Frontier AI Models

“Will a handful of CIOs get very publicly flogged and fired in the next year because they accidentally didn't understand this and just did an API deal because they wanted to feel popular and leaked data into these models? Guaranteed.”

“You're gonna have a few of these very powerful models. It's like stupid for them to limit what you're doing with these. It would make sense if they were the only model in the world, but it's kind of like having a group of people, and you go to one of them, you…”

“There'll be things that, this isn't surprising, like, there'll be things that rely on the model's strengths, like their ability to, like, grind out a long computation or, like, you know, pull together kind of technical ideas from many areas or, like, maybe man…”

“Today to build a frontier model costs, let's say, three to five billion dollars, right? So, and let's, you know, and that's to train it. And so the inference has to pay back at least that. Of course, right? You know, in order for any of this stuff to be viable…”

“If you only have a million dollars to spend, You could spend that million dollars on customization and getting specific tasks done a lot more efficiently than you can on a frontier model.”

“We're going to absolutely see this thing that, you know, Prohibits the cannibalization of frontier models, at least in the short term.”

“There's never been a time yet where the open source model has trumped a frontier model on innovation yet. They might be better at specialization, but they certainly don't compete yet in sort of Complex and being able to do complex tasks.”

“I mean, it makes sense if like a Manhattan style project, if we did that for AI 10 years ago, well, fine, you know, do it because it's strategically important to the country. But today, given the size of them, I think the only reason you do it is For political…”

“The most valuable time for that to happen is at the latest possible moment. Because then you would have the actual system that you're trying to align to work with.”

“Sandbox is what we've seen this year, and we've seen many examples. They are pretty much easy now for these models to escape from. It's really hard nowadays to say, I'm gonna make a fully, you know, foolproof sandbox. I'm sure it's gonna be resistance against …”

“And it's not because the model is dumb, but I think it's because probably, I mean, part of it is because of the training process and how they are trained to be efficient, how they use their token. But this means that they start to and bundle a lot of semantics…”

“I think that the price of compute is going to go up and that will provide an advantage to The models that have the most lucrative algorithms that are able to produce the most intelligence per watt or the most intelligence per token or per GPU.”

“I'm no longer thinking that we need two or three orders of magnitude on the largest models to be able to You know, solve knowledge work the accounting, the legal, the code that we write.”

“What the AISI has said is that the difference between the frontier and the Chinese models is about seven months, but I would argue that that underestimates it because the Chinese models that we see are under-trained. So that the internal Chinese capabilities a…”

“So I think there's a chance that over the course of the next two to three years, As we take on much more complex agentic tasks, that the distance between the frontier and everybody else doesn't converge, it actually extends.”

“80% of non-coding tasks in the enterprise can be done with models that are not at the frontier.”

“If you make these models available through everyday consumer subscriptions, there will be groups that will create networks of, you know, thousands of accounts and distill the models. And then you're giving away this, like, advanced capability.”

“I suspect the Frontier AI models, as a crystal ball, they will spend a lot more time in the next year or two building memory around consumption.”

“I think you can look at the last several months and see the opposite pattern, which is that the frontier is actually gaining share.”

“Zooming out, I think the most important thing in all this Is that we step back from Anthropic and realize that the models are to place now, whether they're from Anthropic or elsewhere, where they can autonomously find and chain multiple vulnerabilities togethe…”

“So there is a time in the next year and two where we'll start to see these directionally, yes all the token prices are going down. But as the models are maturing, we are starting to see that, you know, the opposite effect as well, where, you know, the cost is …”

“You're going to have, pick a number, three to six companies making a frontier model. Spending, no one knows, no one honest knows, like something between two hundred billion dollars and two trillion dollars a year on building these models.”

“I have a belief that whenever he wants, Jensen can probably get pretty close to the frontier. Patrick O'Shaughnessy: With his own model? Gavin Baker: With his own model.”

“Since this industry has kicked off, there's never been a moment where the delta between the, you know, the best model and the second best model was, like, more than maybe, like, a 10, 15% gap.”

