The Ledger
Every statement that passed quotation and attribution checks. Mix any filter with any other: certainty 1/5, debate potential 5/5, or both at once.
why aren't all 2,573 resolved? a statement only gets an assessment when the public record can support or contradict it. opinions and what-ifs never can, and 0 checkable ones are still open, waiting for their date. predictions held up or didn't; assertions are supported or contradicted. on every card: ▮▮▮▮▮ certainty · ▮▮▮▮▮ debate potential. speakers are clickable
Tan: Agentic engineering fails when developers put Markdown logic into brittle code
“All of the difficulty in energetic engineering today is when people try to do things that should be in markdown in code and it fails because code is brittle.”
Tan: Technical people with taste benefit most from tokenmaxxing AI agents
“It's very significant for people who are technical because it actually raises the bar on like what you are capable of doing. Like all the people who are attacking me about lines of code, they particularly are, The people who are most likely to get wings if you…”
Tan: Founders should willingly spend $500 daily on AI tokens
“One of the key maxims for YC is You know, how do you find good startup ideas, live in the future and build what's missing? Right. And so this is a profound version of that where all you have to do is commit your brain to look at, you know, spending 500 dollars…”
Chaubard: Recursive models discover problem-solving strategies without human teacher forcing
“That's the most important part, is that if we had Sudoku, and we know how to solve Sudoku, because like we were just, you know, dumb homo sapiens that didn't know how to solve Sudoku, like it would just have solved it. And that's why it's cool, because it actu…”
Hassabis: Active agent systems are the necessary path to AGI
“You have to have an active system that can actively solve problems for you to get to AGI. That was always clear to us. So agents are that path.”
Tan: AI Scaffolding, Not Model Intelligence, Is the Coding Bottleneck
“The bottleneck here is not Not the model's intelligence. As long as you set the models up right, they are already smart enough to do extraordinary work on your code base.”
Vuong: Full autonomy requires an incremental mixed-autonomy approach
“We think that it's going to be more like a peeling an audience analogy, where you start from a really strong base model that have all sorts of common sense knowledge and already works to some extent on your robot, and you have then a Mixed autonomy system. Ver…”
Vuong: Reactive AI models eliminate the need for expensive, high-precision hardware
“You don't need a incredibly expensive robot that is capable of very precise motion today to be able to do this task. And the reason why is this model really reactive? And so they can compensate for some of the inaccuracy in the actual robot movement”
Vuong: LLMs lack a fundamental understanding of the physical world
“This only works for simple cases today, and the reason why that's the case is because I think it's pretty fundamental limitation of the model that we have today, which is that they are not at the core model that take action in the world and see the consequence…”
Bilgen: MRD cancer detection shares same technical challenge as population screening
“If you can detect a microscopic level of DNA and be able to say that that is actually cancer, that is the same really technical problem as being able to detect those in healthy patients or, you know, general population.”
Mellata: AI Agents Eliminate the Need for Specialized Fraud Classifiers
“You don't need a classifier anymore because AI agents are able to read a set of standard operating procedure and reason over an image or reason over sort of unstructured data and know that this is possibly chargeback fraud. You don't need a specialized classif…”
Chollet: Current LLM Stack Can Fully Automate Any Formally Verifiable Domain
“And I think right now we're in this situation where any problem where the solutions you've proposed can be formally verified, and you can actually trust the reward signal. It's not just some guess made by a model. Any domain like this can be fully automated wi…”
Chollet: General Intelligence Is Human-Level Skill Acquisition Efficiency
“General intelligence is human level. Skill acquisition efficiency on the same scope of tasks that humans could potentially learn to do.”
Chollet: Human-Engineered Agent Harnesses Prove AI Is Far From AGI
“I mean, to me, the fact that you need humans to engineer these harnesses is also a sign that we're short of AGI today, because if we had AGI, you know, AGI would just make its own harness. It would not need to be told how to solve a problem.”
