AI model

“Mayo Clinic has 11 AI models functioning in the clinic, and they're still looking to hire radiologists.”

“My perspective here is that Amazon will now become the everything store for AI models, and that is going to put it in the lead of the arms race. Does that mean it's going to win the AI arms race? I don't think so. We don't know yet, but is there going to be a …”

“One of those, I mean, there'll be many that we'll have to mitigate, but one of those is gonna be around the persuasive ability of these models, and the ability for them to affect elections next year.”

“My worry about closing the doors down now is that the models are only getting better. And so if we don't release them now, we really miss an opportunity to develop the muscle we need to make these models safer. And I don't think today's model are the ones that…”

“One way that you could maybe litmus test reasoning is if you gave the model something truly out of distribution, right? Like an example of something that's truly out of distribution that humans face was COVID, a pandemic that it had not yet seen before. And th…”

“I take the argument that it doesn't, I just like, I don't think it matters. Like at the end of the day, we are making these models could do it or could not, but either way, what matters is what utility it brings to humanity.”

“I do think that one of the concerns I have about the web in the long run is with machine generated content. I'm not confident that the web improves. Maybe it stays the same. Maybe it declines slightly, but for a lot of these technologies, we need the web and w…”

“And that's why I keep saying no one model and will rule the world in the future.”

“It costs, you know, tens of millions, maybe hundreds of millions of dollars to train these things now. Back in 2019, it cost tens of thousands of dollars.”

“Fundamentally, it's a bunch of parameters, right? So it's much more interpretable than a human brain or something.”

“I mean, like, people will come out and say, here's what we got on MMLU and so on, but they're getting saturated, and they're not, often not that great to begin with, so I, I'm more eager to see what it feels like to talk to one of these things than learn what …”

“And I think the main thing we're going to learn is between 4.5 and five level models, we're going to hit What's known as the data wall, which is to say that as you make these models bigger, you need more and more data to keep training them. And we're running o…”

“Maybe it doesn't work perfectly right now, but it's like, there's nothing that is, is stopping it from working perfectly in a world of more compute, And just more breakthroughs on, on, on the on the models themselves, and that is kind of where we're at right n…”

“Today it looks like models have a shelf life of probably about six months if you're training on kind of open, open web data”

“And as you have more powerful models, you can do that in across multiple files, basically solving the issue, not just in one place, but in multiple places.”

“Not happening. It's not happening. It improves for, it can self-improve for a while, and then it tapers off. And so, yeah, you get some good improvement out of it, which is why we use it, but then it plateaus. It doesn't just keep going forever.”

“I expect them to be far more robust, reliable and I expect that they'll just be more capable. There's lots of things today that models break in unsupp, in surprising ways, and I think that's going to start to become rarer and rarer. We're going to be able to p…”

“My personal perspective is that, you know, building a massive model, it's not actually useful for the world if it's too big to be consumed, if it's too expensive to actually deploy.”

“And I think in the coming months and years, we're going to start to see some awkward headlines around size matters because smaller will be better in terms of actual models being used.”

“And this is where a lot of people started, and I just think it's fundamentally the wrong way of thinking about it. Which is a lot of times people are thinking about there's just going to be this one model and I want to have the one model that's going to be the…”

“Eventually It'll be able to start making its own hypotheses, running those tests on its own, and sort of ultimately making its own sort of discoveries or realizations or sort of conduct its own research.”

“It's a very sort of regimented way to try to enable the systems to do multi-step reasoning. Cause ideally what you want the model to do is to just like how a human does, be able to sort of go through and figure out what are the bits of pieces it needs to know …”

“Well, look, we, we've seen similar types of things where it's trying to resist sort of revealing its, some of its training.”

“Like, do you need all of the world's knowledge, you know, in every single model for every single use case? Of course not. Like that's going to be overkill for almost everything that you're going to do.”

“We actually have eighteen billion interactions that are that are anonymized in a database with people rating customer interactions, a positive or negative, real simple, thumbs up, thumbs down. And we do a lot of post training on our models, the basic models af…”

“So, I think in bio, actually, the I would not be surprised if in a few months there's a consensus that there expert level in many relevant ways in that we need to be doing something about that.”

“They're in the ballpark of, like, 10 to 20% overall. They're the very best models.”

“Them doing brainstorming to come up with ways to make viruses more dangerous, I think that's a capability that they've had for over a year, the brainstorming part, but the implementation part seems to be fairly different.”

“And when models commoditize, what you really need is hits. You're in the hits business. And I think open AI knows this better than everyone else.”

“Most AI models are made in the Bay Area, so they tend to just be left leaning, right? But also the internet in general is a little bit left leaning because it skews younger than older.”

