI and automation, their effects on the rest of the economy, and how governments and markets will respond.
While this basic analysis reflects all the inherent limitations of ChachiBT, it represents one potential future in which corporations decide to aggressively push forward with artificial labor.
The report also noted that this change is far more rapid than previous economic disruptions, giving a greater sense of urgency.
Staffers wrote, "...the agricultural revolution unfolded over thousands of years.
The industrial revolution took more than a century.
Artificial labor could reshape the economy in less than a decade."
Now, the point of the report ultimately was not to generate an accurate number of jobs under threat.
It was to stir up conversation and provoke a policy response to the looming issue, something that many AI leaders have called for as well.
The report recommended adopting a 32-hour workweek, increasing worker protections, a $17 minimum wage, and elimination of tax breaks for companies that automate their workforce.
In an accompanying op-ed in Fox News last week, Senator Sanders argued that, quote, the rapid developments in AI will likely have a profoundly dehumanizing impact on us all.
We do not simply need a more efficient society.
We need a world where people live healthier, happier, and more fulfilling lives.
I've said before in something that might surprise some people that I've actually appreciated over time Senator Sanders' approach to this particular conversation.
And the reason is spelled out right here in the first line in this essay.
He writes, everybody agrees that AI and robotics are going to have a transformative impact on our country and the world.
And yet I've seen in the past how when it comes to a new technology, the tendency for the side that doesn't like the technology is actually to try to strangle it in its crib before it gets out and impacts the world.
Now, it may seem obvious that AI is beyond that stage, but I don't mind someone like Bernie Sanders taking the position that AI is here, it's real, and trying to bring up this conversation around what the new social contract in the context of AI looks like.
I don't agree with a lot of the foundational arguments that he has about the motivations for why this technology is being created, and I don't agree with a lot of the remediations that he's suggesting.
But this conversation that presumes that AI is here and that it will have an impact on real people's real lives in ways that are so significant that they could change the shape of the economy in ways that demand a new social contract conversation is something that I agree with.
We've forgotten this recently, but the foundation of democratic society isn't everyone agreeing.
It's everyone being able to have good faith conversations that start from some shared consensus about what reality is.
So TLDR, I don't think we're likely to see 100 million jobs ripped away, but I don't mind the starting conversation being, should we nudge what we consider a full workweek down to 32 hours?
Now, one interesting article that I also noted from last week that I also think optimistically shows just how little we know about how this is all going to play out and why we can't make too many assumptions before we see it in the real world.
Back in May, a business services company called Housecall Pro surveyed 400 home service professionals in an attempt to figure out how AI adoption was playing out in blue-collar professions.
They were so struck by the level of adoption that they named the report the AI-Assisted Trades Pro, How the Field is Leading the Future of Work.
The survey found that 40% of these pros actively use AI, and 60% were using AI at least somewhat.
The pros were using AI for content as well as administrative tasks.
The pros reported saving an average of 3.2 hours a week using AI.
That's 160 hours for a year for professions that are largely small business or owner operated, enough to really move the needle.
When you are saving the equivalent of four full weeks of administrative work per year, that is unbelievably high impact.
Now, in that report, cleaning professionals were the most common users, while electric professionals were the most satisfied with AI.
Another big takeaway was that AI was not replacing blue collar workers at all, even though it was delivering huge time savings.
73% of the pros surveyed said that AI had not impacted their hiring rates.
Now, CNN recently covered the survey and tried to track down some of these AI-augmented plumbers and electricians.
They found Oak Creek Plumbing and Remodeling in Milwaukee, who now have 20 plumbers all using AI.
Company President Dan Calley said, "...it's definitely been worth the investment.
Some of our older guys have learned to ask ChatGPT the right questions, and they're kind of amazed with some of the answers it comes up with."
He noted that the AI boost is now showing up in on-the-ground troubleshooting just as much as behind-the-scenes admin.
He commented, "...it's affecting both sides of our company, out in the field and internally within our office."
Another company, Gulfshore Air Conditioning and Heating in Niceville, Florida, has implemented a fully AI bookings and request system.
Once the technician arrives, they use AI to diagnose the issue and pull up the relevant technical information in seconds.
The process used to mean sifting through multiple lengthy manuals searching for the right fix.
Gulfshore has also used AI to optimize their marketing campaigns, which caused a huge bump in revenue.
Surprisingly then, these trade professions have turned out to be a perfect testing ground for AI.
They require an immense library of technical knowledge, as well as having the experience to know the tricks of the trade.
Being able to access the entire internet and every technical manual ever produced isn't a replacement for decades of experience, but boy does it help that experience figure out what's actually going on much more quickly.
These trades also require a ton of tedious booking management and administrative support.
Going back to that original study, is work that most workers would happily automate away.
Laura Ulrich, an economist at Jobsite Indeed, commented, People go into the trades because they like doing the hands-on work itself.
And if some of the administrative tasks can be automated, then that should help those workers lean into the parts of the job they like and do smarter work.
Krista Lander, the marketing and IT manager for Gulfshore, commented, All of our technicians are running more efficiently and they're less stressed.
I feel like I'm a real-life Jetson living in the future.
Now, I am very wary on this show of being Pollyannish about the real challenge that the AI transition is going to represent.
As I've said before, and I will continue to say, I am extremely bullish on the long term.
I think that AI is going to unlock more creation, more business, just more of everything, including more jobs.
However, I think that the disruption along the way is going to be enormously painful.
And I do think beyond a shadow of a doubt, it is going to be significant enough that we have to have a conversation about a new social contract and what we expect from people to be full contributors to society in a world where AI can just do much more of the work.
What's encouraging to me is that between these studies and these real on the ground lived examples, we're starting to move beyond blithe, generic, could be theoretical future fan fiction type scenarios and actually understand what's happening in practice and what people think in aggregate and in specific.
That leaves us in a much better position to actually have the conversations we need to have to make this AI era our best one yet.
For now, though, that's going to do it for today's AI Daily Brief.
Appreciate you listening or watching as always.
And until next time, peace.
Today on the AI Daily Brief, what more than a thousand executives told us about AI agents.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
We're sold out for 2025 and about halfway through the beginning of 2026.
Send us a note and we can get you all the information you need.
Now with that, we turn to today's topic.
And since this is a long read slash big think type of weekend episode, even longer for many of you in the US, I thought it would be fun to go beyond the headlines and actually dig into the data around what we are hearing about AI and agents in the enterprise based on our conversations with thousands of executives.
Now, the context for this, for those of you who don't know, is my startup Superintelligent.
Superintelligent is effectively an AI business intelligence startup that focuses on AI and agent planning in the enterprise.
We use voice agents to totally transform the process of discovery, opportunity mapping, and figuring out what AI and agent opportunities are most pertinent for your organization and what you're going to need to do to get ready to take advantage of them.
Over the last six months, we have done thousands and thousands of these interviews, and this analysis represents a subsection of what we've learned.
We're going to talk about the most common challenges we see, what the biggest blockers are, as well as some of the interesting opportunities and what the biggest enablers are.
Hopefully, this is the type of presentation that can be extremely practical and useful for those of you who are inside businesses figuring out how to harness AI and agents.
And to kick us off, let me give you a few grounding statistics.
As part of this opportunity mapping, we curate what we call an agent readiness score.
It's on a scale of 100, and it's divided into quartiles, with the bottom quartile being called agent initiate, the next quartile being called agent explorer, the third quartile being agent pilot, and the fourth being agent ready.
Now, there's no shame in any of these quartiles.
They simply reflect different levels of preparation and organizational development when it comes to AI and agents more specifically.