Chollet: RL Benchmarks Like Dota and Atari Test Memorization, Not Intelligence
“If you look at Atari games, for instance, or even Dota, you're training on, on the same environment as what you use for testing. So effectively, you're just trying to memorize the best strategies. You're trying to at training time, explore the full space of po…”
Chollet: Human generalization stems from symbolic, causal program synthesis
“I do believe the human mind does at the highest level something that looks a lot like programs in this, like we're currently building Causal models of our surroundings. Like we are describing our surroundings in our mind as, you know, a set of objects and agen…”
Chollet: Deep learning guidance is necessary to break combinatorial program search
“You have to break the combinatorial wall, and the way to do it is to add deep learning guidance. It's actually very similar to the principles that analyze something like AlphaGo or AlphaZero.”
Chollet: Scaling genetic algorithms could automate scientific discovery
“If you try to scale up genetic algorithms, I mean, I'm sure you can do incredible things with that. You could, in fact, probably do new science. Because that's based on search, and search is the best fit for automating the scientific method.”
Chollet: AI Architectures Requiring Human Engineers to Scale Will Fail
“If you're working on something, but the only way to increase the capabilities of the system Is to have human engineers and researchers spend time on it. It will not work because even if the idea is very clever and very elegant and works really well, capabiliti…”
Mukund Jha: Solving verification enables full automation of software engineering
“If you can solve for verification, which is essentially, you know, you can solve the testing part you can actually automate all the software engineering. That was sort of our key insight that like, you know, verification is the loop, which sort of keeps agent …”
Adding deep developer power after building simple UX is structurally difficult
“And I think fundamentally it's like, unless you start from, you know, a starting point, which sort of solves all of these problems along the line, the whole software development life cycle, it's actually really hard to come from the other side and solve these …”
Hodak: BCI progress is limited by hardware, not neuroscience complexity
“Brain-computer interface research and development is limited by your ability to record and stimulate these signals. The neuroscience, comparatively, is actually pretty simple. As soon as you can record the signals, we've very quickly figured out what, we talk …”
Epstein: Poor demo design signals a YC applicant didn't try
“Especially as we're reviewing YC applications, if you look at a demo and that demo does not have a base level of quality design, it seems like the person just didn't even try because it's so easy to do it now.”
AI is replacing human engineers for dataset understanding and failure-mode detection
“Historically in machine learning, you always, you know, it's like the rule was you have to know your data set really well. But now we're kind of outsourcing that to the AI itself, where the AI is the, it's the AI's job to understand the dataset and figure out …”
Scaffolding gains get wiped out by next-generation AI models
“There's this other general principle that I think is maybe interesting where you can build for the model and then you can build scaffolding around the model in order to improve performance a little bit. And depending on the domain, you can improve performance,…”
Delete oversized CLAUDE.md files and start fresh rather than compacting them
“If you hit this, my recommendation would be delete your QuadMD and just start fresh.”
Steinberger: Local AI agents are dramatically more capable than cloud AI
“I think my big difference is that it actually runs on your computer. Like everything I saw so far runs in the cloud. It's like, it can do a few things. If you run it on your computer, it can do every effing thing, right? So that's really more powerful.”
Steinberger: Coding AI model capabilities transfer directly to solving real-world tasks
“Coding is really like creative problem solving that maps very well back into the real world.
I think there's a huge correlation.
They need to be really good at creative problem solving, and that's a skill.
That's an abstract skill.
You can apply to code, but l…”
Steinberger: Converting MCP tools to CLIs scales better than native integrations
“If you need to, you can use MCPs on the fly. You don't have to restart, unlike, unlike Codex or Cloud Code, where you actually have to restart the whole thing. I think it's way more elegant, and also scales way better.”
