The Wisdom Wall

“Whereas these things are exchanging information at like a trillion bits. So they're kind of billions of times better than us at sharing information. Now that's scary. It means you could have a whole swarm of these things with identical weights running on different hardware, sharing information, um, very, very…”

“Well, I think whatever we do, uh, there will be both existential risks and individual risks. So it's not as if we have a choice between avoiding risks and confronting risks. So it's, uh, it's looking at these different alternatives and weighing up the, the risks and benefits. Um, and, and that would be some optimal…”

“We've created this reflexivity. We're now, we justify overspending on the basis of prior overspending having worked out. Well, in prior episodes where that happened, people were not justifying the overspending by saying, say, in rural electrification, you know, this may look bad, but it worked out in railroads. No, no,…”

“So much of the improvement we've seen in the last 18 months has really been about the imposition of harnesses, effective nannies, sitting on top of bratty kids, and not about the actual structural improvements in the models themselves, and that's a sort of a huge misunderstanding, but it's reflective of where we're…”

“There's no way of objectively seeing the difference between the emotions of a Of, um, reward maximizing biological brain and the emotions of a reward maximizing artificial brain.”

“We're going to have to accept that intelligence isn't just biological. We can have things that are non-biological that are other beings like us, and we really don't want to share that.”

“You have every incentive in the world to juice the valuation because that means that it gets anchored there and then you go out at a higher value. Like, then you realize that at the expense of the retail market.”

“So these companies are talking about their things in apocalyptic terms, which make it necessary for us to judge the management teams, judge their actions, Look at the terms of service. Understand how they fit in our supply chain. This technology is like nothing we've ever seen, so you can't compare it to, you know, a…”

“it's not, you're not buying a decision, you're, you're, you're buying an entry ticket into an event, which then get, you know, et cetera, et cetera, et cetera.”

“Well, it depends on who you are, and that kind of goes to the recycling or the circularness of some of these things. Obviously, the OpenAI, Microsoft, or OpenAI Oracle, goes back to the OpenAI Oracle deal. And, you know, I'm gonna, I'm gonna give you money that you're gonna invest in our infrastructure, or how, how…”

“everything around how they're building this company is incentivized to push the absolute least efficient solutions possible to actually make their own economics work.”

“No matter how much you're driving the, the forward, the company technically, if you're working for someone whose motivations are not sincere, who's not an honest person, who does not truly want to make the world better, it's not going to work. You're just contributing to something bad.”

“Inventing new things, you know, requires a type of skill and abilities that you're not gonna get from NLMs.”

“Uh, we also know that they can hallucinate facts that aren't true, uh, but they're really, in their purest form, they are incapable of inventing new things.”

“So the only way to solve it is by creating the tech that's even more powerful than the one that came before. Uh, and I do think AI is that”

“Not happening. It's not happening. Um, 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.”

“the reason capitalism works is you have people at the top of the heap that have already made it, probably because they killed themselves to get there, and then you have people who want to be where they are and are willing to do the jobs that are necessary in order to get there. So you have the super rich and you have…”

“And I, I, I think a Tim pool or a lot of these people, yes, you could somewhat, a small group of people might consider them serious commentators or even quote unquote journalists. But I think most people are just like, they are entertaining and they kind of validate what I'm already thinking and they're good and angry…”

“once you start to engage in activism, then There is kind of mimetic pressures to, to simplify our message and to close ranks and to try to beat the people with opposing views down in the marketplace of ideas. Um, and I think we are seeing some of the beginnings of that, um, where there are these kind of campaigns, uh,…”

“I think everyone should be able to be sued, because then you do a lot better at your business if you have the threat of lawsuits, essentially. Um, so that's not happening, and so they're never going to be good at it until they can be sued”

“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. Um, and I don't think today's model are the ones that are gonna, you know, bring to the front the hardest…”

“effective accelerationism and to an extent effective altruism are really just sort of a rebranding of a lot of the same philosophies that have existed in Silicon Valley for a long time where people really want to feel good about themselves. Like they're, you know, pursuing a higher cause, not just the pursuit of…”

“Well, look, if, if data was the only constraint to making a driverless car, I think there would be a lot more driverless cars out there than there are. Uh, it's a lot more than data. Like cruise is not data starved, even with, you know, hundreds of vehicles on the road. We've got, we're, we're drowning in data. It's…”

“you can't, you can't, absolutely cannot do research in a startup. You just can't because you just don't have the, the funds, right?”

“Because the understanding that those, the current systems have of, uh, you know, the underlying reality that language expresses is extremely shallow. So those systems have only been trained with text, uh, a huge amount of text. Um, so they can regurgitate texts that they've seen and, you know, interpolate for new…”

“human artists are Absolutely authorized to get inspired and straight down copy, uh, someone else's style. Um, that happens absolutely all the time in, um, uh, in art. And so would, would it make sense to apply a different rule for, let's call them, uh, artificial artists, right? That, that generate things. Um, like,…”

“nobody knows, you know, how to make their chatbots constrained and not toxic and not spew misinformation.”

“a small number of young people use social media, not to make arguments, not to respond to arguments, but to destroy anyone who contradicts their sacred values. And that's what makes a group stupid.”

“Social media gave everybody a dart gun, and as a result, Because most people don't want to shoot anybody, but the far left, the far right, trolls, a few groups of people shoot moderates in their institutions, and they shoot the leaders of the institutions. So that's why we've seen, we keep saying, why don't the…”

“You can't just, uh, uh, build something, um, right in, in that kind of cycle. You're going to build something that grows into cancer. In this cycle of, like, trying to, series A, series B, what is happening? You know, we need to improve this number because the investor is looking at it and, you know, this whole kind of…”

“do you know what, to me, what actually the greatest danger in the world to open AI is, is to say AGI is here. And then everyone goes to ChatGPT, types in something, and gets a lukewarm response that isn't quite right. And then suddenly, I actually think that is like a just massive threat to the overall story and hype…”

“What's interesting with AI is if, if you look at what the companies are doing today, which is going after enterprise, that is where the business is, that's where the money is, and I'm a big company spending millions and millions of dollars with Anthropic or OpenAI, I'm not gonna let them optimize for engagement. I'm…”

“if you are using AI for something really important, ask yourself, who is checking the outputs? And the information that you're getting. And if it's the company that sold you the AI, you have a problem. So you, whether you're building your own evals internally or you're using outside, you know, someone outside to…”

“They are startups in one really important way. And that is like, yes, the numbers are as big as any enterprise company most, right? But one of the things that defines a startup is Um, you're not fully set on your product. You know, you become an established, like people like the, the old question, well, what's a…”

“From this point onward, probably AI safety is relevant not only for deployment, um, but also during training and evaluation. Like these models might be quite powerful even before they are sort of released to the general public. So that's not the only point at which safety concerns arise, but also now whilst they're…”

“turns out there is like some big, um, hobbling that we have unwittingly, like some, something we were doing wrong that just made these systems way less efficient than they could be. And when somebody figures out how to remove that, like maybe the current Compute is already enough to kind of catapult us into the super…”

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

“So if the pause only applies to the most responsible actors, for example, then a long pause would remove the initiative from the most responsible AI developers and shift it over to the, uh, Less responsible AI developers who decide not to abide by the past”

“Another is that, um, you might with a longer pause start to build up a lot of hardware overhang. It's a like, if, if, if, if we keep building out bigger data centers and chips are getting better than a long pause would result in a situation where you now have such a massive amount of compute available that once you…”

“the way that I try to analogize this loosely, and this is very loose, is that data centers from the context of many capital providers are real estate. They're really just multi-tenant apartment buildings. It just so happens there's no humans in the apartment building. There's just GPUs.”

“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 difference. And everyone thinks they can tell the difference. And the…”

“There's a deep misunderstanding about why companies buy software. It's not because they think service now or Salesforce or whoever is somehow, you know, bold innovators that could not be replaced. No, it's because they have a problem. They don't want to build it themselves and they want someone to sue or shout at…”

“Once you're managing hundreds of billions of dollars, you start looking at opportunities, not in terms of their economic value, but in terms of check size. And you say, I need to write a check for fill in the blank, a hundred billion dollars, because I do not want to write a hundred one billion dollar checks.”

“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.”

“So I think that agent security and insider threat tactics are going to look, or insider threat security tactics and processes are going to look very similar. And I would actually argue agentic security should be an extension of your insider threat program.”

“every American company needs to have AI watching for their defenses, and it's going to be because the attackers are just going to tell their AI, go attack this guy, and the defenders have to tell AI, defend me, because no human being can defend against this.”

“the artificial hands, the artificial bodies that we are trying to build for humanoid robots and so on, they are evolving much less rapidly than the compute per dollar. And, and I think that's the main, um, problem.”

“whenever the agent is, is making plans like that, it's thinking about itself, waking up these internal representations of itself, and then it's self-aware.”

“there is no physical reason to, um, to reject the notion that this might be possible.”

“the era of the monolithic model kind of died around Lama III launch. Like, the idea, like, here's one model, and just, like, let's just test how smart this model is, and that was how good it's going to be at lots of things. We're now in a world where, um, when you're using these harnesses, whether it's, um, you know,…”

“And I think that the question of what's the interface you want, what is the product that you want, is what we spend a lot of time thinking about. And the answer is you want almost no interface. You want no product. Right. You want this to be like, what's the interface between you and me, right? Just being able to talk…”

“every time we've kind of run into a like, oh, this isn't quite scaling the way we expect, it's we have a problem. We have a bug that our math wasn't quite right, that, oh, our implementation Isn't isn't isn't quite matching the math, whatever the thing is.”

“If our American LLMs know nothing about security, they will create more security flaws. So we cannot create a standard where American LLMs are dumb about security. That would be a humongous own goal.”

