Watch mode

“The companies that have been most successful that I've been involved with are ones where the pivot foot was about the technology, not about necessarily market understanding”

“And distillation is a way to effectively spend more money to get what is compute equivalent. Like outputs from a language model are a function of compute.”

“And they were open AI and the 01 was batting a 10 out of 10, which was amazing to me.”

“every great product, every great leader, every great decision comes from understanding that trade-off and realizing what does not matter.”

“Where I get off the train is I think takeover is a lot harder than maybe the literal statement suggests.”

“What that means is that you just point your web browser at one malicious website, and the attacker will own your entire Windows PC.”

“What's really hard is that the customers believe or validate that your robot creates enough value that they want to deploy your robots across many of their sites.”

“it took a few hours to like outperform the two years of work which we'd done as with human in the loop”

“If you had a thousand dollars in your pocket and someone took out five and ran away, you wouldn't then say, Hey, here's the rest.”

“It's like the opposite of lean startup. Like lean startup is like, yeah, you like see if you get something working and then you double down”

“The modern quantum computers run at a clock speed of hundreds of hertz or thousands of hertz, not at the gigahertz that we're used to.”

“if you have a system where the improvee, the things being improved, is also the thing that's doing the improvement, you now have a new way to think about it.”

“as context expands, and as individual user has a lot of personal context, the context window increase dominates the batch size.”

“what served you can also screw you. The dose makes the poison.”

“Anytime Greptile gives any sort of feedback or a score under five out of five, the grep loop fires. It takes in the feedback. It writes the code. It pushes to GitHub.”

“They very rarely, at least in my experience, have come from design sprints or times where you set aside on a calendar”

“the promise of AGI is that we're going to spend almost none of the time doing things that we have to do”

“you have to decompose their number. And so, okay, you did 2 million. That's great. How much of it was inbound? How much of it was outbound?”

“There's no announced plan for new revenue. Right now, it's shrink. It's cut costs. It's when we mush these two companies together, we're going to save $6 billion over the next three years.”

“presentations fail from too much information, not too little information. And I think that's true with books too.”

“The real question for AI is where can you apply that loop model in a way that you can make a better prediction or better idea, each loop”

“when we train a protein language model, we train on metagenomic sequences, which are not the highest quality data.”

“You can kind of think about it as for every generation of model, there's a 4X multiplier in compute, roughly.”

“I can design the thing in three months, but I can't actually get it into real silicon at scale for nine months.”

“the church does not support machine consciousness. The priority is what he calls safeguarding the human.”

“This is different because we're getting this capability before we've really characterized the shape of the threat.”

“I have a bribeable chief of staff. Like I have a corrupt butler. Is that what you want? That is a great term for it, a corrupt butler.”

“So I think there's going to be potentially a 10 to 100x increase in volume and the number of queries that are going to hit up against databases.”

“We found that if someone is politically active as a teenager, but CZs to be in their early 20s that tends to correlate highly with someone's drive”

“the glitch allowed us to get what one entrepreneur, Paul Alstrom, who wrote the book Nail It and Scale It, he calls it your beachhead.”

“You've got to break stars. You're going to launch stars. You need stars to put it behind the paywall so that people buy the paywall.”

“the most at-scale workload, so the largest workloads where you're spending the most money, is actually where you're going to want to go and post-train first, right?”

“if you have failure recovery, if you have whatever it is that's interrupting your work, That's part of the issue.”

“I think the majority of actions on the internet will be taken by agents. Models are going to become more cheaper and faster at rates that are, you know, quite incredible.”

“So the attempt to try and centralize and manage the control of the software, I think is a fool's errand.”

“The other thing I learned was when you just vibe code something for yourself, you have to maintain it.”

“but fundamentally, like AI progress is about data and about writing better software to serve your models.”

“So the key thing is I need to call you right when you're taking the blood pressure measurement because now you're in that mode.”

“The hard problem to solve is control, is once you're in the air, can you stay in the air?”

“When we see those indicators, we should probably pause and try to figure out what went down before going forward again.”

“the benefit of pooling is we get to essentially load balance across these different customer types”

“whenever you're predicting abstract representation of the world you're encoding your own bias into what's important to predict”

“the pursuit and optimization of a very simple, easy to describe quantities can often give rise to incredibly hard to describe complexity and emergence.”