French-Owen: The value of writing basic integration code has dropped to zero
“Segment is a funny business in that where we started was building these integrations, right? And so it's like, oh, you need to wire up, like, the same data going to, like, Mixpanel and Kissmetrics and Google Analytics, et cetera. And I think just writing that …”
Taggar: Never give personal AI agents access to emails due to prompt injection
“Do not give it access to emails would be my number one piece of advice or probably anything. Cause it's not clear how safe it is. And it's probably almost certainly gonna probably a lot of people being prompted injected by it right now”
Taggar: Incremental model gains make finding AI startup ideas difficult again
“The models themselves have incrementally improved this year, but there haven't been like major steps forward that have shaken everything up, which is, has a knock on effect. Many episodes ago, we talked about how it was It felt easier than ever to pivot and fi…”
Taggar: Rising AI software expectations force startups to maintain pre-AI hiring levels
“AI is going to reduce the cost of, like, producing the time to produce things, and so then the expectations from your users and customers will just go up, and you'll need to keep hiring more people to satisfy, like, the growing expectations. I feel like this y…”
B2B brands should be weird enough to polarize audiences intentionally
“Like, it needs to be so weird that it polarizes. Every day we'll get a whole bunch of people talking about that website, and we monitor all our brand mentions online, and it's going to be like either this is literally the best thing I have ever seen, and we al…”
Skates: AI product building must be technology-first, not customer-driven
“With the capabilities of AI, because they're so jagged, you, it's a technology first understanding of what is possible. And so if you go to your customers and tell them and ask them what they want, like, they're not even going to be able to describe what's pos…”
Skates: The top filter for founders is persisting when quitting is rational
“One of the really clear takeaways from any of these journeys is there is a point that you get to a year, maybe two years in, where From a rational standpoint, you probably should quit but for whatever reason those successful ones don't, and so that is the numb…”
Skates: Most successful founder CEOs leave after a decade due to scale
“Even if you look at the successful founder CEOs, after about a decade, most of them leave. And in my mind, it's actually for this very reason, because being a large company executive is different.”
Tan: Micro-creators with 1,000 to 10,000 followers are mispriced distribution assets
“When you're paying someone else, like the reason why that's interesting for consumer products is it's a mispriced asset generally still. I mean, Mr. Beast is not a mispriced asset. He's like getting his value from it. Like he's, you know, but It does seem like…”
Heller: AI application TAM is 1,000x larger when capturing labor budgets
“Today, the actual amount of money that we already know people and companies are willing to spend is the combined salaries of all the people they're currently paying to do the job. And that number is like a thousand X bigger. You pay 20 dollars a month to solve…”
Semi-automated companies allow single bad actors to act without human whistleblowers
“In a semi-automated world, that's no longer true. And it could be the fact that a single person could make a decision that changes the entire impact of a product. And there's no single person that might be aware of that except themselves. It gets extremely eas…”
Fisher: Users choose honest AI over sycophancy when asked at principle level
“But I think if you take a step back, and you say, hey, hold on a second, here's two principles, and you can choose the AI that follows this principle or this principle, The first principle is that we're never gonna blow smoke up your ass. We're only gonna tell…”
Anthropic's Joseph: Compute matters far more than pre-training objective details
“I think that, like, the one sort of general intuition I have is, like, compute is the thing that matters. So, like, I think if you throw enough compute at any of these objectives, you're gonna get something that's probably pretty good, and can kind of be fine …”
Joseph: LLM training requires collaborative infrastructure work over publishable research papers
“And to do a project like training a large language model requires a lot of people to collaborate on like a really complicated piece of infrastructure that isn't going to be a paper, right? Like you're not going to publish like, oh, I got a slightly, I got five…”
Joseph: Sparsely linked long-tail data may be most valuable for frontier AI
“And it might be that like, that data ends up more valuable because you, everything that's linked to a lot, you've already got. Like at some point, you're maybe like going for the tails, or you're going for the stuff that no one's ever, like, you know, it's onl…”
Joseph: Training purely on raw LLM generations cannot produce a better model
“Theoretically, I shouldn't be able to train a better model than that. Like, I'm just going to get the same thing out. So I think that's-”
Nick Joseph: Third parties can steer frontier AI labs by publishing evals
“Like, it is the case that, like, the labs right now are really driven by getting good eval scores. And it's hard to make them, and anyone can do it. There's no comparative advantage to having the model to making an eval. So I do think it's actually, like, an i…”
Anthropic's Nick Joseph: Frontier pre-training teams primarily need engineers, not researchers
“The thing we, like, most need is engineers. Almost always, like, throughout, like, the entire history of this field. It's, like, the case that you throw more compute, the thing kind of works. The challenge is, like, actually doing that.”