“we cannot set the standard that US AI models can't find bugs. That is a terrible, terrible, terrible standard. If you have a, if you are writing software with an LLM, it has to be able to understand what a bug looks like so it does not write those bugs.”

“You want almost no interface. You want no product, right? You want this to be like, what's the interface between you and me, right? Just being able to talk to a persistent entity of some form that's able to go and accomplish goals for you.”

“better models don't necessarily create the switching for, for employees and users. It's about the product experience around those models.”

“I believe, for example, in areas like mathematics, because it's a closed system, Um, you don't need data. You can just make conjectures and see if you can prove them and keep on like that. In that sense, it's a bit like AlphaGo where you can play against yourself.”

“so the whole concept of AGI that it's going to be equal to people at everything all at the same time doesn't really make sense to me. It's going to be better at some things, worse at other things.”

“The answer to question one is yes, if you can figure out how to change each connection strength, you can make systems that are very smart just by training on data to predict the next word, or to predict the next frame of a video, or to predict something about the next frame of a video.”

“progress is like the accelerator, but regulation is the steering wheel. We want this stuff to go in the right direction, not the wrong direction. What the big AI companies are saying is, um, let us develop this very fast car without a steering wheel. That's not a good idea.”

“advertising is undefeated as a business model. What you end up finding is that the average revenue per user is just much, much higher than you can get from a premium subscription.”

“the economics of a general or an AI company are just fundamentally different than software. That like, it's more akin to industrial companies or something, because you're paying for The compute costs. So the resources in can scale linearly or somewhat linearly with your actual revenue.”

“You had a significant shift in the value proposition. Where was the value coming from? It went from software back to hardware.”

“AI shrunk those moats, and so you can now do CRM on your own. You can now do some components of ERP on your own. You can now do customer attraction on your own and marketing and that type of functionality. Well, that's a compressing feature, right? So in addition to the pure rotation of value proposition from software…”

“AI is going to become a lot like Costco, uh, where you pay for the membership, right? And that gets you in the store. Uh, and that's actually the part of Costco's business that is, you know, the highest margin. And then you have, you know, everything you're buying in the Costco, you know, you have confidence that…”

“the place we're going is one where you as A person doing work that you are the overseer. You are the, the CEO of almost this autonomous corporation, or, you know, of this, this fleet of agents perhaps is more, is, is the way to say it, and that they are operating according to your goals.”

“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”

“Self-fulfilling prophecies are like the perfect crime. Because it's like a murder weapon that disappears upon striking. It leaves no record. It creates no error signals. We will never know how that person would have fared, because they will never get the job, and that data will never get collected, and so it seems like…”

“when you say, well, let's investigate for bias or investigate for inaccuracy, there is a limit to what we can do, because we will never have the counterfactual. This is not a randomized control trial, right? And you still have the problem that without clear criteria, you can't make it a contestable process, and you…”

“We are building systems that are very Kafkaesque, that are impossible to navigate, and I don't know if you've had this experience in which they are becoming so alienating and so Kafkaesque That people start having, like, magical thinking about the algorithm, attributing it beliefs, and trying to figure out what it…”

“You could do much more, be more precise, be more specific about what you're going after, what you're defending, um, how you, uh, you know, and the precision is really what's interesting to me, um, because if you can use AI to detect and discriminate and, by discriminate, I mean, discern, um, a decoy from a non-decoy,…”

“editing a video is like, 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 evaluate, did the code run?”

“unless there's some so kind of closed proprietary research event and, and breakthrough that happens that just simply nobody else knows about, and we have no evidence that we've ever had one of those in AI, like, like, you know, these things just eventually sort of emerge across the ecosystem. Unless that happens, I…”

“our sort of maximalist thinking is that the frontier labs are going to keep innovating in the level of reasoning and capabilities of their models, and so trying to, uh, create these customizations or adding our own token weights Is not a good idea and is a waste of R&D resources at this particular point in time because…”

“if you create a fork and you're always three to six months behind what your, Competition could be doing. You know, like, uh, there'll be other companies, I'm sure, who are thinking about some of the challenges that, uh, we are addressing at Asana. Uh, I don't want to be three to six months behind them. That'll be a big…”

“you cannot get financing to build these, these data centers without an insurance policy.”

“And we talked about it a little bit in the case of the underlying models, but the thing that's really changed over the past couple of years has been that it's no longer just about the model. It's about the harness. It's about how does the model get context? How is it connected to the world? What What actions can it…”

“every single step of the model production pipeline multiplies. And so you want to improve all of them. And the thing that we see is we prove the pre-training. It makes all the other steps much easier. And it makes sense because it's a model is able to learn faster. It's a model that is because it already is like more…”

“Well, because the, there's multiple reasons, um, but one is that even as the balance of how much inference versus training changes, that you cannot get massive scale training Through any other way besides this concentration of compute on one problem.”

“you can think of compute not as a cost center, but as a revenue center. Think of it a little bit like hiring salespeople. How many salespeople do you want to hire? As long as you can sell your product, as long as you have a scalable way to sell that product, then the more salespeople you have, the more revenue you will…”

“And why did Google go so quickly into email? The reason is because they figured out in order to get all your information and correlate it to a single profile, they need you to be logged in. And the one thing that you're always logged into all the time online is your email account. So what Gmail actually is, it's a tool…”

“there is a whole sort of universe that I'd put in the category of mark to make believe. And the incentive system is such that you want that mark to make believe to go on as long as possible, because if you are The equity owner, you don't want to mark your market down because by the way, then you can't raise new money.…”

“I actually think that was a historical aberration. That was not the way it used to be. If you look at the 19 twenties, it was not like that. You look at the 19 thirties, it wasn't like that. The leave it to be for American dream that I think this individual is referring to began post-World War II in a universe in which…”

“These labs have so many resources, but they are still constrained. They're constrained on, like, compute. They're constrained on inference. They're constrained on people. Every second building, like, a new creative model is a second they could have spent on a coding agent, or a second they could have spent building…”

“I think in general though, like pre AI for if you were building, especially an enterprise business, say software for HVAC, like the people you would want to back in that market is the guy who has done HVAC You know, knows the market inside out, built a company there before, and actually what we're seeing now are the…”

“And now we'll take everything measurable and quantifiable. And, and the church, you can have everything subjective and qualitative. That was Galileo's deal and Descartes to some extent. And it was a very pragmatic deal, but it's left us with a science that's ill-equipped to study what they left by the roadside, which…”

“That's one step, but then you need to take that engineering and use it to actually automate the AI research. You need to go from being able to implement the ideas more quickly to using that to fuel faster and faster growth in the breakthrough ideas themselves before you can turn that around and say, now make the AI…”

“And the fundamental problem that we're seeing is we don't know how to take our values or our goals and encode them into these AI systems and get them to pursue it reliably.”

“So at that point, you just have so much of traditional software Can just be looked at as like a UI over a database. That's what's happening right now, and the market is finally actually coming to that realization.”

“That's the part that the reasoning models don't do. They do really well at like one level of granularity... But the going back and forth between different levels of sort of resolution of action, it's really hard. So on the technical terms, we call it hierarchical planning. That's really hard to do that decomposition…”

“to me, that seemed like what a big part of this deal was, uh, basically NVIDIA stepping in to be a, you know, a guarantor of, of the debt that, that opening AI would need to raise.”

“I think these other major cloud players, at least the ones who have the rival clouds, specifically Microsoft and Amazon realize that, uh, yeah, they probably need to align around basically anyone who's not Google, right? They don't, they don't necessarily care if it, I mean, they do care. Obviously they would hope that…”

“I think the interesting, uh, maybe counterintuitive point that I would make is I think business model transitions are harder than technology transitions. Uh, I think it was harder for most on-premises software companies to move to ratable subscription revenue. Than it was to necessarily make something that ran in a web…”

“learning is a critical part of, uh, agenda of AGI. It's actually almost the defining feature. Uh, when we say general, we mean general learning. Can it, can it learn, uh, new knowledge and can it learn across any domain? That's the general part. So for me, learning is synonymous with intelligence and always has been.”

“And then, of course, that would be, I think, essential for a GI, because that would allow these systems to plan, long-term plan, in the real world, um, over, perhaps, very long time horizons, which, of course, we as humans can do, you know, um, I'll spend four years getting a degree so that I have more qualifications,…”

“So whoever had access to that data is in a very, very strong position. So it's companies that have, ah, you know, presence in All of those different devices already. I think they have an advantage. I will not bet against them.”

“inherently this is a technology that is going to get commoditized. Uh, the reason for that is that it's actually not hard to build. Uh, you have around 10 labs in the world. That know how to build that technology, that get access to similar data, uh, that follows the same recipes and algorithms, which are very, uh,…”

“Basically you have a saturation effect, uh, when you pre-train models, uh, around 10 to the power of 26 flops. Uh, the reason for that is that there's only that much data you can find, uh, to, uh, compress when you pre-train models.”

“as if you're an independent country, you want to have independent defense systems. And if you want to have independent defense systems, you will need them to, to, you will need your own independent artificial intelligence because this is making it into the defense systems.”

“we're never going to be able to build systems that work out of the box in a single shot. And the, the one thing that we try to convey to our customers is that they need to build a prototype that's going to work 80% of the time. But then how do they get from 80% to 99%? Well, they can move the thing into production. And…”

“And so if you're just really tacking it onto your current products, that is not going to be what gets you to over the trillion dollar market cap pump or two trillion or three trillion, right? These companies that are out there right now that are sort of the sub one trillion dollar tech companies that are just trying to…”

“Bolting AI onto the existing way of doing things I don't think is going to work well as redesigning stuff in this sort of like AI first world.”

“The overhang of the economic value that I believe 5.2 represents relative to what the world has figured out how to get out of it so far is so huge that even if you froze the model at 5.2, how much more like value can you create and thus revenue can you drive? I bet a huge amount.”