Overall, the average readiness score was 52.1, which is in that agent pilot category.
And 58% of all the organizations we work with are at the low end of that agent pilot category.
This means pretty much exactly what you think it would mean.
There are probably some AI pilots and AI infrastructure there, but it's not necessarily super robust.
Maybe there's been some dabbling even with agents, but not necessarily a lot of full-scale deployments.
There are probably lingering issues around either governance or data or something else.
But there is also a foundation to be built upon.
That's the agent pilot stage.
Explorer, which represented 39% of organizations, is a step back from that.
Maybe there's even less infrastructure ready and even fewer pilots going on, but we're still not at a null state.
Initiate is really on that low end and Reforster organizations that are just getting up and running when it comes to anything AI and agents.
Again, no shame in that game.
There are plenty of organizations that fit in that mold.
One other interesting overview is that on average, we find that big organizations are a little bit farther along than small organizations.
Now you might think that this would be the opposite, given that smaller organizations feel like they are better able to dynamically adapt, but it turns out this is hard for everyone, and large organizations have just had a mandate from on high to get out ahead of this compared to others.
When it comes to the top use cases...
Once again, I don't think that if you are a regular listener here or if you spend any time on these issues inside your company, it will seem super surprising.
At the very top is enterprise knowledge search.
That came up as a recommendation in 48% of our audits.
There's just a ton of information locked inside corporations that, if it were more easily accessible, would help people inside those organizations do their job better.
Agent-assisted coding is, once again, very high on that list.
In fact, the only reason that I think it's as low as only showing up in 45% of the cases is that in addition to auditing full organizations, we frequently work with just specific lines of business or departments or even functions.
And on average, those have tended not to be coding organizations.
Some others that you might expect are showing up pretty frequently, customer service agents, sales support, back office reporting automation.
You can kind of get a feel for the type of distribution of use cases that are coming up.
It's pretty wide.
It's pretty cross-cutting.
Let's move now into the challenges that we found.
You always gotta start on the hard stuff and then come out to the opportunities, right?
One really interesting note is that in general, tech readiness has not been the issue.
And what I mean by that is that most organizations at this point, even smaller organizations,
have relatively modern, sophisticated technology platforms.
They're working with some of the major cloud providers.
They've been updated for new possibilities over the last five years.
Basically, the issues with AI are not, in general, the clunkiness of legacy platforms.
Where the issues do come in, and by far the number one blocker and the number one challenge across a huge cross-section of these audits, is data fragmentation.
And really, this could be data everything.
Fragmentation is an issue.
How usable the data is, even if it's not fragmented, is an issue.
Compatibility is an issue.
Data access is an issue.
We often see organizations that have spent a bunch of time trying to organize all their data, but still have huge issues around who can access what types of data.
This is particularly the case in highly regulated industries where there are really strict barriers between different people having different types of data access within the organization.
We also see it in the finance industry where different data sets are relevant for different types of investment or financial activities.
We talk a lot right now on this show about context engineering, context orchestration, context in general as the big unlocker of all the next opportunities of AI and agents.
I think this is going to be a massive theme for 2026.
And part of why I think so is that we just see this over and over and over again.
Even companies that have spent time on this challenge, no one is fully there.
It remains an issue.
And thinking about it as very foundational and important, I think is going to benefit organizations in the year to come.
Next up, we move to some of the perhaps more surprising findings.
One is that enthusiasm can inadvertently create resistance.
What I mean by this is that in many organizations, executives have jumped out into the lead when it comes to the AI opportunity.
They are pushing their people to go out and learn AI, to use AI tools, to integrate AI into their workflows, to design entirely new workflows around AI.
And even when those executives have good intentions, that rapid top-down push for multiple new AI tools and processes can create change fatigue.
It can create a feeling of overwhelm for employees who are just trying to do their job while also trying to adapt to this whole new reality.
Which brings us to a related and very important point, something that we see echoes of in far more than half of our audits.
Employees very frequently report some variant of being too busy to learn the thing that saves time.
This is one of the great paradoxes of AI inside the enterprise right now.
There is, in many cases, actually, not a skepticism to the idea that AI tools could make one's life better by making their work more efficient.
The challenge is that there is a learning curve.
There is a barrier to entry.
And unless an organization has structured the time to give people the time to go learn those new processes, it just becomes another category of work to do and another item on the to-do list that was already too long.
It is far more often employee bandwidth rather than budget that is a constraining factor to AI adoption right now.
And there is such a common mistake of executives providing excitement, providing the tools, but not providing mandated time to learn and experiment with those tools.
And it's that third missing element that becomes the big blocker.
Another challenge that comes up really frequently in well over half of our audits is some type of a policy awareness gap.
Now, the outcome of this, in other words, of people not knowing exactly what AI policies are, how they're supposed to use the tools or how they're not, what they are and aren't allowed to do, is either one, avoidance, or two, shadow usage.
In other words, using the tool without telling colleagues or managers.
Now, shadow AI isn't always bad a priori.
In some cases, it's just a chance for people to learn how to use new tools, sometimes on their own time and even on their own budget, and they bring that knowledge back into the workplace.
The problem is that if people are using these tools, and especially if they are bringing sensitive data to those tools outside of the enterprise ecosystem, there can be real challenges in that sort of usage.
What we find over and over and over again is that people are not using these tools externally because they specifically want to be breaking the rules.
It's because they don't really know what the rules are, and they're not exactly sure how to go find out.
Now, there are other reasons for shadow AI, probably the biggest one being that historically there has been a gap between the quality of the tools that are available to the enterprise user, as opposed to those that consumers get to use on their own.
More recently, that gap has started to close, thank goodness.
And now this policy awareness gap is one of the biggest drivers, at least that we see, of that shadow IT.
A couple other patterns that come up pretty frequently.
The first is that we see organizations stuck trying to figure out whether they're supposed to buy or build when actually in the context of AI, that is a false dichotomy.
There really is no such thing as off-the-shelf when it comes to AI, and especially when it comes to agents.
Even the most off-the-shelf agent is still going to involve some amount of customization, some amount of wiring into your existing systems.
And so getting out of the mindset of buying versus building and just thinking about an integrated process that's going to live somewhere in the middle there is really important.
Relatedly, one of the most interesting anti-patterns that we see, in other words, something where you would think it would be good for an organization but actually ends up really dragging them back, is an overabundance of a DIY mindset.
We will frequently see really strong resistance from the people that you might think would be the most tech-forward, IT.
And formerly, at the beginning of these surveys, although much less so now, engineering departments more broadly.
But whatever the reason for it is, a pride of purpose or a past experience or a sense that you know your systems best, this DIY mindset that we always have to build everything on our own, when we see this DIY mindset, those organizations tend to be less far along on their agent journey than those who don't have it.
Overall, unsurprisingly, across all of these interviews and all of these audits, co-pilots and assistants are increasingly table stakes.
They are everywhere.
They are proliferated.
There have even been some agent pilots and agent experiments.
But full agent platforms, in other words, thinking systematically about agents, is extremely, extremely rare.
Let's sum up some of the big blockers that we see.
Fragmented data, like I said, remains one of, if not the biggest problem for all organizations.
Even the organizations who are higher in agent readiness still deal with issues of fragmented, unstructured, unorganized data.
There tends not to be platforms.
Instead, we still live in pilot world.
There is a governance fog where people aren't exactly sure about the rules of using these tools.
There is the change fatigue that we talked about and consequently, skills gaps.
Over 70% of the organizations that we surveyed reported some issues with big skills gaps with their workforce when it came to AI.
