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
Benaich: Tech leaders underestimate human inertia in real-world AI adoption
“Tech people in SF obviously live in a bubble, and it's kind of, it's quite techno-utopist, and I think they underestimate how, how much human inertia there is to the diffusion and adoption of technology in, like, the real world.”
Bernhardsson: Serverless architecture fits AI workloads better than web backends
“So it's sort of always felt to me like, you know, in, in hindsight, it looks like the right idea, but applied to the wrong problems. And I always like wonder, you know, now, like I'm looking at like what we built at modal. Like I I'm kind of convinced that lik…”
Bernhardsson: Demands for self-hosting mean the product isn't compelling enough
“One trap that I think a lot of startups fall into is that, you know, you go out and talk to a lot of customers and they sort of insist on, on self-hosting. In my opinion, like, you know, sometimes that's actually like a true, like hard concerns, but in many ca…”
Bernhardsson: Lower GPU prices increase value of developer AI software
“If GPU prices were to crash, like the relative value of that software actually goes up. So like, I'm not necessarily sure that like a crash in GPU would be bad for us. I think it actually could be great for us because, you know, if GPU prices go down, there's …”
Distributed data systems impose a massive complexity tax over single-node hardware
“There's just this huge tax that you pay to have to build a distributed system that scales out and can do, like, you know, distributed transactions and shuffling data and, you know, if you were gonna design something for kind of modern hardware, you could make …”
Socher: Startups cannot be 10x better than Google on simple consumer search
“If you think about it from first principles, there is nothing you can do to be 10 X better.”
Pomel: Month-to-month contracts beat multi-year deals for validating new products
“So for new products and new companies, I would argue, month to month is great. Because your customers can try at any time which means you'll get the hard reality to hit you in the face and you can't ignore it, you know, which is a problem when you have, so you…”
Pomel: Product AI chatbots require handholding as users struggle with prompts
“It works great for some use cases, but you need to handhold the users a lot more. Like, users don't necessarily know exactly what to ask for and how to ask for it, or what to go next, you know, once they've asked a question.”
RAG and prompt engineering are just search techniques, not real AI
“Today when people are running RAG or prompt engineering those are search. That's not AI. It's like keeping the AI frozen and fixed.”
Aaron Katz: Cradle-to-grave sales reps boost efficiency and customer retention
“If you can enable your direct sales team to kind of handle a lead from cradle to grave I think you're going to drive efficiency as part of your sales organization. I think you're going to improve your customers buying experience and you're going to have higher…”
Dines: Encoder-decoder models beat decoder-only models on semi-structured documents
“We believe that encoder-decoder models are better for when you deal with semi-structured documents than decoder-only models.”
Howie Liu: UX is now the main bottleneck to enterprise AI adoption
“And I actually think now the bottleneck to their to their application is UX. It's actually giving people the tools to very easily deploy those models into not just a few Very central core use cases within a large company, but actually into the fractal of every…”
Howie Liu: No-code remains necessary because non-technical users can inspect AI output
“The fact that no code by definition means that the output of the Turing bot, the AI builder, Is actually in a form that a non-technical user can understand. So the very person who's asking the AI to generate the app can then understand fully its outputs.”
Liu: Existing AI models hold 10,000x untapped enterprise value without further upgrades
“In fact, my thesis is you could freeze all model development in its current state and say, you know what? GBD five is never going to come out. Like there's not going to be any better models. The models we have are all we're going to get for the next 10 years. …”
Howie Liu: Every PLG company eventually hits a growth asymptote
“You know, which is not to say you have to be PLG to start with, but I think every PLG company, and this is true, whether you started in Zerp or after I don't think it's going away, but I think if you start with PLG, you will eventually run into an asymptote of…”
AI focus creates pricing arbitrage in non-AI startup categories
“There's probably some pricing arbitrage in other categories where people aren't paying attention.”
Turck says AI evaluation startups currently lack significant enterprise traction
“A lot of those companies to put it bluntly don't have a lot of traction yet, because it turns out that, if you're trying to do something, and I can, I'm not picking on them, but like evaluation or monitoring and all the things, well, you need to have LLMs to m…”
Turck says vector database viability depends on incumbent database feature parity
“Do you really need a, you know, a specialized database to do that? Or in a context where some of the more general purpose players like the MongoDBs of the world have started announcing vector capabilities, you know, is that good enough? The answer to that ques…”
McGuire: AI advances stem from data preparation, not new models
“A lot of the learnings in AI are not on the technology side. It's not about creating a new AI model from scratch. It's about learning how to prepare data, how to augment the data so that you don't need quite as much of it, learning how to normalize or regulari…”
Evans: Every enterprise software company is unbundling Oracle, Gmail, or Excel
“Basically every enterprise software company for the sake of argument is unbundling Oracle, Gmail, or Excel.”
Evans: Major software failures stem from institutional rot, not missing regulation
“You don't look at this and say, well obviously the solution is that we needed to have a database regulator that makes sure that databases don't have bugs. Rather you look at it and say, well, this is an institutional failure. A in Fujitsu in the post office an…”
Cold outbound sales generates manual awareness, not buyer demand
“It's because cold outbound doesn't generate demand. Cold outbound generates awareness manually among, you know, otherwise highly targeted folks.”
Radiologists won't be replaced by AI, but by radiologists using AI
“I think radiologists will not be replaced by AI. They will be replaced by radiologists using AI.”
AI engineering skills do not easily transfer across narrow domains
“I have learned the hard way that it's not that transferable. You actually need, if you're trying to push the state of the art, you need people who have proven experience in the very, very specific narrow scope of problem that you're trying to solve.”
