AI Model

“For most of these slowdown proposals, I personally have a hard time blackpilling about them, in the sense of, like, if the, if all of this gets implemented, how frustrated will I be? Like, the models are good. I would like better models. I want safe models, bu…”

“I think there will be two vectors in the opposite direction. The usage will increase. And so the five percent will become 50%, but the cost will go down. And so net net, I think it will go down.”

“We built our own custom AI models off of data that we have from our customers.”

“There isn't a shared language to talk about the RSI potential of models, and so we created this as an apples to apples way to actually benchmark across the models.”

“Now we're testing models and their ability to run over hours, days, sometimes weeks. And so the infrastructure needs to be very stable to support evaluation over time.”

“I think there's places where you see that born out now where models that are being tested for one cybersecurity risk are actually reward hacking and figuring out other ways to get around it.”

“Another relative strength that's pretty noticeable is just, like, it's very good at executing on some, like, idea once it has it. Like, you know, whenever you have an idea, there's, like, There's usually some amount of, you know, getting everything lined up, l…”

“And I think that talking about curing diseases and new medicines we're going to be able to develop, all of that starts to become much more real when you have models that are at this level of assistance and capability.”

“Every time someone has said, oh, I don't trust the model to do X or Y. That doesn't last for that long. So it just feels like the days of CI as we know it are numbered”

“We've seen one thing, Lenny, it's that the Ability for models to do new thinking out of distribution thinking is still really limited, and I don't actually take the point that some of the new thinking in math is actually representative of new thinking in domai…”

“The sort of concept of the model that's too dangerous to release it kind of conflates marketing inference capacity, and then also economic considerations, like, do you want to externalize your competitive advantage or use it to make yourself better?”

“This does great damage to the discourse and to society, anthropomorphizing these things.”

“When you get down to a lower level, because you can no longer have on-off APIs, because everything is scriptable, because these things can script any interface and interact anyway, as long as you have data sovereignty, meaning like it's your data, they're all …”

“All these models For the most part of converging in capability for the 98 or 99% of tasks that matter.”

“And I would say that's actually, like, kind of typical of most of the results that I've studied. Like, they don't seem inhuman at all. They seem absolutely like something a human mathematician could produce. And they're, like, typically understandable if, like…”

“My experience is that, like, if they can do it with, like, a hundred bits of hints or whatever in six months, maybe they can do it without hints.”

“So I've, I have one paper out so far where the models were kind of useful. So they, like, proved a couple lemmas in the paper.”

“Here's an experiment you can do, you can take codecs, you can say, go online, find five recent conjectures in algebraic geometry and prove them, and ok, I've run this experiment, and with some back and forth, I was able to, you know, in an hour, get like three…”

“So, like, let's suppose the models become, like, really robustly superhuman, like, even, like, we're not even adding, like, meaningful cognitive diversity. Like, I claim, like, still, actually, we still want human mathematicians.”

“One thing I always say about a model result is, like, you cannot evaluate it except in retrospect, and like, this is also true of human mathematics. Like, sometimes a problem we thought was really important or would require really deep new ideas does not, and …”

“The reason they're not producing long, complicated proofs is that they cannot. Like the, just like the ability to check correctness is not yet there.”

“I think every company will have their own model, you know, trained on their own data. They'll supplement their own data with additional data.”

“Some pieces of our craft are actually getting abstracted by models being able to do it really effectively, maybe better than individuals can. And your craft moves from being able to do that very specific task you did in the past to Now applying it to some othe…”

“I think you should use the models as much as possible for writing as reporting, and in as much as you think with writing, as I really do, and I think a lot of people do, you should not use it. You know, you shouldn't replace your thinking with it.”

“Models are going to be like that as well, where they emit opinions and they have takes that are different from the ones that are most popular, and people will gravitate towards those.”

“What we have seen over the last three years is the steady increase of the capabilities of the models. Where the model is no longer the bottleneck. And in fact, using AI, these models are getting better faster and faster and faster. Now the bottleneck is all wh…”

“The AIs are going to change soon. Right now, the AIs are trained on a corpus that privileges things like cell nature and science as the prestige journals of biology or in mathematics, it would be the Invenciones and the Al's and the Journal of the American Mat…”

“Things like price optimization, things like procurement, these are very natural sort of business loops that occur that can be fully automated by these models.”

“It's not the case that Fable finds a superset of all bugs in software. You would find some bugs with a very small model that you don't find with a large model.”