I actually have a thesis that because there has been such a big emphasis on agents this year, automations, that it replaced entire categories of work or workflows, that the two-year journey towards upskilling people that had started right after ChatGBT and was proceeding bigly at the end of 2024...
got waylaid and kind of kicked back over to HR and learning and development, as opposed to being a front and center focus for the main organization.
I think one of the things that you'll see in 2026 is a recalibration where you're not going to have such a strict divide between augmented AI, i.e.
employees using tools, versus agentic AI, in other words, entire workflows being automated.
Organizations are going to try to do that all at once.
And to do so, one of the things they're going to need is better documentation around their processes.
This is the last big blocker that came up in something like 44 or 45% of the interviews that we've done.
Turns out it's very, very hard to automate workflows when those workflows live exclusively inside people's heads and are not articulated anywhere that an agent can read and learn.
Chatbots are great, but they can only take you so far.
I've recently been testing Notion's new AI agents, and they are a very different type of experience.
These are agents that actually complete entire workflows for you in your style.
And best of all, they work in a channel that you already know and love because they are purpose-built Notion superusers.
Notion's new AI agents completely expands the range of what Notion can do.
It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days and get them complete in minutes.
These agents don't just help with work, they finish it.
Getting started with building on Notion is easier than ever.
Notion agents are now your very own super user to help you onboard in minutes.
Your AI teammates are ready to work.
Try Notion AI for free at the link in our show notes.
What if AI wasn't just a buzzword, but a business imperative?
On You Can With AI, we take you inside the boardrooms and strategy sessions of the world's most forward-thinking enterprises.
Hosted by me, Nathaniel Whittemore, and powered by KPMG, this seven-part series delivers real-world insights from leaders who are scaling AI with purpose.
from aligning culture and leadership to building trust, data readiness, and deploying AI agents.
Whether you're a C-suite executive, strategist, or innovator, this podcast is your front row seat to the future of enterprise AI.
AI isn't a one-off project.
It's a partnership that has to evolve as the technology does.
Robots and Pencils work side-by-side with clients to bring practical AI into every phase.
Automation, personalization, decision support, and optimization.
They prove what works through applied experimentation and build systems that amplify human potential.
As an AWS-certified partner with global delivery centers, Robots and Pencils combines reach with high-touch service.
Where others hand off, they stay engaged.
Because partnership isn't a project plan.
It's a commitment.
As AI advances, so will their solutions.
That's long-term value.
Progress starts with the right partner.
Start with robots and pencils at robotsandpencils.com slash AI Daily Brief.
But let's shift over now to opportunities, first talking about some of the observations we saw from organizations that are doing well, or even where organizations are struggling, the parts that were bright spots.
One thing that came up pretty frequently is that AI is so high leverage that you can, in many cases, get massive ROI from a single individual.
A number of times across these thousands of interviews, we saw or heard examples where
where someone inside a particular function had figured out a new AI or agent enabled workflow for a core piece of what they did that was not only able to help them do their work better, but because there were a bunch of other people like them with a similar role, was able to transmit across the organization, leading to hundreds of thousands or even millions of dollars of benefit,
Again, just from the genesis of a single individual.
I think this speaks to the need to have good information sharing systems where people are actively encouraged not just to use AI, but to share their successes so they can be disseminated across the organization.
Here's one that people are sometimes surprised about.
There are so many flashier ways to use agents that it might surprise you that internal support bots are very frequently one of the use cases that really get internal teams, including skeptics, on board with AI.
This comes back to that idea that there is a ton of knowledge, information, data
all locked up in various silos and pockets across an organization.
And if you had better tools to access across that information, perhaps in an interlocutory chatbot sort of way, it could make a big difference to people's jobs in the here and now.
And that is exactly what we see.
The organizations that did some version of an internal support bot as one of their early use cases had real success putting that early win on the board, and it had a double ROI.
The first ROI was the benefit that it provided for people who were just trying to do their jobs and who were able to do so more effectively and more quickly.
But the other benefit was less resistance to other AI deployments that would come later.
Next, we have sort of the inverse of that anti-pattern we talked about before of the DIY mindset.
Zero prior automation is, in our experience, actually an advantage.
One of the things that has come out across these interviews is that if organizations had spent a bunch of time with previous approaches to automation, think RPA, they in many cases had to unlearn those systems, both in terms of the human knowledge unlearning, but also to rip them out and unstructure them.
Organizations that were leapfrogging any sort of automation 1.0 type of strategy like RPA had an advantage because they were able to go straight to these new UI UX patterns and new experience patterns that are different about AI and agents.
So if you haven't done any automation and you think that makes you farther behind, you actually could turn that to an advantage.
Now, if we see the psychological benefits to people and reduce resistance after internal support bots, the place that we are starting to see the first ROI in actual practical business and financial terms is in and around back office automation functions and around support roles.
So finance and support is where we're seeing the first ROI.
This does not mean, of course, that that is the right place for your organization to start.
But to the extent that you are trying to put early wins on the board, especially early wins where you can actually calculate the benefit, there's a lot to be said for looking into those categories.
have another inverse here if one of our big blockers was undocumented processes the reverse is obviously true as well where better documented processes mean faster pilots more results more quickly one of the things that's great about this too is that ai can help in addition to there being services like super intelligent that do at least some part of this and other services that go even more directly at having ai observe what people are doing and turn that into workflows that agents can understand you can also do this without any custom software at all
For example, have someone fire up a loom that just screen captures them as they go about doing some particular job or task that they do day in and day out.
When they're done, paste that loom into ChatGBT and A, have it turn it into a document explaining the process or B, even from there, start to make suggestions on where there might be efficiencies just in that human version of it.
Point is that there's no such thing as too documented when it comes to processes.
The better documented the process is, the faster you're going to start to see some value from the early implementations of AI.
Now, moving over into the governance idea, if we see the organizations that don't have clear governance struggling with shadow AI, there is a very clear pattern among organizations that are more successful with their governance, which we might call a sandbox with guardrails.
The idea here is pretty simple.
You can't be so restrictive with AI and agents that people can't go figure out how to get value out of these things.
There has to be some tolerance for and even encouragement of experimentation.
That's the sandbox part.
The guardrails are the systems you put around it to make sure that there aren't potentially deleterious effects and negative externalities or consequences that come out of that sandbox experimentation.
This doesn't have to be super complicated.
It could be as simple as saying, you can play around with this type of tool with this portion of our work and our data, but not this portion.
These data sets are excluded from that sandbox.
Whatever the specifics, we see this pattern in over two thirds of the organizations that have established governance regimes for AI and agents.
Now, once again, we did top blockers before.
Now let's talk about the top enablers.
Although they can be over-exuberant, and we talked about that before, committed execs still are one of the key enablers across all the organizations we had.
Without that, it is very hard to get momentum around an AI strategy.
A second top enabler, organizations that tended to score higher on agent readiness were much more likely to have some sort of AI task force or center for excellence or other centralized organization that could both absorb internal needs and disseminate potential AI strategies, that could serve as a resource as new tools were rolled out, that could think about the different experiments that were happening and try to see what should be scaled across the whole organization.
Number three, you've heard me talk a number of times about quick wins.
Like anything, AI is a momentum game.
When people see the benefit of it, they want to do more of it.
So looking for those quick wins where you can quickly realize either really tangible ROI for the organization as a whole, or just really clear benefits to individuals and how they work, those sort of things can build the momentum that you need for broader, more sophisticated, and more comprehensive strategies.
related to the task force is the idea of AI champions in training.
The training part is obvious.
We're way behind on this.
Organizations simply put just need to spend more resources on training their people.
Now, some of this is the market's fault.
I've complained numerous times about some of the deficits that I see in training products right now.