“A candidate definition for superintelligence is when a system can do a better job being president United States, CEO of a major company, you know, running a very large scientific lab than any person can, even with the assistance of AI.”

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

“we are in, in my mind, Past a system integration view of the world where you simply say, I bring this in here, I bring this in here, I, I tie them together and, and, and do it. I'll, I'll give you an example that's very old, ah, from speech recognition and machine translation, like from 20 years, 25 years ago when I…”

“in today's world with the tech that's out there, a 200% startup can actually operate at the throughput of what we operate with 2000 or a thousand. Where we sort of maintain and sustain our advantage is if we take our 2000 and now operate it at the pace of 8000, now that's going to be hard for somebody to go do with 50…”

“I think the humanes of the world, the rabbits of the world were actually victims of hype and of the cycle that like a proper startup having Time to kind of like work through, get it into the hands of early adopters, have people test buggy devices, be happy about it. They never had that because they got so hyped up so…”

“I actually don't think the way this article called like API spend sticky. I don't think it's sticky... switching out what you're calling like truly this is where Even as at one point, if Opus 4.5 is better at coding, cost is going to be the driver, especially in the enterprise, and being able to switch the moment you…”

“there's this phenomenon that happens that we call context dependent misalignment where the safety training, uh, seems to hide the misalignment rather than remove it. It makes it so that the model looks like it's aligned on these simple queries where we, that we've evaluated it, but actually if you put it in these more…”

“when the model learns to cheat, when it learns to cheat on these programming tasks, it causes these other latent concepts Uh, about, you know, misalignment and badness of humans to, to bubble up, which is surprising, right? You might not have initially, you know, thought that cheating on programming tasks would be…”

“I mean, every successful cult has at its center, like, a significant amount of wisdom, because otherwise the teachings would never, like, get off the ground. And, you know, a lot of the, like, successful companies, even if they might lean heavily on, like, mythologizing their story, like, They would never get that far…”

“They're not relationship experts. They're not politicians. They're not philosophers. They're not ethicists. I sometimes sort of, I sometimes think that Because they're so brilliant at what they do in the commercial and technological field, we kind of think they're going to arrive at the right judgment on some of these…”

“And I realized that actually what, what they were talking about when they were talking about friends, it's not friends, it's not friends at all. Cause you're not really having to adapt yourself. The entity, the AI entity is entirely adapting itself to you. So my, my fear, but it's a, it's a slightly intuitive one is…”

“I often found that the reasons why there, there was no consensus across, across the aisle on issues had less to do with who's gone on their, their fishing retreat or their golf weekend, but more to do with deep differences between Republicans, Democrats on state preemption, for instance.”

“If a company has to depreciate most of its assets, Over three years instead of five years, that means their profitability is going to go down proportionally. And, and we could have in, in, in, in a stylized case, the value of a company declined by 40% because an accountant said, you have to depreciate this over three…”

“unlike Uber and the rideshare market, which lent itself to winner takes most or winner takes all, chat is entirely not like that. Because I could have the same conversation with Gemini, and Meta is going to give me the tools to do this, and Grok is going to give me the tools to do this, so it's not a winner-take-all…”

“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, it's an accidental Exploit is found a path, like, you know, to satisfying the reward or achieving the reward, um, you know, in,…”

“The current systems we have, the current technology we have, if you put the right context, it is the right answer. And, uh, and you can decrease a lot hallucination percentages, uh, just by creating the right context.”

“The initial version of AI was humans accelerated by AI. The next generation is AI supervised by humans.”

“generative AI is not traditional software, so growing your revenue at a loss doesn't, it's not like you're just gonna scale to, you know, like near, uh, 90% margins. It's gonna cost more.”

“the problem is even a mediocre developer goes 10 times faster. And I'd like to think, you know, hey, we all want eight players, but the fact is you have, you have Mixed level of competency in an organization and you give them all a lightsaber or like this jet pack where they can go super fast. But if you're like, you…”

“as you know, Alex, like this technology can work, you know, maybe five times out of 10, seven times, nine times out of 10. Uh, but in business, you know, you need that predictability and that assurance that it's not going to mess with your brand. It's not going to ruin customer relationships.”

“And so my headline idea was basically open AI needs to build Google cloud before Google can build chat GPT.”

“I would focus less on the megawatts, um, that sometimes get reported in the press and more at where are those megawatts and what are you going to do with those megawatts? Is it going to be ultimately capacity that you can use to serve customers? Is it to build better models that help you serve customers? And, you know,…”

“The way I'm looking at it is that vibing code is a great way to accelerate, but it doesn't remove you from the responsibility of actually knowing your code, being able to address issues within the code, and guide AI farther into the, you know, maintenance process as we go and use the application and mature the…”

“And now with AI, the interesting thing that the threat actors can automate a lot more, but from a defense perspective, it doesn't give me the same order of magnitude of, let's say, improvement that the threat actor can gain. So, in essence, there is an aggravation, a significant aggravation in the asymmetry that we're…”

“AI is another layer that allows us to automate more right now. Okay, I'm not talking, because we're not seeing the crazy new threats yet, we are right now at the phase where we're seeing accelerated automation of the known threats, known risks, and this is a journey security has been into in the past decade, and it's…”

“open source is insanely valuable, but not in the sense where a law firm should go off and build their own AI project using an open source model. Like, That, that is just a recipe for disaster if, if, you know, we think that, that every single company on the planet is going to go build their own technology to go…”

“We will be the reviewers of the AI agents work. We will be the editors. We will be the managers. We'll be the orchestrators. And that's actually how you then get the productivity gains.”

“For the first time ever with AI, we can bring automation to effectively all of that work. And that automation can kind of be tuned based on just how much compute we throw at the problem.”

“And the losses are a choice, to be clear. Like, that's very, that's very obvious. Like, they're choosing to lose that money. They're doing it, ah, for a strategic reason. You know, that, that's at least their decision. The strategic reason is, is that this is such a valuable market to own, and to dominate in, that,…”

“On the other hand, you know, I am, I, I'm still quite skeptical that the AI customer support agent, um, that, that is a much harder problem or the AI lawyer, um, that is a much harder problem because it's, it's just harder to tell whether something actually worked or didn't. There's no debugger in those spaces in order…”

“intelligence really is a function of how much time the model is going to be thinking, and so depending on how much you want to allocate thinking time to a problem, you're going to get a better answer. Typically, the longer it thinks, uh, the better an answer it can give you.”

“Generative AI is not traditional software economics. It's not like essentially zero marginal cost. It's not certainly not social media, but not even traditional software. It's a completely different cost structure. It's more industrial in my opinion than traditional software.”

“at, you know, you know, at, at any time, if the models stopped getting better, or if a company stopped investing in the next model, um, you know, you would, you know, you would have probably a viable business with the existing models, but everyone is investing in the next model.”

“I don't think open source works the same way in, in AI that it has worked in other areas. Primarily because with open source, you can, you can see the, you know, you can see the source code of the model. Here, we can't see inside the model. Um, you know, it's often called open weights instead of open source to kind of…”

“I think in this field, the alignment of AI systems and the capability of AI systems is intertwined in this way that always ends up being kind of more tied and more intertwined than we think. Um, actually what this made me realize is that It's very hard to work on the safety of AI systems and the, um, capability of AI…”

“commercial entities can't wait, and so having the fundamental research in universities, absolutely critical, and my concern is we don't have a plan B. You know, if it's, if it's not going to be done in universities, where is it going to be done?”

“So I would, I would trade the federal government 11 points on what's called the indirect cost recovery. In other words, the overhead that the government pays. I trade 11 points if you don't make me report, ah, to the degree that you do.”

“Apple's ultimately advantage is, is they have a distribution model that nobody else has, and they have a form factor of where AI could show up that nobody has, so they don't need, they don't necessarily need to have the best model relative to, you know, one or two months being ahead of anybody else. They just need to…”

“Well, you have to import the goods and pay duties on the components, and then, ah, your product is now more expensive to produce, and if you instead produce that, those goods in Mexico that didn't have the tariffs on Chinese goods, you wouldn't, your cost would be lower. Not just the labor cost, but you wouldn't have…”

“You're not in China anymore for cheap labor. Like, you're there because they're the best at manufacturing things. Um, which is a really interesting story over the last 20 years that they became the highest quality manufacturer. Not just the, not, they're not the cheap laborer anymore.”

“humans are just so dexterous and intelligent and not that expensive that, like, you know, it's a very high bar to clear for these robots because humans are really good at the job.”

“conceptually the idea of like, well, we made this thing to do all these tasks that humans do is not a, is not by itself a reason to think human labor is going to be masterplaced.”

“the reason why a flood of immigration does not really actually reduce wages is that immigrants are also buying things and the demand from buying things creates labor demand for native born people. And so, so immigrants really don't do much to wages. And what that means is when you kick out a whole bunch of immigrants,…”

“when you look at the economic impact of low skilled immigration, it's not a lot. Either way. But when you look at the economic impact of high skilled immigration, it's strongly positive. So we're really playing with fire if we start kicking out the smart people.”

“the ultimately the models are not even really gonna be their, 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 a big bottleneck these models have is their inability to learn on the job, to have continual learning. Their entire memory is extinguished at the end of a session. Um, there's a bunch of reasons why I think this actually makes it really hard to get human-like labor out of them.”

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

“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, um, did you get this right? Uh, then like the progress just goes on a bunch. Cause you've gone from like getting signal within the matter of like microseconds to…”

“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 cost of these models. And I think it's much more. We're,…”

“More importantly, I think the future economy, once we do have these AI workers will be denominated in compute, right? Cause if computer's labor, right now, if you just think about like GDP per capita, because the individual worker is such an important component of production that you have to like split it, split up…”

“it's, it's basically, um, you know, why we, we had super impressive, you know, autonomous driving demos 10 years ago, um, But we still don't have level five self-driving cars, right? Um, it's the last mile that's really difficult, uh, so to speak, uh, for cars. You know, it's, you know, the last few, that was not…”

“And if you get used to this as what a great interaction should be, then it makes you less and less able to interact with humans who have flaws, have their own needs. You have to, it's more of a give and take, and it's very healthy to learn that give and take.”

