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 5,759 resolved? a statement only gets an assessment when the public record can support or contradict it. opinions and what-ifs never can, and 49 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
Masad: Software companies must shift from building apps to directly solving problems
“And for Replit, and I'm sure a lot of other businesses to survive, at some point Replit needs to stop being focused on making applications and start being focused on solving problems with software.”
Masad: AI model training has reliability limits without fast environment feedback loops
“I think there's a limit on how much the training can increase reliability, but I think the environment feedback and the ability to try things really fast is the way to get to the upper echelon of reliability.”
Masad: Next-gen coding AI will train via AlphaZero-style RL, not human code
“My bet is that Pretty soon we're going to move into more of the alpha zero style of training where you have a more traditional LM that's trained on all of the internet. But then the way to train the next generation of it would be to give it a reinforcement lea…”
McGrew: FDE-led product discovery beats traditional sales-led discovery
“Sales-led product discovery, you're talking to people from the outside. And again, this is important very early on, but it's not as effective as the FDE-led product discovery, where you're solving these problems from the inside.”
McGrew: Building what customers ask for turns startups into consulting firms
“I think one of the other failures, by the way, that's even prior to that, and more, the easier failure to become a consulting firm, it's where you build the product in the field that the customers are asking for, rather than the one that's actually valuable to…”
McGrew: Startups should avoid FDE strategy unless no alternative works
“My first, second, and third pieces of advice to people who are thinking about trying an FDU strategy is like, don't,
Don't do this at home.
If you can avoid it, like it's probably bad for you.
Probably you're going to end up doing services.
And then only if yo…”
McGrew: In five years, AI agents won't be a single category
“Probably in five years, we'll look back, we'll be like, well, AI agents, there wasn't even a thing at all, right?
We were actually doing all these different things.”
McGrew: FDE strategy succeeds by growing contract sizes rather than cutting customization
“In the product market fit strategy, you want to be doing less work for every customer, you want to be driving down costs, you want to keep the contract size the same. In the FTE strategy, you want to drive the contract size up. So you're doing more and more va…”
McGrew: AI capabilities will outpace adoption over the next five years
“What the world is gonna look like over the next five years is that the capabilities just race ahead and race ahead and race ahead.”
Truell: All Software Development Will Flow Through Models Within Five Years
“And I think we took a step back and realized that if we were being really consistent with our beliefs, you know, there was going to be an opportunity for all of coding to change in the next five years and for all of software development to flow through models.”
Legal AI innovation is scaling parallel queries, not prompt engineering
“And the big innovation there does not really come from, you know, how do you prompt and work with the model, but it's how do you make this run at scale? You know, how do you run a 100,000 queries in parallel at the same time and make sure nothing breaks.”
Legal due diligence commoditized as clients refuse paying for manual review
“And now it's becoming almost a commodity where you're expected to do it, but clients are also not really that excited to pay for Very simple contract review when they know that AI can do, you know, 99% of it.”
Blomfield: Discussing pricing early almost never scares off enterprise prospects
“Founders are often very afraid to have a willingness to pay conversation. They feel like if they bring a dollar amount up, it might scare the customer off. But honestly, that's almost always not true.”
Blomfield: Early startups should accept customer contract terms unless company-ending
“So I'd be pretty flexible on signing, frankly, whatever your early customers want, as long as it's not going to expose you to like unlimited liability or container clause that's going to transfer all the IP in your product to your customer or something company…”
Field: Design and craft are the primary differentiators in the AI era
“If you really believe that development gets easier and it's more simple to create software, it's faster to create software, then like, what is your differentiator? It's design. It's craft. It's attention to detail. It's point of view.”
Field: Critics should not dismiss Sam Altman and Jony Ive's hardware collaboration
“Sam is one of those people that, you know, he's right about a lot of stuff, so I would encourage you if you just dismissed it outright, To ask yourself what you might be missing.”
Field: Designers and PMs, not just researchers, should build AI evals
“As you're, you know, doing developing a model or you're developing research ideas, You have to have good evals, and usually the researchers are the ones building those, and I think that's kind of just the wrong model. For us, at least, designers, my point of v…”
Dylan Field: Figma does not view Cursor as a competitor
“I really don't think of Cursor as a competitor. I think of them as someone that, I mean, we just launched our MCP server to explicitly make it so that you can get your designs into Cursor and Winsurf and all these other, and VS code, you know, all these great …”
Chas Englander: AI models outperform humans at structuring public filing data
“You know, if you look at some of these data providers, they've got humans reading information from say public filings and put it in that structured format. You know, the bulk of the work that we're doing, what we're seeing is models are already more accurate t…”
Chas Englander: Top financial firms have already fully automated basic data gathering
“But I think some of the more lower level, more kind of data gathering and presenting types of tasks are fully already being automated at the top firms.”
Englander: AI workflows will become fully autonomous in 2025
“We think the biggest shift this year is going to be that even that element won't happen, and therefore elements of the user interface we think will be less important. In other words, these tasks will happen Entirely autonomously. You know, as you arrive in the…”
Englander: Global finance technology buyers are surprisingly based in San Francisco
“There's a surprising number of decision makers for global firms around sort of technology implementation in San Francisco. It's not out of New York or out of London, it's out of San Francisco.”
Kaplan: AI scaling trends are as precise as laws of physics
“We found that there's actually something very, very, very precise and surprising underlying AI training. This really blew us away that there are these nice trends that are as precise as anything that you see in physics or astronomy.”