But if you really want your teams to be operating at highest function now, you have to give them resources to learn how to use these tools, even if that means some amount of customized design and consultation.
And one of the best strategies around training is to specifically elevate the people who are getting out ahead and getting farther than their peers when it comes to AI.
Out of sheer personal interest, almost every organization has some cohort of people who are naturally and natively and without being asked, spending their nights or weekends learning these tools.
Maybe they're learning them in a different context and they're not using them in the same way they would use them at work, but these folks are getting used to the patterns.
They're learning prompting.
They're learning how to organize context.
Organizations that score higher on the agent readiness exam tend to have some formalization of those people as, for example, AI champions who can be a resource for their peers as their peers get caught up to speed on AI as well.
Lastly, while we found in general that a solid API-driven core modern tech foundation is pretty prominent across all these organizations and that tended not to be the big blocker, for those small number of organizations that didn't have that, they were dead in the water.
So our assumption is that while most organizations are there, if you happen to be working off of really legacy clunky tech infrastructure, you got to change it.
There's just no way of getting around it.
Now, one really interesting and practical takeaway from all of this, the single biggest lift that we see, basically the single biggest factor that shows a difference in the average agent readiness score from an organization that has it to an organization that doesn't, is an established governance framework.
Organizations with established AI governance frameworks were 6.6% more agent-ready on average than those without.
And I think that this speaks to the idea that governance is not just about the rules of the road.
It's about creating safe space where people can experiment and build and try new things within their core workstreams rather than off on the side without people knowing.
So if you take away anything else from this, if you don't have a governance framework right now, you can probably make a big dent in how well AI is working in your organization by just focusing on that.
Now, the last thing I want to go through is sort of a summary of some of the archetypes we see of organizations in case these help you place yourself in context.
The first we'll call the visionary bottleneck.
These organizations are vision rich, but plumbing poor.
They tend to have strong executive intent.
They have modern SaaS systems, but they have weak data set up.
They're dealing with those data fragmentation issues.
The risks for this type of organization is change fatigue that we talked about, and practically pilot purgatory.
Lots and lots of things starting, but not ever adding up to something greater than the sum of its parts.
Next we have the cautious incumbent.
These organizations are often in a regulated or conservative industry, where they almost by definition prioritize governance, risk management, and security above all else.
These are almost the inverse of the last, where they have prioritized formal AI policies and a diligent review process, but to the detriment of just getting out and trying things.
In these organizations, their caution creates a culture of low trust and slow experimentation.
Employees are often wary of AI's accuracy to the point where they don't even try it, and their fear of compliance missteps holds them back from doing anything.
The risks here are analysis paralysis, competitive lag, and stifled innovation.
The next archetype we'll call grassroots tinkerers.
This is where there's broad, general support for AI experimentation, but in the absence of a central strategy.
So in this case, you see lots of people being encouraged to use GPTs or co-pilots, but without a real roadmap and without a lot of support in form of upskilling or long-term strategy.
The risks here are really inconsistent quality and ultimately being left behind as organizations more systematically bring on agentic workflows.
The grassroots tinkerers jumped out ahead when it came to some of the co-pilot type uses, but are falling behind in the realm of agents without that central strategy.
Lastly, and aspirationally, we have the foundation builder.
These are the organizations that take a deliberate infrastructure-first approach to AI.
They tend to have a strong central IT or data team that are focused on the plumbing of AI, consolidating data into a unified lake house, establishing a secure AI gateway, developing an enterprise agent platform that others can plug into.
The biggest risk here is that while they are extremely technologically sound, the approach can be slow to deliver tangible business value, which can sometimes create frustration among business units that are eager to just get those quick wins.
And so what you can see about all of these things is that there is no such thing as just purely negative.
In each of these archetypes, the organizations have particular strengths, but in many cases, vis-a-vis AI, those strengths are weaknesses and vice versa.
Understanding which of these archetypes or other archetypes your organization fits into might be a way to help you better identify where your remediations and next steps are best going to be.
Wrapping up, right now, I think that there are two big contenders for year of titles for 2026.
By this, I mean the big enterprise themes that we're going to see.
And it's not necessarily an either or.
In fact, I think it's going to be both and.
The two that I am seeing discussed most often and resonate most with me based on all of these interviews and surveys and also the conversations that we have is the idea that 2026 is going to be the year of context and the year of ROI.
ROI is easy to say and hard to figure out.
One of the things that comes up over and over and over again is that CIOs right now understand that ROI is important, but that it also does not fit into traditional frameworks from a pre-AI era.
Organizations are not only trying to measure ROI, they're trying to figure out how to measure ROI in the first place.
This is something that Superintelligent is thinking deeply about.
preview, we are about to launch a performance pulse product that helps organizations better track the results of their AI and agentic deployments, because I think this is just a complex, challenging area that every organization is going to be thinking about and needing help with.
But I will also say that one of the things that is going to lead organizations to be successful when it comes to ROI is better context.
I think that there is going to be a bunch of narrative cloud cover going into next year for organizations to peel back away from flashy agent pilots and instead see data foundation work as sexy, exciting, something that the organization should be really invested in and thrilled about.
Basically, whereas six months ago it might have been cooler to release some pilot agent than it was to get five MCP servers set up,
I think that flips going into next year.
And I think the organizations that really lean into that year of context and working on their data foundations are going to, by the end of the year, also have it have been the year of ROI.
In any case, that's going to do it for this particular episode.
I hope this was interesting.
Let me know if this type of ground-level feedback and data is useful.
For now, I hope you're having a great long weekend.
Appreciate you listening or watching, as always.
And until next time, peace.
Today on the AI Daily Brief, five reasons AI is a bubble and five it's not.
The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.
All right, friends, quick announcements before we dive in.
First of all, thank you to today's sponsors, Robots and Pencils, Notion, KPMG, and Super Intelligent.
To get an ad-free version of the show, go to patreon.com slash ai daily brief, or you can get it on Apple subscriptions.
To learn more about sponsoring the show or anything else about the show, speaking engagements, et cetera, go to ai daily brief.ai.
One of the things that I'm thinking about for next year are extensions to this core show.
Nothing is changing about this core show, but there is so much more to explore out and around it that I'm exploring three different ways to expand the AIDB universe.
The first is an enterprise edition that would basically take this show, repackage it, reorganize and restructure it as learning materials and discussion and strategic material for enterprise teams.
The second is an operator's cut, which would be additional supplemental material that was a little bit more practical and focused, stuff that would help you apply AI more directly.
And then the third, relevant for this particular episode, is an investor's edition, which would be a supplement more focused on big market themes.
I would love to know if any of these interest you.
So if you would be so kind, please go to ai-dailybrief.ai.
There's a three-question survey there, and that would be ever so appreciative if you would fill it out.
Lastly today, we are talking about the big behemoth topic, is AI a bubble?
And it got long and dominated the entire episode.
So no headlines today.
We'll be back with headlines tomorrow.
And without any further ado, let's dive in.
Welcome back to the AI Daily Brief.
Today, we are once again exploring a theme that has been with us since the beginning and will be with us for a long time to come, I think.
However, it has just gotten too loud, too unignorable.
We have to dig in again because this is all anyone, at least in the mainstream media, is talking about right now.
The question, of course, is AI bubble, and we are going to talk about this latest report from the Information and Oracle, which restarted this set of conversations, and look at a set of reasons why AI is or is not in a bubble right now.
But just for a little bit of context, and to really reinforce the point that this has been with us since the beginning, ChatGPT came out at the very end of November of 2022, and already by February 7th,
people like Josh Brown of Ritz-Holtz Wealth Management were calling an AI bubble.