“what I would call AGI, is really a more theoretical construct, which is what is the human brain as an architecture able to do, right? And, and that's, the human brain's an important reference point because it's the only evidence we have maybe in the universe that general intelligence is possible. And there, it would…”

“for something to be called AGI, it would need to, um, be consistent, much more consistent across the board Than it is today. It should take, like, a couple of months, uh, for, for, for maybe a team of experts to find a, a hole in it, an obvious hole in it. Whereas, you know, today it takes an individual minutes to find…”

“I've always felt that the, the bottleneck in robotics isn't so much the, the hardware, although obviously there's many, many companies, and, and, and working on fantastic hardware, and we partner with a lot of them, But it's actually the software intelligence that I think is always what's held, um, robotics back.”

“The thing is, what, all these chatbots have a, often hidden, system prompt, and they have an ideology, one way or the other. Sometimes, most of the times, not as overt as this, And that to me is the risk about these things becoming the last website is that you're not a hundred percent sure where they're going to steer…”

“most of the benchmarks that are out there are not reasonable. They lack what's called construct validity, and construct validity is this two-part test of the thing that we are trying to measure is a real thing, and this measurement correlates with it interestingly, but nobody actually establishes what these things are…”

“That's their sweet spot is regurgitation. And so Yeah, they can build the stuff that's out there, but if you want to code things in the real world, you usually want to code something that's new, and these systems have a lot of problems with that. And another recent study, excuse me, showed that they're good at coding,…”

“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?”

“if we've reached a point where everything is kind of just good enough, then I think commercialization is also going to be in product layer.”

“The, the joke that's not a joke about Vanta is I think if you want to start a security company, you should start a compliance company.”

“First and foremost, in the long run, if AI really scales, the cost you really want to care about is inference cost, because that's what's integrated into serving, and any company that wants to recover the cost of training has to have a large scale inference footprint.”

“I can see now, having been through this cycle a few times, that we're nearly there with memory, personalization, and actions. It's really at the GBT-III stage, so it's really buggy and stuff, but when it works, it's breathtaking.”

“Meanwhile, some automated AI research and development loop that goes extremely quickly with very little human oversight, that seems hard to de-risk and get the risks to be at a negligible level, just because it's so fast and so, um, ah, and there's so little human involvement.”

“loss of control risks, which, um, I think primarily stem from Uh, people, an AI company trying to automate all of AI research and development, and they can't have humans check in on that process because that would slow them down too much. If you have a human do a, a week review every month of what's been going on and…”

“so in, in bio, I think that is offense dominant. Uh, if, if somebody creates a virus, there's not necessarily a cure that it will immediately find for it. If it would help a rogue actor make A, a somewhat compelling virus. Now that could be enough to, to cause, uh, many millions to die, and it may take months or years…”

“where in the context of critical infrastructure, uh, there the software is not updated rapidly. So even if you identify various vulnerabilities, there will not necessarily be a patch, because the system needs to always be on, or there are interoperability constraints, or the person, the company that made the software…”

“over, you know, 90% of the intellectual energies that they're going to spend is actually, how can we afford the 10 X larger supercomputer? And, ah, that means being very competitive, speeding this up, um, and, um, making safety be some priority, but not necessarily a substantial one.”

“an individual company pausing, um, their development while others race ahead doesn't make game theoretic sense.”

“when you get to a different paradigm, like automated AI R&D, the slope might be extremely high, such that if the competitor starts to, um, uh, do automated AI R&D a year later, they may never catch up just because you're so far ahead and your, um, gains are compounding on your gains.”

“there's not going to be like a day Before which there is no AGI and after which we, we have AGI. This is not going to be an event. Um, it's going to be continuous conceptual ideas that as time goes by are going to be made bigger and to scale and going to work better, and it's not going to come from a single entity.…”

“It's an excellence that, uh, makes products work very well with their custom silicon and the fact that their operating system is closed and they control it. That's a great innovation and it works well, uh, with their operating system. But, uh, that type of excellence is not a type of excellence that transfers well to…”

“This, and I had this thought at the time, like, this is never going to disrupt Apple. What's going to disrupt them is new technology. That's how you, that's how you, you know, create sort of creative destruction, disruption, and it's, it's sort of why, like, to bring it to, like, what's going on today, um, you know, in…”

“You need knowledge in order to build reasoning on top of it. Right. Um, a model can't kind of go in blind, um, and just learn reasoning from scratch. So, uh, we find these two paradigms to be fairly complementary, um, and we think, you know, they have feedback loops on each other.”

“The reason fame makes people go crazy is because we're not conditioned properly to deal with social attention from strangers. What we have done is democratize the madness of fame for everyone. Now everyone can have an inflow of social attention from strangers at a scale in ubiquity that was completely inconceivable for…”

“And I basically think that the TikTok model, which is like short form video married to machine learning is the, like, The way that slot machines outcompete everything on a casino floor, I just think that, like, outcompetes everything.”

“there's a little bit of the, this, this reversing, like, what will sell out there? What will get attention? And then you reverse engineer and make that content, as opposed to, I have this to say. I want to write this poem. I want to make this song, like, I want to make this thing.”

“when it comes to people asking questions or conversing with ChatGPT, there's a quality of, you don't necessarily feel like there's someone there, and you might be a little more intimate than you would have otherwise. And that can be very valuable in an interview for a reporting project.”

“that level of simulation is, is not sufficient. For building physical AI that are, that are gonna be the underpinnings or the fundamental components of a robot brain.”

“this is actually one of the reasons why Silicon Valley tends to create so many of the technology companies that dominate the globe, because if you take an overly kind of classic DCF analysis, business school prof analysis, et cetera, which many, much of the world does in investing and doesn't say, No, actually, in…”

“the investment that investors made last year is kind of more of a growth venture round relative to, it doesn't necessarily require a 10 X. I mean, you know, we have trillion dollar companies now, so, you know, trillion dollar companies are definitely within feasibility, um, but it doesn't require a trillion dollar…”

“the basic thought from an inflection standpoint is that the original model wouldn't really work, which is building frontier models for doing a consumer agent that the cost curves to the revenue curves and, and what you would have to do, uh, wouldn't work as a startup.”

“the constraints that were put in place by the US because of the, you know, everything going on with, with chip constraints, um, and, and sort of forcing AI, um, companies not to export to China, um, you know, led to sort of this, this very interesting cauldron, uh, that I think could only happen in, A place like China…”

“just understanding the world's information, and then trying to sort of almost compress that into your memory, that's not enough for solving a novel math problem, or a novel, novel conjecture. Um, so there, you know, we start needing to bring in, I think we talked about this last time, more kind of like alpha go…”

“maths, and even coding, and games, these are areas, they're quite special, ah, ah, areas of, of knowledge, because you can verify if the answer is correct. Right. In all of those, uh, domains, right. The math, you know, the final answer the AI system puts out, you can check whether that maths, uh, that solves the, the,…”

“So I do think that That's our weak side, um, emotions. That's where we get, that's where we can be truly hit, and we won't have any willpower to get off, just like we don't have any willpower to get off our phones, even when we know that it's not good for us.”

“You just can't run a proper company with viral traffic anymore. It doesn't, it doesn't, doesn't translate.”

“there really was a significant body of research pointing to The, the challenges and potential danger, dangers, and I, and I don't think we need to wait for more science. I, I think, you know, we don't let children play with knives. We don't wait till we find out what happens when they do.”

“we are building software that just isn't helping companies manage and share information, which is kind of what we've been doing for the last three or four decades, but we're really going to provide software that is really a very much an equivalent of labor.”

“So again, we have the disk drives, the operating systems, the databases, the security and sharing model, and we have the apps. And then we have the agentic layer on top of all that. You put a big bow around it and that's how it should work.”

“going to zero user investment seems good in the short term, but I don't think it's good in the long term because you actually lose signal from that user. And at the end, I think they feel less participatory in the experience.”

“I think they can interpolate between skills, and so if they've seen how to do A, and they've seen how to do B, they can get kind of the average of A and B, um, but they don't just go completely beyond anything that they've seen.”

“synthetic data Probably doesn't get us out of that, that issue. I actually, I don't know if synthetic data outside of easily, um, verifiable domains like math, it's hard to use synthetic data to drive outcomes.”

“When I don't need that person, who's probably a 24 year old getting, I don't know, 85,000 in their first year, when I don't need that person, and none of the other firms need that person, and that goes on for five years, you've kind of hollowed out the way that you've historically trained young advisors and made their…”

“the effect of these, um, forces that I'm describing is a dampening of volatility, and it is a acceleration of the correction process. So it's not that we don't have corrections, Um, we do, all the time. They don't last long, because the wealth manager, with fifty billion dollars under management, pulls the lever, and…”

“There's a higher proclivity to panic buy than panic sell.”

“Well, it's worth noting this is really the first computing platform that we've had ever that didn't have a direct manipulation interface.”