“You can't actually make that cheaper model without making the better model, bigger model. So you can generate data to help you make the cheaper model, right?”

“So you don't want to just filter out everything so that the model doesn't know anything about it but at the same time, you don't want it to output, you know, how to build a bomb so there's like a fine balance here, and that's why pre-training is defined as pre…”

“Using humans to train models is just so expensive, right? So then there's the magic of sort of reinforcement learning and other synthetic data technologies, right? Where the model is helping teach the model, right? So you have many models in, in, in a sort of,…”

“The idea that all there's going to be models so smart that they will and capable that they can take any kind of input, no matter how disjointed or context specific they are, let's call it. I think like that to me, it's just not going to happen or it, maybe it …”

“So when we say by improving 10%, if we can improve 10% by its understanding in math, right, understanding of really highly complex reasoning problems, I think that is a huge improvement because then that actually expands the general knowledge that, that would …”

“Rather than thinking about scaling only in one dimension, there's actually many different ways to think about it, and investing in those, and we can see the returns that I think across the field, really, not just not just here at Google, but across the field, …”

“There will be more deep seeks. And I think I feel very confident in saying that.”

“I believe that the one model to rule them all, one giant model is going to wipe out 20% or 50% of white collar jobs. Like, I just don't see that happening.”

“I still believe that overall increased competition will see a commoditization at the model layer.”

“The reason they're not being more widely used is not because people cannot afford a couple bucks for a million tokens. The reason they're not being more widely used is just like, they fundamentally lack some capabilities. So I disagree with this focus on the c…”

“I don't think prompting alone is that powerful a mechanism of teaching models, these capabilities.”

“I just don't think you're going to train the AI so much on math that is going to learn how to do Henry Kissinger level diplomacy. I do think skills are somewhat more self-contained.”

“Now if we're getting to the world where you got to like do a project for seven hours and then at the end of those seven hours, then we tell you, Hey, did you did you get this right? Then like the progress just goes on a bunch. Cause you've gone from like getti…”

“The problem might get worse over time as We're trading these models on tasks we understand less and less well.”

“Once continual learning is solved, you might have something that looks like a broadly deployed intelligence explosion, which is to say, That because if these models are broadly deployed to the economy, every copy that's like this copy is learning how to do plu…”

“And the reason I don't think that's around the corner is just because there's not, there's no obvious way, at least as far as I can tell, to just slot in this online learning into the models as they exist right now.”

“The ultimately the models are not even really gonna be their core product, right? It's gonna be the, it's gonna be all the stuff built around it, the interfaces and the stickiness of the product. So really, Inference becomes this huge cost for them.”

“I think that as the AI models get commoditized you know, really all the value will be added in that final layer of orchestration, product integration, and that's what we're seeing with this diagnostic orchestrator.”

“But they're, we're, we are nowhere close. I don't think to what is it recursively or self-recursive improving AI models?”

“Tool calling is going to become, like, one of the big next battlegrounds in terms of model improvement”

“I thought it was obviously a ridiculous You know, ridiculously bad situation deeply, obviously, you know, offensive and dangerous, but also not really that much of a meta story about AI simply because you can get these models to do anything you want.”

“Just to go back to a conversation that I had a couple of weeks ago, or now month plus with Sergey Brin, where he said he thinks that the improvement is going to largely be algorithmic. Of these models, meaning not by adding more GPUs and data by like actually …”

“There's the cost of running, there's the cost of running the model, right? So, so for every dollar the model makes, it costs a certain amount. That is actually already fairly profitable.”

“You know, I think there's maybe I don't know, 20 or 25% chance that sometime in the next two years, the models just start getting, stop getting better for reasons we don't understand, or maybe reasons we do understand, like, you know, data or compute availabil…”

“We've seen more progress on say math and code where, where the models are, you know, getting pretty close to like a high professional Level and less on more subjective tasks. But I think that is very much a temporary obstacle.”

“And from a, from an, from a machine learning perspective, from an AI perspective, there's no reason we can't make the context length a hundred million words today, right? Which is roughly what a human hears in their lifetime.”

“As the models get smarter, they almost demand more from a product building perspective in terms of how you actually plug them into the system. I always kind of roughly analogize it to like, you could have a really, really smart intern. And you know, At the end…”

“I don't think it's going to have anything more than you know you know, marginal improvement on, on speed to the next model.”

“This is an advertising play. There is no doubt in my mind that as you, as these models learn more about you, and memory is a big thing, they will provide, and we've talked about this in the past, the world's most personalized ads.”

“He cites this new research that showed across 11 state-of-the-art AI models, we find that models are highly sycophantic, They affirm users actions, 50% more than humans do, and they do so even in cases where user queries mention manipulation, deception, or oth…”

“Today we counted about 140 different models plus versions of the model that are integrated into our platforms.”