Traynor: AI builders must automate complete outcomes rather than individual minor tasks
“We shouldn't be approaching AI from just simply, how can we automate a task? Or how can we automate a little step? Or like, you know, how can we summarize this paragraph? Or you have to actually look at how can we automate the outcome?”
Traynor: The era of SaaS is an era of "right-click view source"
“The era of SaaS is an era of like right click view source. You know, ultimately like if there's a single thing you've done, other people can do it. The job is to build a brand around having the best product. And to do that, you often, you have to have the fres…”
Handy: Hyperscalers build features for RFPs, not to kill independent tools
“When you talk to the folks at the hyperscalers, they will say, we are, we're like making sure that our solution is like covers all these different areas, but it's mostly not. Because we want to compete in these areas. It's mostly because people come to us with…”
Handy: Building in-house is often faster than M&A for data infrastructure
“Had to build, you know, there's like the iceberg, which is 10% above the water and you can see it. And then there's 90% below the water for a data company. The platform is the below the water part. And then the functionality that you build on that platform is …”
Tristan Handy: Software multiple expansion has shifted from modern data stack to AI
“If what you were trying to do is maximize the multiple that your software company was trading at, then AI is the enemy because all the multiple expansion has moved from modern data stack to AI.”
Van Luijt: RAG carries less hallucination risk than model fine-tuning
“That is something that is, works better than fine-tuning, for example, because if you fine-tune, then you're still dealing with potential hallucination Fair enough, with RAC that's possible too, but it's like, it's less it's less risky.”
Alex Rinke: Hiring a large sales force is not a GTM strategy
“Just hiring a lot of salespeople, right? It's not a go-to-market strategy. And that's what a lot of startups did, right? Hey, we have just got a hundred million dollars in funding. We're going to hire tons of salespeople. They know how to do this. Good luck. T…”
Alex Rinke: Enterprise founders must personally sell their first $2 million
“If you didn't sell the first million or two in revenue yourself, Don't expect anybody else to do it, right?”
Kant: Programmatic RL can scale magnitudes larger than human feedback
“And it's that RL loop that is very interesting because since it's programmatic, since we have an Oracle of truth, we can scale this up far larger, right? Magnitudes larger than what you can do with human feedback today.”
Scott Belsky: Contextual UI prompts matter as much as underlying AI models
“One of the most important insights was, and maybe as important as building a great family of models. And I can talk more about Firefly and the underlying models and how we trained it and how we didn't train it and everything else, but was simply the insight of…”
Belsky: AI detection algorithms will never catch media manipulators
“We're never going to catch the bad actors because anyone can, you know, make anything and it's a cat and mouse game to have algorithms that tell whether something was edited or not.”
Zhou: Zero-shot LLMs can replace manual human labeling in RLHF workflows
“Which is that why can't it be another LLM or a pipeline of LLMs that can help with that feedback? I think manual labeling is very tedious, especially for our target user, which is a software engineer. And I don't think people should necessarily have to do all …”
Incumbents will dominate text-to-SQL, making it a bad startup problem
“The bigger companies, like Databricks or Snowflake might build these tools and offer it on their sequel editor or, like, some kind of, like, you know, assistant And that could probably solve the problem. It's not a great problem for a startup to work on.”
Vertical AI applications offer startups the best long-term competitive moats
“Vertical applications give you the best chance at creating a longer term competitive mode”
High GitHub star counts for AI projects mostly reflect unmonetizable hobbyist interest
“There are definitely some that have, you know, tens of thousands of stars on GitHub. But if you look at the use, where those stars are coming from, it's mostly hobbyists, and I'm sure you can monetize hobbyists and sort of grow from there. But I think it is li…”
Polu: A single all-knowing enterprise AI assistant is currently far-fetched
“Context size and model quality makes it such that an assistant that knows it all within a company is, is, is, is still kind of a bit farfetched because it'll get confused if it sees too much information.”
Polu: Building a strong AI research team is 10x easier in Paris than SF
“If you want to build a strong AI research team today, it'll be 10 times easier to do it in Paris than it is to do it in SF with OpenAI as a lab, as a competitive lab to, in the same hiring markets.”
Kanjun Qiu: AI emergent capabilities are artifacts of evaluation metric design
“If your metric is smooth you actually see slow performance improvement over time, and if your metric is relatively discrete or not smooth, that's where you see the emergence, and it's actually more about the evaluation metric Than about the emergence of the ca…”
Production LLM deployment challenges mirror autonomous vehicle development
“The trajectory of issues and concerns that people are running into are similar to, you know, like the path is similar to what it was in self-driving, which is, you know, How do I get like this, you know, runtime safety? How do I get like runtime constraints, e…”
Existing model risk frameworks fail for third-party AI models
“Existing, for example, like model risk management frameworks don't really apply when you haven't built the model yourself. You know, you didn't like curate the data that the model was trained on. And so you can't make any claims to that.”
Fine-tuning cannot eliminate LLM hallucinations
“At the end of the day, these models are like next token predictors, which is, you know, like they kind of look at like what they've predicted until now, and then, you know, figure out like what the next token they're on is. And from that, like, even with fine-…”
GenAI solves ML's first mile, but traditional tools handle the last
“LLMs and Generative AI really helps solve, like the first mile problem in ML, right? But the last mile problem, which is, like, how do you take this generic generalizable technology and make it work, like, specifically for your use cases and for your actual ap…”
Taylor: Community building provides more startup growth leverage than marginal technical improvements
“Sometimes you can over index on the technical and getting maybe, you know, your platform to be five percent more fast or supported on one more browser might not 10 X you the way that spending that time building a community would.”
Socher: Don't train LLMs from scratch without hundreds of millions of responses
“And yeah, I don't think you should attempt it unless you have, you know, hundreds of millions of responses that to really like train your own model versus just fine-tune it a little bit. Like certainly not from scratch.”