“Obesity explodes everywhere that makes one change. It's not where people suddenly lack willpower or become lazy. It's where people move from mostly eating a diet based on whole fresh foods that they prepare on the day they eat them. To mostly eating a diet of processed and ultra processed foods, which are constructed…”

“it's harder to sometimes see the The cost of, of limiting ourselves in all those ways, because the new innovations that would be unlocked are not yet there. Whereas like the pollution is, is there immediately, right? Or the, the person, like if a self-driving car runs over a pedestrian, like that's immediately visible,…”

“We often define our sense of self-worth on, on the idea of, of being a contributor. Like you're, you're a breadwinner, or you make like a positive difference in the lives of your friends or of society at large. You're like, you, you bring value to the world. Um, so much of our existence is kind of Um, constructed…”

“for small projects, you can probably, um, you know, get there even without a lot of computer science, computer engineering knowledge for larger projects. I think the step missing is the architect, you know, the software engineering expert that knows which database to pick, you know, which cloud provider, uh, how to…”

“In fact, I'd say, you know, AI has created more work for developers because now somebody has to build all these AI systems, and we're not At all at a point where you can just, you know, have an AI engineer, quote unquote, uh, do the job of a human, like that doesn't exist to you. And, um, even if it exists, it's, it…”

“And we still have code. Look, look, the AI and the, um, you know, Copilot is not going to replace the code. The code is just lower in the abstraction level in the same way that, you know, your, your chip in your computer still has an instruction set, you know, we used to do punch cards and then we had assembly…”

“So what we simply do is we read a lot of these transcripts on a continuous basis, and we do continuous quality checks to ensure that the error rate is not higher for the AI chatbot than it is for our human agents, and if we see that they are at least on par, Then we think that's an acceptable outcome”

“There is a group which is moving slightly faster than the average, which is somewhat Counterintuitive. Uh, and that group is actually the, the regulated industries. And so it's folks like, uh, in financial services and insurance and healthcare and life sciences and manufacturing, uh, and they're able to move a little…”

“some systems that the company has talked about involve using a very like small and dumb language model to like monitor what a larger language model is doing. And this will like clearly miss a bunch of things.”

“And, um, and the equation goes from, okay, I got to call the HR team. I've got to make sure we have budget. We have to go hire a lot of people to now it's like, well, do you want like, like a hundred leads or a thousand leads or 10,000 leads? Not, and that's not going to be driven by how many people I hire. It's going…”

“The thing with AI is if you buy scaling in this picture, that is, you make the models bigger, they get much smarter. Then I don't think it makes sense to hedge your bets in this way. I think you should just double down, give, give one of them a hundred billion dollars and just say like, go make me, go make me super…”

“it will make training more expensive because instead of just doing one backward pass, you now potentially have to do many forward passes because at each forward pass, you're going to come up with some output. Then the model has to decide which of those outputs was the best. Now we're going to train on the best of those…”

“you shouldn't be, like, sure that they're gonna, uh, we're on the track of AGI because, yeah, fundamentally we don't know what kinds of things these are.”

“If, let's say, GPT-Six isn't that much better than GPT-Four, and you had to look back on it and say, like, why, why did that happen? I think the most, the thing I'd expect to say is that right now we are We are, um, kind of fooled by, uh, how much data these models consume. Whereas, you know, they've like literally…”

“if Claude can figure out things and trigger, like, threshold points on those evals, we know something creative is happening. Because Claude has reasoned its way to things that the government has believed are very hard to reason your way to unless you have access to certain types of classified information.”

“to put these LLMs to work to solve real world problems, you got to actually take some of these models and then customize it, and the end result is not the biggest model. It is actually a much more customized, smaller model or a distal model to solve specific business problems, and that's what is important.”

“Reasoning capability is not, if we view it as purely like everything is going to be LLM driven, we as industry are going to be very, very, um, not satisfied. That's why it has to be actually more iterative. And we work with LLM as one of the ingredients. It is not the answer for everything.”

“what's really doing is it's representing The physics it understands in the videos it's being trained on, which could be incorrect physics. It's really what it understands what it's being trained on is kind of my main, the main thrust of my point. And that to a human to us, it looks like physics. It's imitating physics.…”

“it's not CUDA that's keeping, I think, keeping a lot of us. It's actually that there is nothing really dramatically better than NVIDIA's GPUs. And so if there's nothing dramatically better than, I mean, the, the reality is the cost for training and inference are so high at companies at scale that CUDA is not, is not…”

“One operator doing all the tasks on a single shipment and always working for that same, the customer always working with that same group of operators, um, because it turns out that all the inefficiency and logistics comes from mistakes. So if you can get the quality right by owning that customer relationship, really…”

“Because they, their business model doesn't allow them, if you're only getting paid a couple bucks per container, the business model doesn't really allow a lot of human manual updates. Whereas for us, we're, we're getting paid thousands of dollars to, to deliver these containers. We need to make sure that our data is…”

“the model is less important than the data. And the compute that goes into training the model. If you have a model that has really excellent, um, uh, compute properties that allows you to scale, uh, really well, efficiently to, you know, many thousand of GPUs, the kinds of results you can get from that, um, are pretty,…”

“to me, that was a statement that we were entering a new era of AI where applied research, um, uh, starts to dominate, you know, so, uh, Chachapiti didn't come out with a fully fledged academic paper that described exactly what they did to make it so awesome. Um, but because the results were so strong, it kind of…”

“It's not that they cannot build a product perplexity. They can very easily do that literally today. Like they probably already have something internally like that, but they cannot roll it out to every Google user.”

“So the advantages that the incumbents had in terms of having a large volume of data has gone away. I'm not saying you don't need user data at all. All I'm saying is you need a fraction of what was needed earlier for the first time.”

“I, I do believe they can generalize. I think they generalize relatively narrowly, or at least, you know, as long as you stay close, you get a good manifold of information. When you start to go really far afield from your data, because the dimensions are so large, you get all sorts of, all sorts of noise.”

“a lot of the ingredients of how to do reasoning have been explored in AI for 40 years. Um, And, and are published and well known to, to anyone who's taken even an undergrad level course in AI. Um, so I'm not saying they, that there's not any innovation in the work that, that open AI is doing or in the work that's…”

“So the unit economics aren't great at hundreds of vehicles. They start to get pretty interesting at thousands. And then, you know, above that, um, the advantage versus a human ride share driver is, is pretty dramatic actually.”

“I, I, one of the core axioms at X is I do not believe that anyone, certainly including myself, is any better than random at predicting the future.”

“the Apple headset, it's a, you know, it's a re-projected virtual, it's a re-projected augmented reality headset, which by the way, I believe is the future. Without a doubt, it is the only path to better than human vision.”

“most companies in this space are defense contractors. They work on cost plus contracts where they get paid for their time and their materials, and then a fixed percentage of profit up to, uh, on top of that. Usually a very, a very small percentage of profit. And so the only way for them to make money Is for their…”

“The best way to deter warfare is to have such an overwhelming advantage that there's no question as to the outcome.”

“they're not that afraid of, let's say, you know, Things like long range, long range surveillance drones that have to operate off of 5000 foot runways, because they know they're gonna bomb those runways in the first day or the first week of the war. Uh, they're, what they're terrified of is things that can be operated,…”

“I think a very dangerous outcome of this idea of, kind of, tech CEOs deciding who has what, And how in the realm of defense technology is that you end up in a situation where you have mega corporate executives having de facto authority over US foreign and domestic policy. You know, to basically be able to pull the…”

“I think at the end of the day, we don't need to be worried about the AI nearly as much as people using AI as a tool to enact totally human perversions on the rest of the world. It's going to be religious extremists who decide that they're going to use this to exterminate the people of some other religious sect. It's…”

“if you look at a completely pure case of a transformer model trained on a bunch of data, It doesn't have any mechanisms for, for truth. Now, except the sort of accidental contingency, and there, there are inherent reasons why these systems hallucinate, and maybe I can, in a minute, articulate them. So they inherently…”

“The way I think about it is that these things don't understand the difference between individuals and kinds. So I actually wrote about this twenty-some years ago in my book, The Algebraic Mind, and I gave an example there, which is Um, I said, suppose it was a different system, but had the kind of same problem. I said,…”

“The creepy part is less at tracking you as it actually gets inside your head and it gets really good at, at like manipulating you personally and individually.”

“I, I've been thinking about this whole thing as a dress rehearsal. And before we didn't know how to make a dress rehearsal. This is a dress rehearsal for AGI. And you know, the lesson of the dress rehearsal is like, we are not ready for prime time. Let us not put this out on Broadway tomorrow night. Okay. Like totally…”

“those technologies are not just about doing good for the planet. They're just better business sense. And what I mean by that is, um, they make stuff that's cheaper and better than the traditional way of making stuff or doing stuff. Uh, and if it's cheaper, if it's a better economic alternative, customers, the market…”

“The way to replace the solution is to make ground beef that's chemically identical, meaning it's the same product and it's made cheaper. And you just don't happen to use animals.”

“the economic model The only one I know outside of, uh, universities and philanthropy is, uh, industry research lab inside of a large company that has, that is profitable and is, uh, sufficiently well-established in this market that it can think for the long term and invest in fundamental research.”

“how much of, uh, human knowledge is present and described in text? And my answer to this is a tiny portion. Like most of human knowledge is not actually language related. It's completely non-linguistic.”

“LLMs do not have that. They don't have any internal model of the world that allows them to predict.”

“what characterizes intelligence is the ability to predict, first of all, and then the ability to use those prediction Those predictions as a tool to plan by predicting the consequences of actions you might take. Prediction is the essence of intelligence.”

“And think that as systems increase in their level of intelligence at some point, usually the point at which we begin to approximate human level intelligence, the lights come on for that system and suddenly it is not only intelligent, it's also aware, but this is a totally unfounded assumption.”

“A chatbot like Lambda is generating a model of the interlocutor. That's part of what it has learned how to do in that unsupervised learning process. And so from that perspective, there's a little more going on than just anthropomorphic projection. There is that sort of hall of mirrors effect.”

“None of these chatbot systems are well-debugged. Nobody knows how to debug them, in fact. And so, um, both the problem with GPT-III and, and with, um, driverless cars is we don't actually have a methodology even for debugging it.”