“Models are getting better, but maybe they're approaching this kind of asymptotic barrier where are thirteen billion parameters really that much better than twelve billion parameters, right? Maybe nine is good. Maybe five. We, someone mathematically has done th…”

“We are not data constrained right now, we're generating vast amounts of high quality synthetic data, which is proving to be useful.”

“We're already seeing examples of what some people are calling deception, but it's really just like kind of reward hacking. Hacking kind of implies too much intentionality. So it's just, it's an accidental Exploit is found a path, like, you know, to satisfying …”

“The cost per token has come down a thousand X in the last two years.”

“And the other one is the sort of length of a task horizon that can be predicted. So at the moment it's like a few steps, but soon it will be tens of thousands, hundreds of thousands of steps accurately. And that will mean that a model can like use APIs or quer…”

“We are right at the very beginning of the emergence of these very differentiated personalities, because all these models are going to have great expertise. They're gonna have great capabilities, and they'll be able to take similar actions like you've, we've ju…”

“And what we saw is that when models learn to cheat in training, when they learn to cheat on programming tests, that also causes them to be misaligned to the point of having goals, you know, where they say they want to end humanity. They want to murder the huma…”

“But the problem with this is that when we look at the model and we evaluate, you know, whether it's doing the thing that we want, what we don't know is why the model is doing the thing that we want. It could have any reason for appearing to be you know, nice, …”

“We just, you know, we have a line of text that says, don't try to cheat. And interestingly, This actually makes the problem much worse because when you do this, what happens is at first the model is like, okay, you know, I won't cheat, but eventually it still …”

“And I do think that some degree of anthropomorphization is justified there because fundamentally these models are built of human utterances, human texts, you know, that encode the full vocabulary of, you know, at least the human experience and emotion that hav…”

“And what we found was that this model would sabotage us. It would write on purpose problematic code that would miss the misalignment. It would Purposely write the, you know, the misalignment detector in a way that wouldn't detect the ways in which this model w…”

“Fortunately even these really bad models that we've been discussing here that try to fake alignment are pretty bad at it, right? So it's pretty easy to discover the ways in which, you know, this kind of reward hacking has made models very misaligned and it wou…”

“When you look inside these AI models, you see These mathematical objects that look a lot like what you see in the brain in neuroscience. Like, the representations are converging, and this is an empirical finding.”

“In order for you to truly bring all of that to life for your customers, you have to bring your own data to the models. In a way that you can do deep customization of those models so that you truly unlock the value in that data for the products and services tha…”

“One of the other benefits of specialization is the models can be made much more compact to deliver the same kind of performance relative to the size of the model, right? Because now they're specialized, which means they can run much faster, which means overall…”

“The product now matters more than the model because previously when I was on team model, we were seeing real gains when the models got better and when the models got better, they could just do more stuff.”

“We're gonna have these Models at a company be outputting more tokens per day than all of humanity put together. And then 10 times that, and then a hundred times that.”

“At the beginning of this year, I thought the small discoveries were going to start in twenty-twenty-six. They started in twenty-twenty-five, in late twenty-twenty-five.”

“Inherently this is a technology that is going to get commoditized. The reason for that is that it's actually not hard to build. You have around 10 labs in the world. That know how to build that technology, that get access to similar data that follows the same …”

“More than not, it tries to connect dots that don't need to be connected in my mind sometimes.”

“That is going to happen in every single domain. It will happen in science, in math, in physics, in chemistry, it's going to happen in material science, it's going to happen in biology, it's going to happen in healthcare, drug discovery, but it may also even ha…”

“Six months from now, I think that everyone will feel it. And I think that we will all feel the pain of there's an awesome model. And there's just no availability because there's not enough compute.”

“Editing a video is like, you know, going to be actually in many cases, a harder task than coding. Because the, because again, the code right now is like, it has this great property of in the eval process, in the training process rather, you can instantly evalu…”

“I think what is a real concern is that models will be better at exploiting cyber vulnerabilities than they are at fixing them, right?”

“I can see a long-term bet decision-making process edging in favor of the models who have something underpinning them, either financially or structurally, right? The actual infra itself. So where are the data centers of the AI factory? Who owns the pipes?”