Now, of course, we know what has happened since then, at least from a big market perspective.
With the small exception of the tariff tantrum earlier this year, it has been just up and to the right.
And in 2025 in particular, the sheer size of everything surrounding AI has just gotten absolutely enormous.
This has led over the last couple of months to a renewed discussion of the bubble conversation.
We had a big burst in this topic in August as we got the combined factors of initial disappointment with GPT-5 and the renewed arguments that we were hitting a wall, widely over-reported comments about Sam Altman saying there was a bubble, and of course that now infamous MIT study, which I put in air quotes, that argued that 95% of AI pilots were failing and which whipped around Wall Street with incredible frenetic energy.
Still, another interesting thing happened in early September when we got news of the $300 billion Oracle OpenAI deal that sent Oracle stock soaring by 35% in a day.
All of a sudden, people were seeing the financial consequences of missing out on this leg of the AI boom, and bubble talk quickly fell off the menu.
In fact, Deutsche Bank published a report in late September that declared that the first bubble to burst was in fact discussion of the AI bubble.
They found that the peak of the narrative was back on August 21st, with the number of web searches for AI Bubble falling by 85% over the following month.
One of the things they pointed out is that, quote, identifying a bubble is almost impossible, not least because no one agrees exactly what it is.
Typically, it's something like when asset prices rise significantly above intrinsic values, but they also don't agree what the correct intrinsic values are even after it bursts.
What's more, they point out, concern about a bubble may act as a pressure valve, lowering valuations and encouraging a whole new round of bargain hunting.
And yet, despite this report coming out on September 30th, just a little more than a week old, it already feels out of date.
Searches for AI Bubble have tripled over the past two weeks as a wave of new news reignited the narrative.
On Monday, we got news of OpenAI's deal with AMD, which was the subject of yesterday's show, which, based on its non-traditional structure of OpenAI potentially owning 10% of AMD if it hits milestones, felt very bubbly to some.
Then on Tuesday, the Information published an expose on Oracle's financials.
The piece was titled, Internal Oracle Data Show Financial Challenge of Renting Out Nvidia Chips.
They write, Internal documents show the fast-growing cloud business has had razor-thin gross profit margins in the past year or so, lower than what many equity analysts have estimated.
This could raise questions about whether the AI cloud expansion undertaken by Oracle and its rivals will affect profitability and sustain investors' expectations.
So the report states that in the three months that ended in August, Oracle generated around $900 million from renting out NVIDIA-powered AI compute.
Gross profit was $125 million, or 14%.
which, as the information points out, is lower than even many non-tech retail businesses.
For some comparison, Walmart's gross margin was 25% in the last fiscal year.
And then on the other end of the spectrum, closer to Oracle's business, Microsoft's Azure division currently operates at a 69% gross margin.
Oracle stock actually dove 6.9% on the news, but recovered much of the drawdown by the end of the day.
The narrative was absolute catnip for headline writers, with FX leaders going with, "'Oracle stock dives on weak cloud profits.
AI hype meets tough reality.'"
Still, some thought something felt off about the whole thing.
Amit is investing, posted on Twitter.
This doesn't seem like a demand issue, aka Oracle not finding customers for renting their GPUs, but rather seems like a margin issue.
It seems like customers are asking for a really good deal, or they would just leave to an Iran, Nebius, or Coreweave, and as a result, Oracle's margins on this business have not been as strong.
Some people questioned the financial acumen in the article.
Tay Kim, a veteran financial reporter and author of the NVIDIA Way, posted, Today I learned AI infrastructure biz has lower gross margins than software, and margins are bad when GPUs are getting installed and not yet generating customer revenue.
This is basically an argument that the information and then consequently Wall Street investors were turning a molehill into a mountain by not understanding the natural lifecycle of this particular type of product.
Jensen Huang commented on something similar in an interview with Jim Cramer on Tuesday, stating, What Oracle does with our systems is not easy.
We're talking about giant supercomputers.
When you first ramp up a new technology, there's every possibility that you might not make money in the beginning.
But over the life of the systems, Oracle is going to be wonderfully profitable.
In fact, differences of opinion on the depreciation schedule for AI chips is one of the big factors going into bubble analysis.
Most firms are depreciating their chips over five or six years before they need replacement.
Yet some of the loudest bears, like Jim Chanos, argue the depreciation schedule is much shorter, with data centers needing to upgrade to the latest and greatest from NVIDIA every two or three years to remain competitive.
JT of LaCoya Capital picked up on an interesting comment from the information article on this front.
The article stated, "...one silver lining in Oracle's GPU business is the amount of revenue it is generating from older generations of NVIDIA chips, such as the Ampere chips that came out in 2020.
Those chips appear to be helping Oracle's margins, while newer versions of NVIDIA chips strain them."
JT points out, hilariously, this is probably the most bullish data point on the actual versus theoretical useful life of NVIDIA Silicon I've seen anywhere, maybe ever.
They're still printing it on Ampere.
In other words, while the margin story shows how difficult it is to make money initially off of the state of the art, it is real lived evidence that chips from five years ago are still profitable, which has huge implications for how we think about the longevity of the value of this infrastructure that's being installed.
A huge argument from those who think we are in a bubble is that the infrastructure is going to go out of date even faster than people think, and that unlike railroad track from centuries past, or fiber optic cable from decades past, these chips are going to be useless in two or three years.
And this suggests, at least, that life cycle might be a little bit longer than we thought.
Now, the other big piece of news circulating on Tuesday that supercharged this bubble conversation once again was that XAI had raised $20 billion in an ongoing funding round.
The fundraising is apparently a mixture of debt and equity and will be used to fund the expansion of XAI's Colossus 2 data center that's currently under construction in Memphis.
The part that caught everyone's attention was that NVIDIA was investing $2 billion in the equity side of the round, a move that Bloomberg wrote was a, quote, strategy by the chipmaker that helps accelerate its customers' AI investments.
Okay, so this brings us to the main thrust of this piece.
Is this a bubble or not?
Let's talk about five reasons this is a bubble and five reasons it's not.
I think, by the way, I'm probably going to end up at six or seven reasons in each case.
And to be clear, as you can probably imagine, I am very much in the boom, not bubble camp, but I am going to represent the reasons it is a bubble with as little snark as is manageable for me.
By far, and without a shadow of a doubt, the number one reason that some people think this is a bubble is the notion of circular investment, with the XAI deal being the latest in a string of examples.
You might have seen some version of this chart going around showing how interconnected all of these companies are, or maybe this one which shows the actual dollar amounts going between different companies.
To take just example of that new Nvidia XAI deal, Nvidia will invest $2 billion into XAI, which will show up three months later as revenue on Nvidia's balance sheet after XAI uses it to buy chips.
OpenAI's deals with AMD and Oracle work similarly, with investment dollars and revenue getting passed between firms to boost everyone's bottom line.
This is bringing up what are for some uncomfortable questions of vendor financing, and some think it's amplifying and overstating demand in the entire sector.
Stanfill Capital, for example, wrote, none of these circular AI deals would be happening if there were genuine cash demand for the chips at list price.
It's obvious that the economics of this industry don't work unless the chips are hugely discounted, which means the chip makers are hugely overvalued.
Alright, so circular investment, reason one that people think this is a bubble.
Reason two is the risk that AI infrastructure will be overbuilt.
Kip DeVere, the CEO of private equity group Aries Capital Management, told Bloomberg on Tuesday, If you look historically in areas like this over the past 20 or 30 years, typically when this much capital comes online, some of it at the end of the day has to be marginal.