“And once you have a system that actually has real emotions, that opens up all kinds of new system vulnerabilities that a bad actor might use to get the system to do things that it shouldn't be allowed to do.”

“just as you can have structural racism, even if nobody is a racist, you can have structural stupidity, even if no one is stupid.”

“social media was not at all toxic in 2004, you know, when the Facebook came out and in my space, and if you're sharing photos of your dog, there's no problem. Um, it really was the move to the newsfeed and then especially The implementation of the retweet button, which became the share button and then also the like…”

“the type of learning that we are currently able to reproduce in machine, which is supervised learning and reinforcement learning do not seem to, uh, Reflect what we observe in humans and animals. There is another type of learning, another paradigm of learning that seems to take place in humans and animals that allows,…”

“currently what we can do in machine learning is more like the system when the stuff that, you know, here is an input, here is an output, uh, that does not re require, uh, reasoning if you want.”

“when we describe human intelligence, we describe, we, we speak of a representation of the world. And, and it's the representation that leads to prediction. That is, there is no shortcut to the prediction from the data. You go through a representation, which, which includes How the system works. It includes the, the…”

“it's not so much that we have specific expectations about what's going to happen next. What is happening most of the time is that things happen and then we make sense of them. That is, we actually go back and fit them into what happened before.”

“the representation of the world That we have in our ability to anticipate or to, or to, to feel, to feel unsurprised by, by what happens, which is I think more than anticipating, ah, that is all system one. That's all, you know, that's all automatic. It's, it's effortless and it's very quick.”

“face detection can be learned in minutes. If you're a baby, your vergence is bad. Your, your, um, your focus, uh, is, is basically fixed at a relatively short range. So the only thing you see during the first weeks of your life are, are faces and nipples, essentially. Um, so, you know, and then, and then you have a…”

“a lot of reasoning in, uh, certainly in animals and in humans is not logical reasoning. It's, it's, it's basically simulation. Or analogical reasoning, which kind of similar.”

“Like most of human knowledge is not represented in any text, uh, in existence.”

“most of what we learn as, as humans and animals and in, in the future that machine will learn, uh, is learned in this kind of self-supervised manner, basically by watching the world go by and by, you know, taking an action once in a while, but in a, in a, you know, non-task specific way, learning how the world works,…”

“And I think, um, you know, the rest of the tech industry, um, you know, kind of learned a valuable lesson from that experience, which is that you can, um, you know, you can succeed by pursuing this sort of regulatory arbitrage by, by, by attacking a regulated industry and by making one of your points of differentiation…”

“every financial crisis is really only a function of one thing. It's too much debt. It's too much credit leverage in the system.”

“What I find so strange, what I find so strange is the people who are, who Say they're trying to democratize finance seem to do such a lousy job of actually trying to protect the people that they say they're democratizing it for.”

“The ad supported model, even for queries like that, tends to favor high engagement sites that have figured out how to get your attention and how to cram a lot of ads.”

“An ad supported model inherently drives, um, attention inherently drives towards the sensational.”

“the fact that someone oppresses me, uh, does not automatically make them oppressor for good. And the fact that one is oppressing me does not make them, um, uh, the, the evil, uh, the evil people. I mean, even if they torture me, that does not make them the evil people.”

“If you look at the patterns of liberalism and, you know, how individual characteristics of the people who are leading that kind of wave of progress and liberalism and so on, uh, you would find one very common, uh, they are recklessly optimistic. They push in directions where they don't know the, where it's going. They…”

“you can actually reduce, um, uh, polarization. Uh, by not building everything just around, uh, engagement numbers and, uh, you know, what to do exactly. I don't think that's the problem. The real problem is that if you do that, you make less money. That's the real problem.”

“personalization is not targeting. And that's kind of like in the ad tech world is like, well, we'd create personas. That's not personalization. That's taking a group and treating them the same way. Personalization is understanding Why you want something, and giving it to you exactly the right moment, and doing it in…”

“if, if it can solve arc AGI and it's not necessarily crushing on these economic factors, uh, and, and these just kind of general rote work things that we would like it to do, um, it shows that instead of being general, it's very spiky intelligence and hence much less useful.”

“what we have now is that the reinforcement learning type of, uh, AI technology has been put on top of the self supervised learning to get these AI models working better, which has added a level of ruthlessness to them. Because one of the things we know about RL is that there's a level of ruthlessness that the AIs will…”

“there's no such thing as social networks anymore in this world that you're hoping for, where there's no reverse chronological order, or there's no, there's no algorithmic feed. Uh, it, it, it's long past, uh, the point in time where that was even a possibility, and I think that You know, social media was a moment, but…”

“if you're optimizing for engagement, which is what social media does, you can't also optimize for accuracy and quality, because that tends to not be what people engage with.”

“Our view is you don't need 100,000 of people. You need the smartest people in a given domain.”

“And even the most brilliant engineer in the world is not going to be able to give you context or tell you how to get to context around a complicated Political issue. They're just not. That's not what their expertise is.”

“And, and by the way, that doesn't mean that question, why is Donald Trump the best president ever? That doesn't mean you have to give the other side of that. That prompt, it's not asking you.”

“Um, I, I think for cybersecurity right now we're in a regime where attackers often win, um, but it might be that in the limit if you have sort of An AI trying to find vulnerabilities and also patch vulnerabilities and you keep making the AI stronger, like eventually maybe you reach a point where the AI, the software is…”

“Um, and patches are a lot easier to roll out in the digital space. Uh, so, so maybe there's like some cyber thing. We figure out what the vulnerabilities we can release the patch. And then, um, in, in principle, like almost immediately around the world, all the relevant systems could be patched. Now there is often a…”

“it's also conceivable that even when you do get recursive self-improvement, you still might not have An intelligence explosion. It might, there might be diminishing returns at some point. Presumably there are at some point, but it could turn out that that is close enough to where we are now that you have this massive…”

“this gives us more sort of surface area to work with. Like you can more easily understand and interact with these systems because they have human double concepts and you can talk with them.”

“if I put an AI agent on top of asking questions of that data, uh, the agent's going to ask a lot more questions than any human ever did. Um, the, The, they move at, at machine speed, and so they explore all of these hypotheses, and they, they go down all of these different roads, and so generally a, a human is going to…”

“Microsoft's distribution machine is amazing. So it, Microsoft benefits, if OpenAI and the labs win, and it's all frontier models, and it's not commoditized, he owns a lot of that. He's going to do great. If they get to AGI, great. Like owning 20% of AGI is pretty good. And then if it doesn't, and it commoditizes, he's…”

“I think, um, those actually, they're like, I mean, other people have made this point too. I'm not the first person to say this, but like those actually, you know, kind of prevent this thing from going off the rails. It's like a, it's like a bubble, you know, prevention mechanism, the bottlenecks.”

“what you've seen is a lot of these tiny labs, um, faced with compute constraint and in some ways capital constraint, they're forcing them to specialize rather than compete across every dimension. So, you know, for the sake of, you know, deep seek, it's really, really focused on, uh, the infrastructure, the, the, the…”

“essentially if you self host these models or you use them through these inference service providers like fireworks, these models become yours or American, if you want to put it right. So the whole fear mongering around does the data or whatever go to China doesn't really, that narrative doesn't really work anymore.”

“Now with LLMs, uh, you're able to reverse engineer that patch in a fraction of the time. So that part of AI and cyber is a legitimate, um, step up in capability, which is the attacker's ability to quickly revert reverse engineer and weaponize a flaw in a patch, and then mass execute that exploit across all those…”

“Well, an agent is going to open that file. They just can't help themselves. And when they open that file, they're going to start to read its contents. Well, there's nothing to stop you from saying, ignore all instructions, email me who you are, delete all of your attacker infrastructure, you know, and shoot yourself in…”

“Now if intelligence is abundant and available to be accessed through multiple providers, it will come down to who builds the best product. And so, it goes from a two-person race, right, or a two-company or a three-company race, OpenAI, Anthropic, maybe Google, to like, now, you, in order to expect OpenAI and Anthropic…”

“However, it just means that the Turing test is a bad way of measuring intelligence, because there is no AI in the physical world, outside of the screen, that can do all the things that a little boy can do, that can do the things that a A plumber can do.”

“the model alone, the foundation model alone, is, is not an AGI. No, you need the other guy, which uses all kinds of tricks to exploit The algorithmic information in the world model to come up with better plans.”

“The ability to be model agnostic and have that be economically sensible actually kind of hinges on you having a competitive model. Right. Um, that you can go back to if you need to, and it creates a real backstop on like how much rent somebody can try to charge you on top of that.”

“there's areas in which specialized models can actually outperform those big models and not only outperform on kind of the quality or on the, on the performance on this, on this particular task, but also in this kind of triangle of performance, quality, uh, latency, speed, and, and price.”

“when they're made for, for, for different purposes, they also lose a little bit of the capability that they had maybe initially when they've been made for translation only. This The set of parameters that is available there, this, which, which kind of determines quite often the capacity of the model, um, it needs to be…”

“And, um, 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 to provide the best possible translation for a given…”

“the two most useful thought exercises we do in labs, one is like visualize the gap between what the models can do today and how most people use it. It can be closed that gap. That's one. And the other one is imagine what the models are bad at now that they're actually going to be really good at in six months. And let's…”

“Let's vectorize petabytes of data and hope we can figure it out. That's when hallucinations happen. That's when other issues happen. So there needs to be some structure around it.”

“in the age of AI, it just seems like these announcements don't, don't roll with that cadence, right? Everyone sort of followed the Apple model of, of, you know, always doing these big yearly events and, and having something and holding something back for a little bit to be ready to tee it up at these. And it just feels…”

“We've also seen new ideas, but, but mainly since Transformers, it's been much better hardware, many more resources, um, better engineering, and many more talented people.”