“I know to my core, the person I'm going to show you right now is the person I'm going to become 12 to 24 months from now.”

“if AI works overall, plenty of other people want all our compute. So they're willing to pay.”

“Pfizer would sell you a drug for a million bucks that was worth a billion dollars. They didn't care.”

“we have one of our engineers that we named him the Viceroy of Variance.”

“You actually need to have a lot of depth in the users you have. You still need a lot of users.”

“it forces the agent to prove that this bug actually happens. Right. Because sometimes it just might have been user error.”

“What I found is the opposite. The team usually knows before the leader knows.”

“We had many checkpoints that were like green on 80 of the rows, but red on 20 of the rows. We knew that wasn't good enough.”

“my theory on this, is that they're overcomplicating the task massively by trying to train the model to behave based on an 80-page ethical system that they've codified.”

“I liken it to we start typically off renting our apartments, but then as you grow up, you want to own a house for various reasons, even if it's a bit of a headache.”

“the emotional sense in the room was like there's just such a weight.”

“the doom scenario becomes like way more human enabled if you are able to amass capital pretty quickly as a rogue agent.”

“What I know about consumers is that the first time it doesn't work, they just walk away. They have no incentive to try again to overcome frustration or brokenness.”

“they were able to get it up to like 99.5%, which is insane. I think that is, I think that's a viable thing now if you invest the time”

“I find that generally people lose motivation as they gain experience.”

“Give me five seconds and then I get five minutes. They're okay with that. But I've earned it. Yes, because I've earned the attention.”

“Egon goes and says, you're not going to remember this $500 million. $400. $400 million.”

“it is surprisingly easy to hill climb on non-verifiable domains if you turn them into some proxy verifiable domain.”

“if we projected inference to be 2% or 5% of the market, even if that specialized chip is, let's say, 2x faster, it might not make sense, right?”

“we will pay our plugins that are popular and are seeing a lot of usage. They're going to get part of the revenue share as well.”

“although this agreement was entered into voluntarily the governance that follows from it is not voluntary it is now day fast it's fiduciary”

“I think about that as how many people were in a position to learn the thing that you did.”

“All the scaling laws are exponential increase in compute for linear increase in performance.”

“And then for the last 10 years, we've spent about 40 trillion on health care. We have not moved our life expectancy an inch.”

“without the vision of a goal, a man cannot hope to manage his own life, let alone the life of others.”

“they actually had synchronous classification or monitoring as a to-do item, but they didn't actually get to that to-do item”

“semiconductor manufacturing was invented to make one kind of machine, right? It's the semiconductor transistor and circuits built around that.”

“I think data quality in robotics is a non-stationary problem in the sense that what is high-quality data for your model changes over time.”

“I believe that the approach of monitoring the human language that these systems use is going to be increasingly difficult as agents get more powerful.”

“you follow up forever until people buy, die, or they bought something and they don't want to, they don't want to work with you”

“I didn't realize how petrified I was of not being seen as a great trader. What a stupid thing.”

“I don't want an AI that just spits out slop that makes a vibe-coded clone.”

“planning for me isn't the agent writing the spec and me reviewing the spec and me saying, go build it.”

“you're looking at people's eyes. And like, that is the metric. It is not a DAO. It's not a DAO over a Mao.”

“I remembered like the internet rewards risk and it rewards like surprise and like things that like people don't expect.”

“I have not seen a straight shot sprint to an ownership market, but you have to go through the rental stage to become an owner.”

“for every breakthrough in math, you should assume there's a bunch of other breakthroughs in other fields that have more commercial applications”

“the LLMs can get you maybe 85% of the way there and tuning certain things.”

“Every time someone recommends that we should add complexity, whether it's a step in a process, a feature, a team, head count to a team, you name it”

“We tell ourselves stories in our heads that we then believe. And those are not real stories. Those are just stories. They're just not reality.”

“I basically think what's blocking continual learning is basically extremely data-efficient training from sparse rewards, which is an unsolved problem.”

“the demand for accelerated computers going up, but the demand for CPU and networking and storage, the traditional data center CPU, et cetera, is also going through the roof at the same time.”

“The fundamental difference in the way that we've built dots is the harness does not run on the machine.”

“So unemployment is still near record lows at 4.1%, even though labor participation rose 0.2% to 61.6”