These trends tend to lead to overbuilds in certain places, so us being selective and measured in what we build is important.
Now, at this stage, there are trillions of dollars worth of AI data center commitments over the next five years.
And what's fascinating is that while this has financial professionals seeing ghosts of overbuilt telecom structure from the past, all of the hyperscalers and the big foundation labs continue to claim that their greatest risk remains to little compute rather than too much.
JP Morgan reported this week that AI companies are now the largest segment of the investment-grade debt market, with $1.2 trillion in commercial bonds issued.
They've reached 14% of the market in total, overtaking U.S.
banks for the first time.
And although the numbers are getting big in absolute terms, the report was still positive, stating, "...debt tied to AI companies is growing fast, but it trades tight for good reasons."
The report noted that AI companies tend to be cash-rich, not highly levered and highly regulated, which makes them very solid investment-grade bonds.
But for those who think this is a bubble, that of course doesn't speak to the fate of the industry if AI demand fails to keep up with the supply of data.
Last month, Bain & Company forecast that $2 trillion of annual revenue would be required to fund AI Compute by 2030.
Their analysis found that AI-related revenue would fall short by $800 billion.
So that is bubble argument 2 over building.
Bubble argument three is the echoes of dot-com.
Many people feel like they've seen this movie before.
It's important to note that most of the senior people currently on Wall Street had their seminal experience during the dot-com bubble, making and losing their first fortune in the space of a few years.
To them, this bubble has lots of echoes, including overbuilt infrastructure and circular investment.
During dot-com, millions of miles of fiber optic cable was laid to power high-speed internet, but as much as 90% of it lay dormant for years after the burst.
Circular revenue was also a huge problem during dot-com, especially among the companies that were little more than a website and a ticker symbol.
In addition to vendor financing of the build-out, you also had the circularity of what little revenue there was going on in the web space, where most of the revenue was from banner ads, which were largely bought by other dot-com companies, essentially meaning that investment dollars flowed around the ecosystem, adding revenue to each company in turn and massively inflating financials.
OpenAI chairman Brett Taylor recently said,
I think there's a lot of parallels to the internet bubble.
It is both true that AI will transform the economy, and I think it will, like the internet, create huge amounts of economic value in the future.
I think we're also in a bubble and a lot of people will lose a lot of money.
So, bubble argument three echoes of dot com.
Bubble argument four, stretched valuations.
The Shiller price-to-earnings ratio, which is a standard metric to measure stock valuations, hit a high in September.
The S&P 500 was valued at the highest level since, you guessed it, 2000.
Now, most analysts are careful to note that there's no reason theoretically that richly valued stocks can't stretch even further.
Russ Mould, an investment director at AJ Bell, said, the U.S.
equity market looks expensive relative to its history pretty much any way you slice it.
However, he added, valuation does not guarantee an imminent accident.
Still, if you were looking for reasons a bubble might burst, incredibly high stock valuations is one of the core warning signs.
On to argument five, part of the bubble logic is just the sheer size of what's happening and the idea that this scale of activity simply isn't sustainable.
On Monday, the Financial Times published an op-ed from Rockefeller International Chair Rishi Sharma entitled, America is now one big bet on AI.
He claimed that AI investments have accounted for 40% of US GDP growth and 80% of the gains in US stocks so far this year.
Shahram argued that AI is increasingly viewed as a magic fix for anything wrong with the economy.
Backfiring tariffs, consumer debt defaults, and a deteriorating job market, AI productivity gains are the big bet to smooth everything over.
Now, the view is more about the consequences of an AI bust rather than evidence of a bubble.
The idea is that AI investment has become too big to fail, but is also wildly speculative.
For Sharma and many other bubble-concerned, it is of course nonsensical that producing tokens to the exclusion of anything else could be a sustainable premise for an economy the size of the U.S.
Let's throw in one more argument for the bubble, which we might call speaking it into existence.
Remember, in many ways, bubbles are about narratives, and they unwind when results fail to live up to expectations.
When there is, in other words, a narrative fracture or a narrative disconnect.
At the moment, expectations are sky high.
AI is viewed as a technology that has the potential to change the world, a narrative that has been only applicable a few times in the past century.
However, because the bubble is built on narratives, sometimes it only takes a few errant comments from leaders to make it burst.
And those comments seem to be coming more frequently.
Last week, while on a press tour of their Texas facility, Sam Altman said,
However, he assured the press, over the arc that we have to plan over, we are confident that this technology will drive a new wave of unprecedented economic growth.
Altman has made a string of comments like this over recent months, and seems far less disciplined, frankly, in his messaging than the veteran CEOs like Jensen Huang or Larry Ellison.
At Monday's Dev Day, he acknowledged that stocks jumping when they were mentioned on stage was, quote, weird and something they're getting used to.
Altman's bubble talk might not be enough to pop the bubble all on its own, but it is certainly contributing to market nerves at the moment.
Small, nimble teams beat bloated consulting every time.
Robots & Pencils partners with organizations on intelligent, cloud-native systems powered by AI.
They cover human needs, design AI solutions, and cut through complexity to deliver meaningful impact without the layers of bureaucracy.
As an AWS-certified partner, Robots & Pencils combines the reach of a large firm with the focus of a trusted partner.
With teams across the US, Canada, Europe, and Latin America, clients gain local expertise and global scale.
As AI evolves, they ensure you keep peace with change.
And that means faster results, measurable outcomes, and a partnership built to last.
The right partner makes progress inevitable.
Partner with Robots and Pencils at robotsandpencils.com slash AI Daily Brief.
Chatbots are great, but they can only take you so far.
I've recently been testing Notion's new AI agents, and they are a very different type of experience.
These are agents that actually complete entire workflows for you in your style.
And best of all, they work in a channel that you already know and love because they are purpose-built Notion superusers.
Notion's new AI agents completely expands the range of what Notion can do.
It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days and get them complete in minutes.
These agents don't just help with work, they finish it.
Getting started with building on Notion is easier than ever.
Notion agents are now your very own super user to help you onboard in minutes.
Your AI teammates are ready to work.
Try Notion AI for free at the link in our show notes.
What if AI wasn't just a buzzword, but a business imperative?
On You Can With AI, we take you inside the boardrooms and strategy sessions of the world's most forward-thinking enterprises.
Hosted by me, Nathaniel Whittemore, and powered by KPMG, this seven-part series delivers real-world insights from leaders who are scaling AI with purpose.
from aligning culture and leadership to building trust, data readiness, and deploying AI agents.
Whether you're a C-suite executive, strategist, or innovator, this podcast is your front row seat to the future of enterprise AI.
So go check it out at www.kpmg.us slash AI podcasts, or search You Can With AI on Spotify, Apple Podcasts, or wherever you get your podcasts.
Today's episode is brought to you by my company, Superintelligent.
Look guys, buying or building agents without a plan is how you end up in pilot purgatory.
Superintelligent is the agent planning platform that saves you from stalling out on AI.
We interview teams at scale, translate real work into prioritized agent opportunities, and deliver recommendations that you can execute on.
What to build, what success looks like, how fast you'll get results, and even what platforms and tools you should consider.
Visit besuper.ai and book your AI planning demo today.
Okay, so those are five, actually six arguments for why it's a bubble.
Certainly the ones that are at the very top of people's lists are the questions of circular investment, which is related to the concern of overbuilding.
And more recently, the Shiller price to earnings ratio and stretch valuations is really starting to weigh on people's minds as well.
But let's move over now into the top five reasons that bubble talk is overblown.
The first is, of course, that revenues in this case are real.
You might have seen some version of this chart circulating on X or wherever you hang out on social media, comparing AI revenues to the dot-com bubble.