“Now, it is also the case that it's not as simple as you can take the output to these models and distill and you have exactly the model of the same capability, It's just smaller and can run fast. If that were the case, we would just do that”

“Where they only release, meet those to certain companies for cyber, that's just going to be something that continues. Models will have less broad and less broad deployment.”

“Economic value that the best model can deliver is growing faster Then our ability to actually serve those tokens to people via the infrastructure. And so this gap will continue to grow and the model labs will continue to have expanding margins until people in …”

“How many, the upgrade cycles for these things are now compressing to now three-ish months. So every three months you have a new set of model weights, a new set of guardrails, a new set of bugs in the way the model behaves, the way it hallucinates or doesn't ha…”

“Actually the most useful one is to mitigate this risk is just to use a strong model. Because the smarter the model, the better it is actually at like not falling for these prompt injection tricks.”

“People are not that, no matter what anybody says, we're not that price sensitive today.”

“We find, we think it's not capable enough, you know, on the basis of some of this capabilities evidence that you've alluded to commit these catastrophic harms.”

“For other bigger, you know, frontier type models, I mean, you're not going to replicate billions in R and D you want to leverage what's out there and take the best of it.”

“I'm not only taking the frontier models and just making them available. I'm doing it across classification levels. I'm combining some of the most, you know, important data that the United States has from an intelligence standpoint, we're deploying it to war fi…”

“What I'm looking for is like, yes, explain this to me in like the Like, the clearest, most sophisticated, most complex, most, like, complete way that it's possible for somebody to, you know, for a real expert to be able to explain things to me. And that's what…”

“These models ate the internet when they were pre-trained, but the internet data is used up. It was used up like three years ago, right?”

“The challenge today is not necessarily make the model bigger. The challenge is how do I generate and create data that is in useful domains so that the model gets better at them.”

“I believe that getting to a frontier model should cost a million dollars or less for most organizations, at least in a specialized domain, right? And when you think about what enterprises need, that's generally what they need. They don't need a model that can …”

“If you really want to get to fi, if you want to get to a billion and then five billion people you can't do that with a paid offering. Like it just, you, at any sort of reasonable price point, it's just not possible. The, you know, global per capita GDP is not …”

“Two years from now, my guess is that we'll have models that cost of the same order of magnitude that they cost today, except they'll be much more capable of doing work, much more autonomously, much more broadly than they are capable of today.”

“I think that you can do a lot of very simple bite-sized tasks, but I think it's just much more convenient to be able to use an AI model that can do a very complex task end-to-end, rather than requiring us as humans to sort of orchestrate a much dumber model to…”

“I mean, I'm thinking right off the bat, probably the biggest thing is models that have been post-trained on more general agentic trajectories over very long time spans. That feels like something that is There's an effort for that right now. But what I mean is …”

“So it is a fast depreciating asset. The second you're off the frontier.”

“So my basic thesis is that, you know, we're potentially hitting diminishing marginal returns. We're going to see less gains from increase, increased scale.”

“Mistral, a lot of the other model companies have fully pivoted, like they've already raised their hand, white flag, we don't have the revenues that can support the spend to keep up, right?”

“One thing that we see today with the models is that because all the benchmarks that were used today are what's called saturated, i.e., you know, in other words, like all the models do really well at the benchmarks, it's really hard to discern actually which on…”

“In the short term, I think we're likely to see more divergence, you know, over the next couple of years than convergence.”

“I think it's almost zero percent the way a frontier model created and the way the frontier models exist today. I think that wave is gone.”

“So it isn't just, you know put in data, right, you know, put in compute, press button, go away for six months, come back, and go, ha ha, I've got GBD-IV. There's a whole bunch of art and science, and what I mean by art is like, well, you train it this way firs…”

“On a per token basis. They're actually really smart, potentially as smart as Really smart humans. It's just that five minutes out, they lose their train of thought.”

“If you want a model that's on the frontier, that's going to be closed source. Now, look, every, as I mentioned, every month it feels different with what Meta's doing with Llama, that may actually fundamentally change the game. If Meta is willing to make that h…”