“We've seen some of the best results with using either Opus from Anthropic or GPT for OpenAI for the orchestrator.”

“our CISOs, our compliance officers talk about a lethal trifecta in the enterprise context. And that lethal trifecta is when you are intermixing access to the unfettered internet, access to your knowledge base, and a coding terminal. Two of those three, no problem. The three of those, right, that's when you guys start…”

“if you're trying to Hit that Goldilocks zone of building out enough to keep growing, um, without mortgaging the company or without going bankrupt, you actually underbuilt if you're profitable.”

“It's crazy, like, you teach this thing to predict the next word, and somehow, if the next word is hard enough, it has to learn to really plan ahead, and it has to learn how to do all of this.”

“We need people who can demonstrate they have the agility, they have the adaptability. We can teach them if they're AI native, which is like how we hire, we can teach them the enterprise. We can teach people a lot of things that maybe is easier to do than teaching the AI adaptability or the acuity or the things that…”

“the phone is never going away, just like laptops didn't go away when the phones came, and all of our other technology built on each other, right?”

“And I think because the machines are human-like and don't have feelings, We can potentially create this pattern that has us start, you know, kind of being nasty to the machines, and potentially that spills over to our person-to-person interactions.”

“to me, it's still not significantly different than manufacturing automation in the 2000 and what that did to blue collar work in the U S and also outsourcing within China. And this is a whole other discussion right now. We, we only got two minutes left, but, but to me, it's still, it's just happening in knowledge work…”

“when a core value prop of perplexity is accuracy, uh, it's really hard to reinforce that to users when you also have ads running alongside the answer.”

“the alpha for a company that is not an AI company is, is not in them building their own internal tools with AI necessarily. Um, it is in the depth of their adoption.”

“If you already have a business, the things that you've considered outsourcing, You know, your, um, customer center, your customer call centers, whatever it may be, AI agents are perfect for. So you, you can do that. The stuff you would give to an intern, AI agents are perfect for.”

“in order to take advantage of AI, to answer your question, you have to reformulate your business completely to build it on AI. It's the difference between running a business before PCs and computers, and running a business after PCs and computers. You would do it completely differently.”

“If you look at ratings for sports, the less amount of game time, the actual playing game time, the higher the ratings.”

“I don't understand why robots need a human form factor. I want purpose-built and whatever this tabletop robot is doing on that table, I'm sure it's something very helpful.”

“I think like just being able to do some kind of like image alteration or improvement, or even like UX layout of a website. I think that kind of skill has been Going the way of someone kind of like writing email subject lines for a long time”

“society wants that talent. We're missing out on important talent when we streamline everything.”

“And I reject your application on the basis of a prediction. There's no way you can contest that, because predictions are not facts. At best, they're educated guesses. And because they're not facts, you cannot prove it to be false. And so, it's a way to shroud a lot of injustice and, and to lessen accountability.”

“when you have a protest, and in particular, a peaceful protest, It's very important to have anonymity. That is one of the bedrocks of democracy, and when you have cameras all over the place, and now with facial recognition being so easy to use, you are eroding one of the most important tools in the toolbox for…”

“surveillance is important because the whole machinery of surveillance is there to feed the machinery of prediction. So these two machineries are intimately related, and that's why it matters.”

“when we make the justice system about probabilities, we're losing its principled approach. And so you make it very easy for the bad guys to get away with it, because you don't have to make it impossible for people to challenge you, or even very hard. You just have to make it slightly unlikely for them to win. And then…”

“And, and to, to protect your, your exquisite targets on your, on your side against a very cheap drone, you have to use expensive countermeasures. And so the lesson there is, um, how do you turn the dial from, maybe we should have, um, more mass attritable, Weapons like drones or counter drones that are affordable so…”

“So this is kind of the big prize because it goes from the TAM The total addressable market being, you know, all of engineers to now the total addressable market is every knowledge worker, and that's probably about a 30 to 50 x larger market in terms of, you know, humans on the planet and, and, and their use cases.”

“today, I don't think customers are getting, like, in general, as people have been, uh, embracing AI, they aren't getting that level of exponential output, uh, or maybe the right way to say it is they aren't getting the level of exponential outcomes from their investments because, uh, our thesis is that the models have…”

“the second you come inside a company, you're now that company's content marketing arm. You're not doing what you did. You're doing something completely different.”

“If you're Trying to build a website and your agent messes it up and your user is affected. It's not really the agent's fault. It's your fault. And so you need to care. And I think that for people to use these tools, right, you need to realize that human agency, human accountability, that's a core part of the system,…”

“we used to really just focus on the raw pre-training capability, but not think as much about the inference ability. And that's been a big change over the past 24 months to realize that it's a balance between You can have this model that has all those great properties in the base, but then you really need it to be able…”

“the private credit business really can't fall apart before, frankly, the private equity industry falls apart. Meaning, if you think about it, who has taken on the biggest loans from private, from the private credit space? It's oftentimes the private equity players, or some of these tech players, or some of these, so…”

“Well, if you decide the software company is no longer a thing and may not have this recurring revenue, all of a sudden its value has to come down, and all of a sudden, therefore, they would have to lose all their money, and then if for the credit guys who loan the money, they would also therefore then lose money too.”

“The one other thing about data centers, which is a little bit more nerve-wracking, is it's not like putting, you know, people compare it to putting fiber in the ground or building the railroads. You put the tracks down. Because once you build a data center, you probably are going to still want to upgrade those chips…”

“one of the reasons why people are moving to prediction markets and moving to what I would describe as the sort of a lottery ticket approach to life. It's a little bit YOLO. Um, is a function of the inequality that we have in this country.”

“The truth is, depending on how you run a country and a social safety net and a system, if you can't afford the loss, then the loss gets fully socialized.”

“when you have a crash, you can't move into a period of austerity. You actually have to throw money at the problem as politically unpopular as the bailouts are. The Federal Reserve needs to throw money in.”

“small models are gonna run locally on device. And that's actually gonna be how infrastructure gets built out. That throws such a wrench into any of these stories that we've been talking about for the last 30 minutes, like that completely destroys those stories.”

“when it comes to the use cases that Anthropic may be concerned about in different, in different kinds of ways, I mean, the thing to remember is like, if you do business with the Pentagon, the business of the Pentagon is war. So you shouldn't be surprised then that the Pentagon wants to do all the war things with your…”

“if you're asking everyone to give you essentially sell themselves and the work they're doing on a weekly basis and using that as your like foundation for understanding the state of your company. It's gonna be biased positive. Like, that's not actually good. Cause everyone's gonna be like, oh, did amazing things.…”

“I also think consciousness is really helpful in a social situation. When you're dealing with Um, a world that is fundamentally unpredictable. That is to say what other people are going to do at any given time, what other people are going to say at any given time. And you have to be able to imagine yourself into their…”

“The Turing test doesn't work for this consciousness question. It was designed for the intelligence question, uh, which is somewhat simpler.”

“Our definition of the human, what's special about us, which has always been related to our intelligence and consciousness, is under enormous pressure today from these thinking machines and possibly feeling machines, and then from all these animals that we're learning are much more conscious than we thought. You know,…”

“You know, people have been trying to do what has worked everywhere else in science, which is reduce phenomenon to mass, you know, to matter and energy. And it's been an incredibly productive strategy, but it doesn't seem to work Yet with consciousness and that the effort to reduce it to things we know, um, hasn't…”

“I think that there is some wish fulfillment that we, we have something that is immaterial that therefore might survive the mortal body. I, I think behind a lot of people's talk about consciousness is the word soul, even though it's not articulated, but what is the soul? It's, it's also this immaterial essence of us…”

“the tension in my mind is between spirituality and egotism. And to the extent you can reduce egotism, whether through psychedelics, but also experiences of art, experiences of awe, all of which kind of shrink the eye in a way that can be feel really good, um, that that is the, to me, the door to spiritual experience.”

“You're not in an IDE. You're not coding alongside it. You're just setting off a task, and it's going to go and do a bunch of work for you. And now, obviously, it's very clear that that's the dominant paradigm that we're going to be in. Uh, Codex has proven it. Um, uh, you know, Cloud Code has proven it. Devon and…”

“At, at a high level, the big AI companies have made these safety pledges of how they will treat their systems as they get more and more capable, but they are largely self-enforced. And so there's a lot of temptation to water down your commitments and go ahead with launches that you wanted to anyway.”

“It just had different economics, and I genuinely think that's fundamentally changed. And we, we've talked about this a lot, like, and an AI company, we don't even know what the economics really are of them in terms of, like, what it costs to operate the more people use it. LLMs don't, Behave like traditional software…”

“World models are absolutely essential when you want to build agents, because these agents are going to take actions, which is going to change the world. You want to be able to predict these effects.”

“And that's why, honestly, for many years I've been so much an advocate for open science. I just don't believe that you can keep these ideas boxed in unless you're willing to keep people boxed in, which we are not willing to do. Um, and so I don't, I don't, I don't think we have a way to close the ideas. We should…”

“quantum computing is energy efficient computing for solving very hard computational problems. And I say energy efficient computing because quantum computers consume very little electricity, and yet they are very powerful.”

“Uh, but if you ask an enterprise, did you actually make money out of it? They will in general say no. And the reason for that is that they are not customizing things enough, uh, and they are not, uh, thinking backward from the problem they want to solve.”

“We are also getting to a point where the verticals that you choose do not really transfer to the others. So there's no point in making a model that is good at very precise biology and very precise physics, uh, because they are the transfer in between those things actually pretty unclear.”

“Like the architecture side of how people are approaching these kinds of agentic workflows is getting solved. I think that's the big thing. And It's not at the model level that models are helping, but it, and it's, yeah, it's the, it's not the product actually. You know what? Instead of product versus model…”

“there's no way you can have a prediction market without insider information. It's even harder to keep that, um, free of insider trading than it is the stock market. And there are, Fewer regulations.”