It shows a chart demonstrating that Cisco valuations completely detached from earnings growth starting in about 1998.
By comparison, NVIDIA's stock price has been on an incredibly steep climb, but right alongside it has been an incredible streak of earnings growth.
Roe and Paul shared this chart and said, this is not a bubble.
Cisco was a valuation story, price inflated while earnings lagged, then the multiples deflated.
NVIDIA is an earning story, price climbs alongside surging earnings.
Lux Algo recently posted a 90s Time magazine cover that asked, is the boom over?
They wrote, Reminder that the dot-com bubble didn't top for another 18 months after the September 1998 cover.
And that was with companies like Pets.com doing $9 million in peak annual revenue.
NVIDIA did $46 billion in sales last quarter.
It is different this time.
The same is true in private markets.
OpenAI's revenue growth forecasts have frequently been dismissed as a fantasy, but they've managed to outperform every year since the release of Chachapiti.
What's more, thanks to the AI coding boom, Anthropic has come into its own this year and is now growing even faster than OpenAI.
Long-tail startups are seeing the same thing.
Stripe's most recent report on startup revenue found that the top 100 AI companies were hitting a million in ARR in 11.5 months compared to 15 months for non-AI SaaS startups.
Now, of course, these companies are not profitable yet.
This is revenue growth, not profit growth.
But still, revenue growth is significantly faster than previous startup booms.
Keep in mind, OpenAI and Anthropix revenue is not based on advertising or strange structured deals.
Close to $20 billion is now coming in from people to purchase a product for themselves or their employees that did not exist three years ago.
Full stop, we have never seen anything like that.
The second argument for why AI bubble talk is overblown is low leverage.
We heard about that $1.2 trillion in commercial debt tied to AI infrastructure, and anytime you get a number with a T, it starts to sound scary, but that debt is backed by very solid companies concentrated in Meta and Oracle.
At this stage, Google and Amazon seem to be funding their data center construction entirely off their own balance sheet, and Microsoft has a very conservative debt capital strategy.
And so far as NVIDIA's circular investments go, a big part of the reason that they are offering vendor financing and equity for GPU swaps is that they have an absolute boatload of cash.
They can't move the needle by reinvesting into the business, so they need to take bets on the AI ecosystem to set up the next leg of growth.
In a recent appearance on TBPN, Doug O'Loughlin, the president of Semi-Analysis, wrote, None of the hyperscalers are levered.
Microsoft has net cash and has become more creditworthy than the U.S.
government.
And what's really crazy is that Microsoft has a 5-bips premium to the Treasury.
The real issue is how much liquidity the private credit market can handle.
They're sitting on an ungodly amount of capital and they have to deploy it.
The traditional wisdom is that bull markets don't die of old age.
They're killed off, almost universally, by a credit crunch.
In other words, to bet on the bubble bursting right now is fundamentally about one or more of the hyperscalers running into acute debt problems rather than a simple loss of enthusiasm.
Reason number three why AI bubble talk is overblown.
Demand is real and growing.
At OpenAI's Dev Day, they said they were now serving around three quadrillion tokens a year.
Yes, get used to needing to now have quadrillion in your common parlance and frame of reference.
A group of 30 power users across startups and large enterprises have churned through more than a trillion tokens on their individual accounts.
Google's most recent figure showed a 100% increase in monthly tokens served between May and July this year.
As of July, they were serving almost a quadrillion tokens a month, and that number is up 100x since May of last year.
Point is, there is clearly a ton of demand for AI inference, and it's still growing at an incredibly rapid clip.
If anything, in fact, the growth rate has been increasing.
So while yes, one of the big risks that people see is AI infrastructure being overbuilt, when you look at the growth in token demand, it makes more sense why every single AI company seems willing to bet that underbuilding will actually be the larger problem.
Reason number four why the AI bubble talk is overblown is actually less about whether there's a bubble or not and more about what you do with it.
We might pleasantly sum this up as F you on buying.
Basically, one of the more recent counter narratives is that Wall Street simply can't afford to be bearish on AI anymore.
Over the past few years, we've had repeated drawdowns in AI stocks accompanied by some scary new narrative.
The deep seek moment and last summer's too much spend too little benefit note from Goldman Sachs are two prime examples.
Each time, retail traders have bought the dip while Wall Street missed out.
And you can feel the seething in some recent quotes.
Speaking with Bloomberg on Tuesday, Michael O'Rourke, the chief market strategist at Jones Trading said, The market is pricing these deals as if everyone who transacts with OpenAI will be a winner.
OpenAI is a negative cash flow company and has nothing to lose by signing these deals.
Investors should be more discerning, but this is a buy first, ask questions later environment.
Others are simply acknowledging that nothing is anywhere near as compelling as AI investment.
Wells Fargo chief equity strategist Oh Sung Kwan's week, outset of AI, I'm not really excited about anything.
The same is clearly true in VC.
AI has been a hot sector as far back as 2019, but has now grown to represent 60% of every VC dollar.
Ogozo, a self-styled contrarian investor, wrote, for reference, only 40% of the money went to internet companies in 1999.
If those AI companies don't start generating real revenue soon enough, investors will start stepping back.
And yet, there is no sign that investors have stepped back an inch.
Aside from some sputtering at XAI, every venture round for these big ones is still upsized and oversubscribed, and valuations show no sign of plateauing.
In short, across Wall Street and Menlo Park, investors cannot get enough exposure to AI.
A fifth reason that AI bubble talk is overblown ties back to the information's reporting on Oracle's financials and AI corporate accounting more generally.
One of the more thoughtful bubble claims has been that GPUs are going to depreciate far more rapidly than previous types of infrastructure build-out, and maybe even more rapidly than many companies expect.
This was a big knock on CoreWeave when they went public, as they were amortizing GPU depreciation across a six-year period, even longer than the industry's standard five-year schedule.
The logic goes that NVIDIA is releasing new GPU updates on a two-year cadence, and data centers will need to serve the latest and greatest.
That's why it's such big news that Oracle is still seeing fat profit margins from GPUs made in 2020.
It suggests that the millions of H100s currently deployed across the country could remain useful until close to the end of the decade.
And while it might feel like a very technical accounting point, the difference between GPUs becoming worthless in two years compared to five years or more is worth hundreds of billions of dollars to the industry.
Since we did a bonus for the AI bubble arguments, let's do a bonus for the not bubble as well.
We'll call this one the Watch Pot Doesn't Boil.
Bubbles typically don't burst when everyone is expecting them to.
Concerns about an AI bubble peaked in August and they look like they're ramping back up again.
That's why we're doing this show.
Historically, though, the moment you should be concerned is when your neighbor is mortgaging their house to buy the latest hot stock or crypto.
Meanwhile, in our case, ahead of us, we still have an eventual OpenAI IPO, the commissioning of gigawatt data centers, and a ton of energy infrastructure to come online.
Ryan Dietrich, the chief market strategist at Carson Group, posted, "...the bull market turns three this Sunday.
Just a reminder that the five previous bull markets going back the past 50 years that made it this far kept going.
The shortest was five years, the average was eight years, and two made it to double digits."
In a research note last week, meanwhile, Bank of America chief strategist Michael Hartnett wrote, every bubble in history has been popped by central bank tightening, and there's absolutely no sign of tightening anytime soon.
Now, one interesting rising narrative to keep an eye on is the idea that the AI bubble exists, but is good.
Speaking at an event on Friday, Jeff Bezos commented, "...this is kind of an industrial bubble as opposed to financial bubbles.
The banking bubble, the crisis in the banking system, that's just bad.
