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 1,548 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
Kant: Internet data lacks the thought processes needed for AGI
“But the reason that never got us to AGI is because, well, the internet never actually included the data set of the thoughts and actions that created it. It's the final piece of code, it's the final article you write, but not the thoughts that you had and actio…”
Software solving boring problems survives because AI developers hate boredom
“What's not going to be dead is the problems that like these AI people don't want to work on because it's so boring to build that software.”
Pre-training aids AI alignment by implicitly instilling human values
“I definitely think we would keep using pre-training data, not just from an efficiency point of view as well, but also I think there is interesting safety angles, because by pre-training and, you know, all this human knowledge, we're implicitly creating an agen…”
Using chain-of-thought as an RL reward destroys model interpretability
“If you're not careful with RL, you can make interpretability harder. For example, one Common thing with modern models is they do reasoning with the chain of thought. You could look at the chain of thoughts to, you know, see what are the model internal thoughts…”
Tworek: Calling LLMs strictly next-token predictors is inaccurate in RL era
“Language models do on their own, like fundamental level is they are often called as next token prediction machines. And that's not completely accurate in the age of reinforcement learning, but they still operate on mostly on tokens that are mostly text.”
Tworek: Uninformed researchers pose a greater risk than IP leaks
“It is like, yeah, it is some like risk of losing IP, but I think the risk of not doing the right thing and of people not being informed about research and not being able to do the best research is much higher in my personal opinion and how, how I approach thos…”
Jerry Tworek: Pre-training AI models is mathematically simple compared to RL
“The first thing that is important to know and understand, RL is hard. Like, conceptually, if you think about it, and there's still a lot of depth to it, but very conceptually, mathematically speaking, pre-training is dead simple.”
Tworek: Reinforcement learning and pre-training require each other to succeed
“And like, I don't like in terms of a pure RL, I don't think like really pure RL makes sense. RL needs Pre-training to be successful. And I think pre-training, as I said before, needs RL to be successful as well.”
Douglas: Independent technical blogs are the highest AI hiring signals
“The fastest route, or like, the most immediate one is whenever we see a really good blog post where people have, like, done incredible amount of work in an independent fashion, it's one of the highest signal things there is.”
Douglas: AI takeoff speed depends on AI assisting AI research
“We think that one of the most important signals of whether or not we are basically the speed of takeoff, the speed of progress is driven by how much AI is able to assist AI research.”
Douglas: Solving AI hallucinations intrinsically requires reinforcement learning
“Saying, I don't know, or solving, you know, hallucinations is, ah, intrinsically requires reinforcement learning in many ways.”
Douglas: AI reasoning strategies emerge naturally with enough compute and RL feedback
“Give it math questions, tell it whether it got them right or wrong, and the model will learn. This is, it comes down to a bit of lesson in scale and search, is just allow the model to search, have enough compute to run the experiments, and the model actually e…”
Valenzuela: AI software scales through core principles rather than vertical specialization
“It used to be the case that you pick verticals. I think now you pick principles. And the principles allow you to scale much better than any other previous generation of software.”
Cherny: Code is the primary path for AI to reach AGI
“Maybe coding is the way that we get to the next level of intelligence. If you call it like AGI or ASI or whatever, the model needs some way to interact with the world and for a model, the natural way is code.”
Laskin: Solving organizational code context yields all capabilities for superintelligence
“Like if you really solve this oracle for organizations just for coding, you've basically built all the capabilities you need to have super intelligence.”
Laskin: Reinforcement learning makes LLM capabilities jagged, not broadly general
“When you train large language models with reinforcement learning, they become jagged in the sense that they become good at what you wanted them to be good at. And there are some generalization capabilities, but they're much weaker than people think.”
Laskin: Focused AI startups can operate with 10x less capital than frontier labs
“You can't operate at a hundred X less capital than a frontier lab, but you can operate at, say, 10 X, like an order of magnitude less capital when you're really focused.”
Vercel CTO says top engineers are becoming primarily code reviewers
“Malte, among many interesting things, I think one of the most fascinating things that he said was that most of his great engineers who have been with the company for a while, as they've dogfed you know, vZero, and they've used things like Cursor and Windsurf i…”
Walsher: Tech buyers value Cursor's opinion over $80B legacy software incumbents
“Most people will care, at least in our world, What the smartest buyer at Cursor thought about a given tool than what the smartest person at the, you know, eighty billion software business that IPO'd in, in 2012 might think.”
Rauch: Dev tools must be designed for both humans and AI agents
“LLMs.txt or raw markdown is better for agents. So if you're building developers still today, you have to have that duality in your head. You need to think, okay, How do I make my content, my errors, my developer tools great for agents?”
Conversational AI interfaces have natural limitations for visual design workflows
“If you are doing any kind of visual communication, then there are going to be natural limits to a conversational interface. And people are going to want to be able to collaborate. They're going to want to be able to control for brand. They're going to want to …”
AI coding tools pose serious quality risks when used by junior engineers
“There's a challenge for our industry in that we do see kind of slightly bipolar results for these tools in the hands of a senior engineer who can tell good code from bad code. They're very, very powerful, but for graduate engineers, for junior engineers, ah, t…”
Dohmke: Context-aware AI agents with tool calls outperform fine-tuned models
“So this iterative process that the agent does with the help of the tool calls in all the context makes it so much more powerful than a fine-tuned model could ever be.”
Gomez: Creating AI reasoning models is dramatically cheaper than pre-training
“It's easy to create a reasoning model. It's dramatically cheaper than pre-training. And so it's accessible. And so there's this huge intelligence uplift that comes for really quite little effort.”