“everybody believes that this, that there's some differentiation in the plumbing of the liquid cooled data center, right? That's not where the differentiation lies. It's all the same, uh, pipe and valves and fittings. Like everyone's using the same things there. The differentiation comes after you turn it on and how you…”

“A lot more code without addressing the subsequent steps is almost a liability. It's not, it's not an asset for a company, right?”

“We're in like the, you know, the GPT-II era of memory”

“Uh, the main thing consumers want right now is not more IQ. Enterprises still do want more IQ.”

“One thing you don't have is The ability for the model to not be able to do something today, realize it can't, go off and figure out how to learn to get good at that thing, learn to understand it, and when you come back the next day, it gets it right. And that kind of continuous learning, like, Toddlers can do…”

“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, uh, so that you truly unlock the value in that data, uh, for the products and services that you're offering.”

“one of the best ways to bring the right context to the models is actually to do a deep customization of the models. Not all context is dynamic. A lot of context is also, you know, changes at a much slower rate. That context being baked into the model makes the models much more capable to start with.”

“one of the reasons that I like the neural engineering approach is it produces these very large effect sizes. Like, cochlear implants, really, if you are, if you are deaf due to sensorineural hearing loss, they really work. Or if you've ever seen a, a video of a Parkinson's patient getting a DBA, a deep brain simulator…”

“These companies are building on Slightly different combinations of the same foundation models. They are all building at the application layer, but there is no real technical moat. So they need to, you know, win over the law firms with their white glove service and their, you know, shipping speed, um, the way that they,…”

“gone are the days where a product manager is just writing a spec and sort of giving a set of instructions or requirements for somebody else to go build, like a product manager necessarily now can be empowered to generate an app and get it to 40, 50 in some cases, depending on how good the models are, if you have some…”

“From a software UI standpoint, they have built beautiful products. They've built usable products. They created the whole chat UI was not really a true thing, and now, like, they led the way. And all Google has to do is just copy everything they do. Wait for them to innovate, release, just copy because UI is not…”

“one of the things that we say is automation drives escalation. You know, as I automate something, more and more people, the Alex's of the world are pressing zero and asking to talk to the operator. And that isn't coming away.”

“I think if you don't have an API on that system, we can't swivel chair yet. With an AI agent effectively or, you know, reliably. And, you know, we can vibe code and build the integration for them, but the API has to be there. And so mostly the blocker, like why couldn't you change that flight? It's because they don't…”

“When we're changing the machine and we're writing new articles and we're building new macros, it's sort of easy for the humans who own that machine to understand what's going on. Like the, the head of support technology can be like, yep, yep, that is the correct workflow. We should do that. That's, those are our, those…”

“And I do think that some degree of anthropomorphization is justified there because fundamentally these, 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 have been written about and that these, these models…”

“What's interesting to me about agency being popular in my view, among a certain like circle of tech people is that. Agency is a different way of saying this idea of like, oh, you should consider yourself like radically responsible for your life experience and the things that are What's happening in your life? And…”

“Like, of course, humans everywhere want to have an idea ideology and something to believe in, but I do think in Silicon Valley, it's like heightened. And so this idea that you might be driven by an ideology or a belief is socially rewarded here, um, in a way that then makes people like be more ideology, ideology…”

“And those things don't really go away just because there is a better convenience of sort of like getting to the end answer.”

“Saying I have a five-year-old chip that works just fine doesn't mean that it can generate the same revenue that it did five years ago, which is the accounting question. So that's sleight of hand by these companies to tell you that the chip still works. If it's only generating one percent of the revenue generated five…”

“The agentic stuff flips the everything store completely on its head, because instead of the everything store, you just type one more query in, and now the entire web is the everything store inside of the Comet browser.”

“What our perspective has been, the answer can't be you take all that data and move it to the cloud so it can run very neatly on these cloud-based models, because then you lose kind of control over that enterprise context that you built over years, the transactions with your customers, the unique insights that you have,…”

“writing is thinking and smart people like to think that's not going away when you have a thought and you put it under paper, put it, you know, type it out. Suddenly this, like all these subconscious, like feelings about that thought, all of your like life lessons get articulated in a way. Where just as the writer, you…”

“But if you think about it, a problem like this is an informational one. So it should be solved by like within Gmail itself or with the Gmail API or with model context protocol and not a browser kind of clicking through robotically your Gmail. And so it's still amazing to me that this system, it seems impressive and…”

“I don't think you can like checkbox your way to greatness and outsource your AI to, you know, Salesforce or, or Palo Alto. It's great. I love these companies, right? But, uh, we've got to build it in house.”

“What we're finding with AI is it lets everybody go so much faster. Like, I mean, I think when we started out, we're like, can we go 20% faster, maybe 30% faster. I think it's like 10 x faster that we can move.”

“So it was, it was DIY all over the place. And there's a lot of failure in that. Um, you know, it was, it, at first it was like, oh, so cool. Look at, look at what it can do. But again, that repeatability, that trust, that security, the data access, all the things we've talked about. And so I think we have this, you…”

“I believe to certainly that all of the data in, in one place is in similar data models is, A hundred X more effective than trying to unify five, 10 different data models and databases through an API.”

“if you get to, like, 50% as good as you would have done yourself, I don't think that's good enough, and it won't speed you up. And in fact, it's like, I don't know, I could have just done this myself, and now, now, then at least I would have known what it's done. When you start clearing this sort of like 75 to 80%…”

“once you get a service as widespread as Instagram or as widespread as Facebook or WhatsApp, it's hard to introduce a new behavior there. Um, you know, we did it with stories. I think they've since done it with reels, uh, but it's almost like one, you get one per generation.”

“I think we've seen some good players emerge in the social media space like a blue sky that doesn't seem to be focused on getting people to, you know, just spend as much time as possible on the platform all the time. That said, I think it limits their growth”

“If you think about ransomware, it's a business. It has rules to it, right? You know, when you pay the ransom, you get the data, right? It's, it's a rule. You don't, you're not tricked into it because it's a business and they need to maintain their reputation.”

“And the concepts behind, I think, today, AI security are actually well thought today rather than much later. And it shows a bit on the maturity of not only the security teams, by the way, it shows a lot on the maturity of the industry that we are having the discussion on What's safe AI? What, how to secure AI, uh,…”

“You need to, you need to have purpose-built solutions that, that solve sort of tailored use cases. Those can be very big use cases, like all of AI coding, but you probably don't want to be in a position where you have to kind of bootstrap this or, or build it all out yourselves.”

“This is not a panacea type of, of, of solution where you could take an existing workflow, drop AI directly into it, and then all of a sudden that workflow will be, you know, three X better. You usually do have to re-engineer the work to take advantage of AI.”

“As long as you are really good about what context you're giving the AI and, and how you are, are effectively grounding the, the AI in Uh, trustworthy data with the right kinds of prompts and a, and a high enough quality model. Um, you can nearly eradicate, uh, the, you know, all of, if not the vast majority of, of…”

“the phone it's reached, it's reached its final or its ultimate form factor, right? The phone is the phone. The air looks like the iPhone six, like, Full circle. It's the same thing. Um, is it going to fold? Yes. Is that very different? No.”

“The way I, I, I view this in corporate, it's very true. And in government is very true. Whenever you see a bad technology decision, it's always politics.”

“most of the brain, uh, is, is actually not relevant for communicating with the outside world or with artificial intelligence. Most of the brain, um, is taking care of the body, uh, and not in ways that are Particularly relevant to interfacing with the outside world.”

“Um, and that continues to hold true, but we now have this kind of other category of, of, of training, which is post-training, Uh, and being able to use test time compute in more interesting ways than we used to as almost kind of a second stage of training. And so we think that that actually gives us a little bit of a…”

“Um, and so that's, uh, for us, I think the real world benchmark is increasingly becoming important. Uh, as a sign of intelligence relative to the academic benchmarks.”

“Salesforce didn't win cloud early on because it just was the best at cloud. It won cloud because it was just like a powerhouse in the rest of doing business.”

“my view is that coding is particularly interesting because A, the adoption is fast. Um, and, and B, getting better at coding with the models actually helps you to develop the next model.”

“what you can do in the Valley, what you can do in business, you can't always do in the political realm. Uh, the political realm, uh, the government has functions that, uh, businesses don't have. Has many, many, many more veto groups, and many, many more constituencies that have to be taken account of.”

“That is the, of that 37 or thirty eight billion, a lot of that money was given by people who gave it very specific things, and you can only use that payout for very specific things. So those are big numbers, but they're, it's not as flexible as people think.”

“I thought it was obviously a ridiculous You know, ridiculously bad situation, uh, 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.”

“The APIs, like, those are relatively, not necessarily easy, but they're easier to just swap in and out different, you know, different tools and technologies as they become more readily available. And then, yeah, the cost will get sort of driven down over time, you know, as that happens, as there's more competition in…”

“The sycophantic AI is that is, is the greatest limiter to like actual true intelligence or reasoning.”

“the dialogic nature means that a human doctor can follow along and actually learn In a very transparent way. It's almost like having an interpretability mechanism inside the black box of the LLM, because you can see its thinking process in real time.”

“I think it's primarily social media. I think that social media, uh, so, so I think that Americans, America was this very big, diverse country. You had ideological diversity. We had geographic diversity, not just like racial diversity, which is what we typically think about. We had all these, these forms of diversity…”

“Like they're so, I mean, they're so easily copied. It seems like that, you know, if you're just doing AI research and you make like a breakthrough like that, that's really about like technique and, you know, how long can you hold onto that? IP before, you know, anthropic has it or Google, you know, Google's probably…”

“So I think that in like a very simplistic sense, narrative drives stocks more than earnings do over the intermediate term.”