That's like 2008.
Those bubbles society wants to avoid.
The ones that are industrial are not nearly as bad, they can even be good.
Because when the dust settles and you see who are the winners, society benefits from those inventions.
That's what's going to happen here too.
This is real.
The benefits to society from AI are going to be gigantic."
So now if we might for a moment, let me add a few thoughts that I have as relates to this whole conversation.
As I mentioned, while I am in the boom, not bubble camp, I'm not dismissive of all the points that the bubble people are making.
The circular revenue conversation is important if I think it misses a few points.
Watching price to earnings ratios as signs of exuberance, that's important.
Having a sophisticated understanding of how fast these capital expenditures actually depreciate is important.
But I think that there are some meta things that go into these calls of bubbles that are worth calling out as well.
The first is a phenomenon that I want to call the big short generation, where everyone wants to be Cassandra.
That movie came out and made the people who actually spotted the GFC before it happened look like absolute superheroes.
And there has been cachet ever since then in calling out exuberance.
I think this is amplified by social media, which amplifies the historical fact that it's always been cooler to be skeptical than it is to be exuberant and optimistic.
And while I certainly don't think that this explains the phenomenon on its own and would be wildly dismissive to individual investors who do believe this is a bubble to say it's all about this...
I've lived for a long time creating media and I've watched a lot of people call a lot of bubbles, both that came to fruition and that didn't.
And the common thread is that people really, really like calling bubbles when their voice can get amplified for doing so.
A second phenomenon I want to identify is something that we might call the rearview fallacy.
This is the idea that our sense of what's possible in the present and future is constrained by what we've seen in the past.
And on the one hand, this is completely natural.
We've only had our lived experience and the experience of history as the possibility set, so it's very hard to imagine a future that isn't constrained by that possibility set.
However, the entire point of the economic system that we've designed is to be expansionary, is to unlock possibilities that were impossible before.
That's also the mandate of science, the mandate of technology.
And I think a lot of what you see, especially when it comes to comparisons to dot-com, is basically just an inability or an unwillingness to see these massive numbers, these massive growth as real.
Because we're talking about such enormous numbers, it feels like it must be bubbly.
This gets back to the quote we heard earlier about the fact that just because stocks are richly valued now doesn't mean they can't get more richly valued in the future.
It is enormously difficult, in short, for us as humans to imagine futures where the possibility set is bigger.
And the faster those changes happen, the harder it is for us to wrap our heads around.
And so I do think that a little bit of the bubble talk by how anchored to the past we naturally are.
A third factor that's a little bit less undergraduate psychology class is that simply put, I think that most people who are calling this a bubble are radically underestimating both revenue and growth.
I think that people see OpenAI's $12 billion or the aggregate of that plus Anthropix 5 plus a bunch of other companies between $100 million and $1 billion and say, look at the gap between that $20 billion and the hundreds of billions or trillions being committed to growth.
That, of course, however, doesn't take into account cloud revenue growth from the hyperscalers.
And even when you've added that in, I think most analysis of AI revenue doesn't go to the next step and actually look at factors like the fact that Meta continuously talks about how much better their ads are doing because of the implementation of AI.
In other words, there is a lot of AI shadow revenue that's hard to spot that I think is underestimating where we are right now.
Also, when it comes to growth, I think that people are just wildly underestimating what's possible and what's coming down the pipeline.
I talked earlier this week about the KPMG 2025 CEO outlook, which is a study of 1,350 CEOs of companies $500 billion in revenue or more.
83% of them anticipate spending between 10% and 40% of their budget on AI over the next 12 months.
That is such an enormous amount of additional incremental money coming in that we have barely begun to account for as we think about the possibilities.
Effectively, the skeptics are saying, look, for these numbers to make sense, AI would have to be the foundation for an entirely new economy.
And my answer is, well, yeah, exactly.
There is also this other detail, which goes way beyond the scope of an already too long show, but we have not even scratched the surface on a next generation of AI, i.e.
embodied AI.
Figure is debuting its O3 robot in just a couple of days, and there is just about no one who is factoring into their considerations how much inference demand there is going to be from actual industrial robots coming online at mass scale, despite the fact that that too is just around the corner.
The last part that I want to mention goes back to this Stanfill Capital argument.
Remember he wrote, It's obvious that the economics of this industry don't work unless the chips are hugely discounted, which means the chip makers are hugely overvalued.
He said none of these circular AI deals would be happening if there were genuine cash demand for chips at list price.
We already made the point that part of why these circular deals are happening is that NVIDIA has so much cash they have to do something with it.
But there's something bigger here.
These deals and these non-traditional financings are not about artificially inflating demand.
They are about cheating time.
These deals are about betting on the future and paying for that future with resources for the future so you can act today to get to the future.
Now, that does not mean that there aren't the seeds of potentially systematic failure in those types of dealmaking.
That's not what I'm arguing at all.
But when it comes to the motivation, it is just simply incorrect to say that these companies are somehow forced to make up these deals to artificially inflate demand.
These companies are trying to cheat time and move faster than the gravity and physics of markets would otherwise allow them to.
Ultimately, though, we have to be humble and recognize that no one knows who's right about this.
So what are things that are worth watching for over the next set of months or even years as this AI boom or bubble, depending on your perspective, continues to grow?
What are the things, in other words, that would be clearer signs of trouble?
The first and most obvious would be a true and unrecoverable loss of faith in the technology.
We've seen little miniature versions of this with MIT's 95% of AI pilots fail report, Sam Altman's bubble talk, and other events like that.
But if we get to the stage where enterprises are actively withdrawing their AI spend, pulling back from how invested in this category they are, if we get to the stage where CEOs are being rewarded for dismissing AI as a failed technology, that would be a huge indicator of trouble to come.
Another thing to watch should be big failed catalysts.
Imagine, for example, OpenAI struggling to raise new funding, or their stock dropping hard after an IPO.
Those kinds of big visible signs that the market was refusing to put the next incremental dollar into AI investments.
Another thing to watch for is infrastructure failures.
The energy build-out is becoming a huge narrative in AI, and right now it seems as though data center construction is outpacing new energy coming online.
AI being blamed for major blackouts would be a terrible look and could cause enough destabilization to really have a big market impact.
There's also revenue growth.
If I'm wrong and we start to see a tapering, if enterprises don't pour in money like it seems like they're going to, all of these investments start to look a lot more suspect.
So by far the biggest factor, if you're just trying to look historically, the only thing that has ever truly popped a bubble is a credit crisis.
As Bank of America's Michael Hartnett said, all bubbles end from central bank tightening and an inability to continue refinancing debt at a higher rate.
The dot-com bubble, the housing bubble, the crypto bubble, they all popped after interest rate hikes caused a credit crisis to rip through the sector.
And it is absolutely the case that with these circular deals, a few key defaults could spread contagion.
This is certainly the factor that the non-ideological folks on Wall Street who are trying to take a balanced view of this are the most concerned with.
Morgan Stanley Wealth Management CIO Lisa Shallott recently told Fortune, Every morning the opening screen on my Bloomberg is what's going on with credit default swap spreads on Oracle debt.
People start getting worried about Oracle's ability to pay.
That's going to be an early indication to us that people are getting nervous.
However, when asked when the bubble would pop, she said probably not in the next nine months, but possibly over the next 24.
The short answer is, of course, no one knows, but hopefully you now have a better sense of where I sit with this and the arguments, frankly, for where everyone sits with it.
As always, do your own research, or at least prompt deep research to do it for you, and make up your own mind.
For now, that's going to do it for this quite long edition of the AI Daily Not So Brief.
Until next time, peace.