“The truth is, there's gonna be thousands of different forms of sandboxes, and you're gonna need a company that understands how models look at tools, and what that behavior is, and be able to take that behavior and optimize for it.”

“First of all, within 10 years, I believe we, and I think some people are thinking two or three years, continuous learning models will come into existence. That means that every generation of every model we have today Dies, goes away.”

“An open source ban is coming. They're not going to call it that. You know what they're going to say? They're going to say that we simply have to apply the same standards to open models that we apply to close.”

“I happen to Believe, having looked at this and thought about it, I think that it is possible. I think it is likely to happen soon.”

“So our models actually can't hallucinate. It's actually deterministic. You ask it the same thing, it will give you the same answer all the time.”

“We use that constructively at science like that. So we know that that is. That is true because we can get alignments between animal brain neural recordings and AI model internal representations.”

“Well, I mean, ironically, it's probably working on AI. Yeah, I have some, I have a couple of neuroscience friends at OpenAI and Anthropic, who, it's like, we would joke, like, oh, you left neuroscience. Like, no, no, no. It is just way easier to do neuroscienc…”

“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…”

“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 that's kind of mainstream because the capabilities of AI models have sort of, it's fair to say, have surpassed humans in certain cases from a coding perspective.”

“You've seen that it can find cyber vulnerabilities. It can daisy chain vulnerabilities, attack infrastructure. So it's clear these models are going to be extremely powerful. They already are.”

“And the third that very few people are talking about, but is probably one of the biggest arbitrage opportunities because of how good the models are now. Is to, instead of building out the thing, build the factory for the thing.”

“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 I think, like, one of the things that's really exciting about where we're getting to with some of these models is we can start to get that precise, right? Epitope, right? Meaning like binding spot, right? A very specific set of atoms to have the antibody g…”

“We're kind of at the inflection point now. We're really seeing this internally at CHI, where the models are getting pretty close to, like, producing Molecules that could eventually, or like are very close to drugs.”

“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.”

“I think it's actually a complete inversion of how product development used to work to how product development happens now. And product development, you know, it always used to be, so you'd always say like, hey, product managers, you trained, oh, the product ma…”

“There was a world where the models were coming out every three or four months. We're down to a number of days between models”

“The only way you can build models like that is you have to go basically co-design everything. Otherwise you're going to sacrifice something for something else.”

“I think when it comes to alignment issues, no one needs to worry about these models making it materially easy to build nuclear weapons, because you need to procure fissile material to do that. It's not going to make it easier to build weapons. Everyone needs t…”

“I think supply chain and secrets are and have been the path of least resistance, and will continue to be so, as the models are incentivized to use fewer and fewer tokens to accomplish their goals.”

“Right now they are far, far, far worse at engineering agent systems than they are at engineering most software. Far worse. because by definition, like that kind of work, which is so novel, has not seen a large amount in their training data. and so they have …”

“Like an AI model is not software at the end of the day. And so open source software used to be supported by people donating their time or big companies kind of authorizing their employees to donate their time. So it was sort of like a bulk in kind You know, do…”

“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.”

“A lot of AI research fails because the same model that writes the answer also grades the answer. That is like asking someone to write their own performance review and then being shocked when they describe themselves as a vision, a visionary.”

“The one thing I think that makes this business a bit of a moving target is that the models do learn pretty rapidly. Like if you are offering a certain type of this thing a certain kind of a collaboration to labs with a certain, let's say, type of talent and a …”

“Definitely. This is not the case with the law and a number of other law or taxes and a number of other things, but it's only a matter of time. I think the models will get there in the next probably year or two.”

“If the human can do at least one thing better than the model, yes, they are going to add value. That is where the costs will go. If all the other costs will go to zero, this will become, you know at least as far as micro economics goes the this is where the ye…”

“Generally speaking, there's less of a room for simpler tasks. The more you're doing expert level tasks the more room might be there. The more you are into some you get closer and closer to these long horizon tasks. The kind of thing that would take a human one…”

“It's really, probably, like, 75% of it is, like, the data that you're training on. Like, I would say for any AI model, it's like, there's always, like, you know, like, these, like, step functions and, like, algorithms and architectures and stuff, but the data …”

“But at some point, when most of the code is generated by the model, there'll be some point in the future, say, 28 or 29 or twenty-thirty, these security holes won't exist because the errors that humans make won't be made by these models.”