Kaplan: Compute scaling drives AI progress more than researcher cleverness
“Basically you can Scale up the compute in both pre-training and RL and get better and better performance. And I think that's sort of the fundamental thing that is driving AI progress. It's not that AI researchers are really smart or they suddenly got smart. It…”
Kaplan: AI may execute multi-month tasks within the next few years
“And this kind of picture suggests that over the next few years, we may reach a point where AI models can do tasks that don't just take us minutes or hours, but days, weeks, months, years, et cetera.”
Kaplan: Founders should build products that current AI cannot quite support
“One is I think it's really a good idea to build things that don't quite work yet.
This is probably always a good idea.
We always want to have ambition, but I think specifically AI models right now are getting better very, very quickly.
And I think that's going…”
Kaplan: AI scaling curves point smoothly toward human-level AGI
“I think that scaling Really suggests a kind of smooth curve towards what I expect is kind of human level AI or AGI.”
Kaplan: Broken scaling laws usually signal flawed training setups, not fundamental limits
“If scaling laws are failing, it's because we've screwed up AI training in some way. Maybe we got, ah, we got the architecture of the neural network wrong, or there's some bottleneck in training that we don't see, or there's some problem with Precision and the …”
Kaplan: Most AI value will come from end-to-end frontier models
“I think that you can do a lot of very simple bite-sized tasks, but I think it's just much more convenient to be able to use an AI model that can do a very complex task end-to-end, rather than requiring us as humans to sort of orchestrate a much dumber model to…”
Finn: Generalist robotics models may outperform purpose-built models
“And we think that this sort of generalist model may work better and be easier to use than purpose-built models, just like we've seen in the development of foundation, foundation models for language and other applications.”
Finn: Scale is necessary but not sufficient for open-world robotics models
“And so I think the lesson here is that scale is necessary for developing these models that can generalize in open world conditions, but they're subordinate to actually solving the problem. So you need scale, but it's not sufficient for the entire problem.”
Finn: Diverse Home Training Matches Custom Target-Environment Performance
“And we find that if we actually increase the amount of homes, the amount of locations that are represented in the data, The performance increases, which is great. And it actually gets to the same level of performance as if we train on data from that target env…”
Finn: Frontier models struggle with visual understanding for robotics
“In general, we found that these frontier models generally struggle with visual understanding as it pertains to robotics. Which makes sense because in general, these models aren't kind of really targeting, ah, many physical applications and have very little dat…”
Chelsea Finn: Real robot data cannot be replaced by synthetic data
“I think that at the end of the day, there's going to be no replacement for real data. And so we're like large amounts of real robot data. It's going to be a necessary component of any like system that's going to work in a generalizable way.”
Jumper: Equivariance explains only 2-3 GDT of AlphaFold 2's 30-point gain
“The sixth row there, no IPA, invariant point attention, that removes all the equivariance in alpha fold, and it hurts a bit, but only a bit. Alpha fold itself on this GDT scale that you can see on the left graph, alpha fold two was about 30 GDT better than alp…”
Jumper: AlphaFold made structural biology 5% to 10% faster
“I like to think that our work Made the whole field of what's called structural biology, biology that deals with structures, you know, five or 10% faster. But the amount to which that matters for the world is enormous.”
Jumper: Scientific AI will eventually be driven by broad, general models
“I think we will start to see this on more general systems, be them LLMs or others, That we will find more and more scientific knowledge within them, and we'll use them for important, important purposes, and I think this is really where this is going, and I thi…”
Srinivas: Perplexity's primary strategic bet for the future is an AI browser
“The browser. That's the big bet we are making as far as the future of the company goes.”
Srinivas: OpenAI and Anthropic Will Both Build Their Own Browsers
“That said, I'm fully working with the assumption that OpenAI will also build its own browser. Anthropic will also try to build its own browser.”
Srinivas: Google's ad model disincentivizes providing direct search answers
“If you get direct answers to these questions with booking links right there, how are you going to mint money from booking and Expedia and Kayak and like, you know, all like same, same, same thing for shopping. How are you going to take money from Amazon and li…”
Srinivas: Google rebrands and announces the same AI search features every year
“The same feature is being launched year after year after year with a different name, with a different VP, with a different group of people, but it's the same thing. Except maybe it's getting better, but it's never getting launched to everybody.”
Srinivas: Phone Makers Will Offer Multiple AI Search Alternatives
“The phone makers will start offering all of them as alternatives. It's not going to be like a locked-in default search option.”
Srinivas: OpenAI's built-in search features in ChatGPT did not kill Perplexity
“OpenAI has perplexity within ChatGPT that did not kill any of these companies.”
Srinivas: No AI product has in-app network effects yet
“No AI product has within app network effect. Like it's not like WhatsApp where if you build a WhatsApp rival, Meta has a definitely like a questionable brand, right? Like people don't necessarily trust Meta's products. They think like these are ad products. De…”
Srinivas: Browser-native agents offer superior reliability over MCP-only approaches
“That's the key advantage of the browser that you do not have if you commit entirely to just the MCP vision. If you commit entirely to the MCP vision, you require these third-party MCP servers to work reliably. Ah, the data that they send you, ah, on, with the …”
Srinivas: Perplexity can reach billions annually in subscription revenue alone
“We think, like, we can grow at least, you know, a few billions a year in just subs, which is a great business.”
Srinivas: Foundation model labs will copy any high-revenue AI startup idea
“You should assume that if you have a big hit, if your company is something that can make revenue on the scale of hundreds of millions of dollars, or potentially billions of dollars, you should always assume that a model company will copy it. Mainly because the…”