Prompting AI agents will soon become synonymous with management
“As agents are rolled out, you'll actually start to see you know, people that are really good at prompting, really good at defining a process, be the best managers, and actually be the best at extending whatever their agenda is in the organization, or making th…”
Enterprise AI sales are currently driven by FOMO, not traditional pain points
“Right now, it's less around pain, it's less around, like, standard enterprise SaaS cycles, and probably more around FOMO, around missed upside, around value cases that are really hard to define.”
Evans: Consumer chat interfaces are thin wrappers, not vertical SaaS
“The only thin, thin GPT wrappers are what you get when you go to chatgpt.com and claude.com and grok and all these others. That's a thin wrapper on a model. Whereas you know, name your vertical enterprise SaaS company. That's not a thin wrapper.”
Evans: GPTs could lower software creation costs by an order of magnitude
“The analogy that's been floating around, I think, is, is to compare this with AWS, in the sense that AWS was a sort of an order of magnitude change in how easy you could get a startup out of the door. You didn't need to write all this stuff yourself and buy in…”
Evans: Meta and AWS both benefit from AI becoming cheap commodity infrastructure
“In a sense, like, AWS and Meta are on the same page, and then Meta wants this to be cheap, generic commodity infrastructure that's sold at marginal cost, and they will differentiate on cool Facebooky stuff on top. Amazon want this to be cheap, generic commodit…”
Evans: Consumer AI apps currently lack viral loops and network effects
“There's no viral loop. There's no network effect. There's no reason why you should use the one your friends use. There's no reason this one gets better because everyone else uses it, at least not yet.”
Levie: Enterprise AI faces no philosophical resistance unlike early cloud
“The difference is the energy is very different where I think most companies I meet with and happen to be in New York and meeting lots of banks. The energy is like absolutely in the category of how many use cases can I apply this to? How do I start to lean in m…”
Ramaswamy: Open formats like Apache Iceberg lack governance and access controls
“This is also where Iceberg or just a data lake strategy or a lake house strategy is not an immediate answer because formats like Iceberg do not understand governance rules, do not understand role-level access control, they don't understand users, they don't un…”
Chollet: Intelligence should be defined as skill acquisition efficiency
“And yeah, so I, to summarize that, you know, I see intelligence as skill acquisition efficiency. So it's not the fact that you can acquire skills, it's how efficiently You can do it. That's a measure of your intelligence.”
Chollet: AGI means humans can no longer easily create tasks AI fails
“You have AGI when it's no longer possible to easily come up with tasks that, you know, you and I can do naturally, but no AI system can do.”
For AI applications, moats lie in curated data rather than user experience
“Their mode is probably not the user experience, but because it's very easy to copy. Anyone can study the product and copy. Their mode is data.”
Kiela: DeepSeek proved frontier AI models can rely on synthetic data
“We have kind of an existence proof now that it's actually not that hard to do this and so you don't need to invest all that much in, in data, and you can use synthetic data and get a pretty good model out of that”
Kiela: Fine-tuning cannot inject new knowledge into AI models
“One common misconception about fine tuning is a lot of people think that you can inject new knowledge into a model using fine tuning. And that is not true.”
Kiela: Advanced RAG systems break down when scaling to a million PDFs
“You can build a very awesome demo on a couple of PDFs and things will probably work. But then you have to scale it up to a million PDFs, and then everything breaks down. And the reason for that is that a lot of these kind of advanced RAG systems still actually…”
Misra: Video cannot be templated without AI content understanding
“Like you actually can't templateize video because the nature of it is completely, it's unknown, right? It could be saying anything. It could be doing anything. Anything would be happening. The, template just can't fit with, it can't understand what's happening…”
Gaurav Misra: AI hallucinations are creative features in video generation
“I think hallucinations in text obviously can be a bad thing, potentially, because they're saying stuff that, you know, isn't true, but in a video, like, you are delegating a lot of decisions to the model, right?”
Murchison: Ada added artificial email delays because instant AI replies had lower open rates
“So, like, you know, the, in other words, like, customers weren't opening these emails because they didn't believe that the email could be helpful if it came so instantly, right? So it was like an eight-minute delay that a lot of our customers use.”
Masad: LLM reasoning performance degrades rapidly past 32,000 tokens
“What we found is reasoning over long contacts is actually not very good. It's like, not good at all. Like, once you cross 32,000 tokens, that performance of reasoning, you know, just goes down to hell, like, very, very quickly.”
Masad: LLMs are next-token completion engines rather than reasoning systems
“What models are still are today is they're like, they're sort of like completion engines, right? That's how LMs are trained. The sort of auto aggressive models where they try to predict the next token. They're still next token prediction machines. Reasoning is…”
Justin Borgman: Open-source enterprise infrastructure beats proprietary alternatives over time
“I believe, gradually over time, enterprise software technology, at least infrastructure technology, wherever there are two things that are mostly the same, and one is open and one is not, the open is going to win over time.”
Chip Huyen argues post-training is what differentiates frontier AI models
“So, so I do think that post-training is what makes this, like, really big lab models are, like, different.”
Evaluation is the single biggest bottleneck holding back enterprise AI adoption
“So, so I do think that evaluation is the biggest bottleneck for AI adoptions, because unless, like, if we can, like, if we can, like, develop a more reliable way to evaluate the application, that application is not going to get adopted. Like, or maybe, maybe i…”
Chip Huyen argues human-generated plans are poor training data for AI agents
“When we ask humans to generate like what they consider the best plan for an actions, for a task, it's actually like not quite the best plan for AI, because what is what is easy or efficient for humans is not the same as easy and efficient for AI, right?”
Enterprise AI drives 500% capacity shifts, not just 10% productivity gains
“None of this is about productivity, right? You being able to go through Three decades of consumer research on Listerine to be able to develop a new flavor, right? You being able to develop all of the go to market around that new product in literally a third of…”