Reid Albergati
Appearances
Today, Explained
DeepSeek deepdive
And this new R1 model that DeepSeat came out with, it's an advance, but it is not nearly a big enough breakthrough to sort of negate those market dynamics.
Today, Explained
DeepSeek deepdive
Yeah, I mean, they showed that you can do some of these types of queries at a lower cost, but it's just not nearly low enough. You might have seen Microsoft CEO Satya Nadella talking about Javon's paradox.
Today, Explained
DeepSeek deepdive
Basically, as the technology becomes more efficient and the cost declines, the paradox is that you would think, well, okay, that just means that it just gets cheaper and these companies are just not going to make as much money on it. But actually what happens is it becomes more useful and people want to use it more.
Today, Explained
DeepSeek deepdive
You know, it's stolen. I mean, that's a very strong word. We saw David Sachs, who's the incoming AI czar, sort of accused DeepSeek of stealing from OpenAI.
Today, Explained
DeepSeek deepdive
So you need data to train these AI models, but what you can actually do is you can use the models themselves to create a very, very specialized kind of data. It's really synthetic data because it's being generated by an AI model, but you can create exactly the kind of data that you want. And you can check that over with other AI models.
Today, Explained
DeepSeek deepdive
And what you end up with is that's how you make these models much more efficient. So this is also not surprising because that's how all of these models work. I mean, we've seen lots of companies do this. So I just, again, the process of distillation makes total sense. Whether or not it's stealing, I think that's something that's a gray area in the AI industry that we really haven't...
Today, Explained
DeepSeek deepdive
ironed out yet i think it's such a new thing that we'll have to sort of come up with the norms and the rules and regulations maybe even copyright law around this
Today, Explained
DeepSeek deepdive
Yeah. Yeah, well, NVIDIA makes these graphics processing units that are the most powerful, the most advanced in the world. And they're very expensive. I mean, the older models, the H100s, which were state-of-the-art, they cost about $40,000 a piece. And these data centers have about 100,000 of those. So, you know, you do the math there. NVIDIA is selling a ton of these chips.
Today, Explained
DeepSeek deepdive
Really, they can't sell enough. There's way more demand than they can even produce. And it's all because these models take a lot of energy to run. And so if you can have a more efficient one that doesn't require these powerful GPUs, then maybe you don't need to spend $40,000 on a GPU now with OpenAI. But again, that's not really what happened here.
Today, Explained
DeepSeek deepdive
What happened here is there's a bit of an advance in how efficient these models can get. But in order to get the most use out of them, you need to run a lot of inference on those. You're still going to need really powerful GPUs. And, you know, as, as there are more advances in the pre-training part of the models, they'll get even bigger and more powerful. So Nvidia is not going anywhere.
Today, Explained
DeepSeek deepdive
I mean, certainly they have competition. There are chip makers that want to build a more efficient inference chips. Um, there are people who want to get rid of Nvidia's advantage, which is it's CUDA software that the whole AI industry basically runs on right now. It creates a big moat for them.
Today, Explained
DeepSeek deepdive
Those are the things that I think are the risks for NVIDIA, not a company building a more efficient open source AI model.
Today, Explained
DeepSeek deepdive
Yeah, and this is where I think there really are national security concerns about China and AI. And if China does win the AI war, the AI race, let's say, it will probably give them a military advantage. This is far into the future. There's a lot of debate about this, right?
Today, Explained
DeepSeek deepdive
But I mean, I think the conventional wisdom is if you win the AI race and you get your first to AGI or super intelligence or whatever you want to call it, it becomes a military tool very quickly. And I think that's the whole reason the US has put so much energy into figuring out how to curb the exports of the most powerful AI chips to China. They don't want to see...
Today, Explained
DeepSeek deepdive
China be able to sort of control its own destiny when it comes to AI.
Today, Explained
DeepSeek deepdive
Yeah, look, I think Sam Altman said that it was invigorating on X.
Today, Explained
DeepSeek deepdive
This is how research works. You know, somebody comes out with a new idea and it inspires other people both creatively and also competitively. This is a dynamic that we've seen for the past couple of years in AI, or even really more than that. This is why so many of these tech companies have been publishing their research instead of keeping it a trade secret for so long, because the genius...
Today, Explained
DeepSeek deepdive
researchers who write these papers, they want to present them. They want bragging rights at the NeurIPS conference. That's the big AI model conference every year. They want to get pats on the back from their coworkers. And I think there's actually probably a lot of mutual respect between the AI researchers at OpenAI and Anthropic and the ones at DeepSeek.
Today, Explained
DeepSeek deepdive
I think in that world, I think it's possible to kind of put aside all the geopolitics and just say, hey, nice job. You created a really interesting model and we're going to learn from it and try to do better. I think the other way to look at it is, look, if the U.S.
Today, Explained
DeepSeek deepdive
doesn't win the race to AGI, then what you could see is a Chinese military advantage that leads to something like an invasion of Taiwan and maybe potentially a hot war between these two countries. superpowers, and that would be very, very bad. I think the most fervent China hawks, what they really want is a US military advantage that is so big that there just will be no war.
Today, Explained
DeepSeek deepdive
And I think if you look at it through that lens, then yes, this AI race is very, very consequential geopolitically, and really there are dire consequences if the right outcome isn't achieved.
Today, Explained
DeepSeek deepdive
Well, I was really kind of slapping my forehead because I think it was a complete overreaction. People knew that this company existed. And in fact, this whole idea of distilling these larger models into smaller, more powerful ones that are more efficient, this is something that had been going on really since ChachiPT came out.
Today, Explained
DeepSeek deepdive
The biggest takeaway for me is that the market really does not understand the AI industry yet.
Today, Explained
DeepSeek deepdive
With hundreds of thousands of graphics processors, tens of billions of dollars. In fact, you probably heard last week there was a deal announced for $500 billion.
Today, Explained
DeepSeek deepdive
And what that investment is for is running these models because there is so much demand. These companies really can't meet it right now. And what we're also finding is that inference, which is just the fancy term for running these models, actually can now increase capability of the models a lot. That wasn't the case before when ChatGPT first came out.
Today, Explained
DeepSeek deepdive
It was just you prompt ChatGPT, it comes back with an answer. Now you prompt the most advanced model of these models, and they are doing a whole bunch of stuff in the background. They're running over and over and over again. They're trying to find the best answer. And that is exponentially more expensive. And that is just going to continue.