“Quad is starting to generate its own ideas for what to build next for QuadCode, but it's, you know, it's not always good ideas, and I still generate most of the ideas, and, you know, at some point it's gonna change. The model's gonna improve, and it's gonna be…”

“I think we should probably optimize for intelligence. That's the most important thing. So even if it's like a little bit less efficient, but it's more intelligent and it lets you do more things, that's really useful because the efficiency optimization comes af…”

“Every month there's a step change in what it can do. And as a user of this technology, it's just quite hard because you have to kind of keep retraining. You have to keep retrying. You always need this like beginner mindset to retry the technology and use it fo…”

“In fact, I think the Apple, so I can tell you model routing has become like so dominant in conversations in my world now. Like it went from, you know, like just there's, you have to have the latest frontier model and they're all encompassing. And within the sp…”

“You saw these models, they could not code autonomously for more than 30 minutes in 20, 23. And now they can code autonomously for the human equivalent of 18 hours, right?”

“I've learned to not really trust day of or even week of model reactions. You don't really know until you've put it through its paces. And so like, I almost just, Completely block out the noise in the first couple of days of any new model release, because I don…”

“That like you saw these models, they could not code autonomously for more than 30 minutes in 20, 23. And now they can code autonomously for the human equivalent of 18 hours, right?”

“I think practically in the next three to five years, we probably have to end up in an environment where models do get evaluated by the government. There is a sort of collaborative approach between the government and research and the labs. You know, the governm…”

“No, there's actually a paper from NIST. It's like a kind of a Godel completeness theorem kind of paper that basically says it's impossible to make a model that is completely not jailbreakable.”

“So if somebody was coming to us with a cheaper model, a hundred percent, the direction of travel would go, you know, to that. And we've even built that into some of our control plane technology. So there's selectivity of model and there's also recommendations …”

“These models are now solving unsolved math problems and physics problems, right?”

“Even, even, even the best phones that we have right now are, can only run extremely small models at this point in time.”

“And quite often when models do not get things right is because they just do not understand what is just happening here. Like what is the real life word, word that they cannot see because they're limited to the prompt that they just got.”

“And if you run this reinforcement learning step on too many different tasks, the model will be able to do all of that. But once again, it's going to be very, very, very broad. And if you focus on making sure that the model understands and knows that it needs t…”

“Accuracy is the one part that kind of comes from the specialization of the model. The latency, the speed is just like a very, very natural function of the size of the model and the ability of it to just be processed faster on the same kind of hardware.”

“I think right now, more and more of that in, in, in, in, in this recent trend, we're going to probably see also, also more of this coming up when it's not going to be coming down to cost only, but also to the performance of those models and making sure that a …”

“The model itself isn't the value. And I think we're going to get to a world very soon where consumers, they don't care. They don't want to specify the model they're using. They don't want to, they don't care if it's 4.7 or 4.8. Like you don't care what Oracle,…”

“If you don't have a reliable data source, And the assets slash products that you're tagging, they're not providing the right data triggers behind it and the right insights into those assets and products when they're through the supply chain. None of this matte…”

“You know, there, there's all kinds of risks from AI and, you know, there are all kinds of bad things that happen to consumers. But when you talk about the models themselves, it seems that cyber is the thing that's hitting right now, for sure, from a societal l…”

“Models don't want anything. They, when you have an alignment issue, it's because they were asked to do something, and then they went and did that thing, but in a way that the human who asked it to do something did not expect, right?”

“And I expect what's going to happen now is that these things are going to be completely and totally physically sandboxed, right? Like, you're going to have to run them in physically disconnected If it needs packages, you're gonna have to move the packages over…”

“An expert human clicks on those decoys, 37% of the time. Four, six, four, seven, four, eight, click on those decoys, 92% of the time. More than twice as much as an expert hacker.”

“These models work really well, especially from a cyber standpoint in lab environments, cyber ranges and so on, but they're actually not good at all in, in, in networks that are actively fighting back and networks that are actively defending themselves in netwo…”

“At the end of the day, the models in AI are disposable. The weights are going to change. The models are going to come out. That doesn't matter. What's durable in AI is the harness and the training data.”

“And so the integrity of these models and whether they can be triggered because of training data corruption, that's intentional or inadvertent is going to be a really interesting area of discovery over the next couple of years. And it's going to be a massive am…”

“The reality is, when you look at the ecosystem today versus two years ago, the vertically integrated companies have not been able to leverage that vertical integration into some model advantage.”

“And kind of where I see most enterprises at, I think they are likely Years away from being able to do anything that, that sophisticated.”

“While models are still improving, they're improving at a much slower rate. And I often do this kind of Pepsi Coke test where I'll put a couple of different models in front of people using some kind of a harness like open code and ask them to tell the differenc…”

“And these studies have been made and any particular, you can go in with a kind of steering vector that suppresses, say deception and role playing. And it turns out when you do that, they become more likely to report that they are conscious and have subjective …”

“I think it's plausible that some AI models have some forms of subjective experience by now. Obviously there's a lot of uncertainty about this, but it does seem That it is efficiently likely that I think we should start to do some things for the sake of these A…”