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 81 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
Diana Hu: Supabase is the default backend recommendation across all LLMs
“Whenever someone asks how to set up anything that you need, some sort of backend, Firebase type of transaction, the default answer from all the LLMs is actually a Superbase.”
Hu: Anthropic surpasses OpenAI as top API for YC applicants
“And, shockingly, in this batch, the number one API is actually Anthropic. It came out a bit more than OpenAI, which who would have thought?”
Diana Hu: AI's future Google- and Facebook-scale startups have not started yet
“Startups like the future Facebook or the future Google are yet to be started because those come in in the deployment phase, because right now I think these things, things are still getting built up.”
Hu: A YC startup's 8B model beat OpenAI on healthcare benchmarks
“There's this particular YCE startup that told me that they collected the best data set for healthcare. And they ended up performing better than OpenAI and a lot of the benchmarks for healthcare with only eight billion parameters.”
Hu: Cursor scaled from $1M to $100M revenue in one year
“We've seen crazy growth unlike anything, only possible with, right now with AI, like all these companies that we work with, zero to twelve million in 12 months. The cursor story where they went zero to one in one year, the next one to a hundred. This is, like,…”
Hu: Software engineers will transition to product engineers
“I think one of them that I can read verbatim is, I think the role of software engineer will transition to product engineer. Human taste is now more important than ever as CogenTools make everyone a Tenex engineer.”
Hu: YC founders rarely use Devin for serious features
“Notable, Devon does get mentioned, but the drawback of Devon not really being used for serious features is that it doesn't really understand the code base. It's being used mostly for small features and barely, it's like barely mentioned.”
Hu: Vibe coding works for zero-to-one, but scaling requires systems engineers
“Zero to one will be great for vibe coding where founders can ship features very quickly. But once they hit product market fit, they're still going to have a lot of really hardcore systems engineering where you need to get from the one to N and you need to hire…”
Hu: Top 1% engineers will still require deliberate practice despite AI
“I think there's gonna be a generation of software engineers that are, like, good enough because it's so easy to retool there with all these cogent tools. Like, the barrier is so low. You're gonna be good enough engineers. There's gonna be tons of those. But to…”
Hu: 10T parameter models will spark a GPT-3 level innovation leap
“I think the type of level of potential innovation could be the same leap we saw from GPT-II, which was around one billion Parameters that was released with the paper of a scaling loss, which was one of these seminal papers that people figure out, okay, this is…”
Hu: AI has effectively passed the Turing test for phone calls
“At this point, AI has passed Turing tests and is solving all of these very menial problems over the phone.”
Hu: Customized open models will beat frontier LLMs in specific domains
“So there's this other world where the open model that's customized, I think, is gonna win and compete versus the big one for specific domains.”
Diana Hu: Command-line interfaces are becoming the right UX for AI agents
“CLI seems to be becoming the right UX for agents. This is why I think going back to ClockCode versus Cursor ended up doing so well. It's just so much more freedom being fully on the command line.”
Hu: Confluence Labs Saturated the ARC-2 Benchmark at 97% Accuracy
“I actually worked with a company in the winter 26 batch not too long ago called Confluence Lab, which actually ended up saturating the V-two results with 97%, and I think their task cost was a lot more efficient too.”
Hu: Google Gemini adoption in YC batch rose to around 23%
“I think last year was probably single digit percent, or even like two, three percent, and, ah, now for winter 26 is about 23%.”
Diana Hu: Series B AI startups are orchestrating and arbitraging models
“Startups are doing as well, they are actually arbitraging a lot of the models. I had some conversations with a number of founders where before they might have been loyalists to, let's say, OpenAI models or Anthropic, and I just had some conversations recently …”
Hu: Voice AI Startups Meta-Prompt with Frontier Models Before Distillation
“I think that's a common pattern sometimes for companies when they need to get responses from elements, elements in their product a lot quicker. They do the meta-prompting with a bigger, beefier model, any of the, I don't know, hundreds of billions of parameter…”
Hu: Happy Robot Closed 7-Figure Deals with Top 3 Logistics Brokers
“Happy Robot, who has sold seven figure contracts to the top three largest logistic brokers in the world. They're built AI voice agents for that.”
Hu: OpenAI reasoning models are catching up to Claude 3.5 Sonnet
“The thing about CodeGen, the big game in town that we saw, ah, six months ago was Clawed Sonnet. It's still actually a big contender. Most are still using it. But O-one, O-one Pro, and O-three meaning all these resilient models are starting to see it's almost …”
Hu: GPT-4o sees virtually zero use for code generation among founders
“The other one is four-oh, virtually no use for CodeGen.”
Hu: Diode demonstrated that OpenAI's o1 can automate PCB system design
“But the thing that they demonstrated now with O-one was actually able to do the system design and component selection, which is crazy. So it would be able to read all the data sheets and select the right components.”
Hu: OpenAI's o1 solves Navier-Stokes equations for airfoil design
“O-one was actually able to write All of these equations, all these partial differential equations, and solve basically Naive Stokes questions to actually solve airfoil.”
Hiring engineers at the MVP stage slows launches and blocks insights
“It's gonna actually slow you down in terms of launching quickly, because if you're hiring from a pool of people and engineers that you don't know, it takes over a month or more to find someone good, and it's hard to find people at this stage where it's very ne…”
Early tech stack choices matter less if users adopt the product
“If you build a company and it works and you get users good enough, the tech choices don't matter as much. You can solve your way out of it.”
Pre-product-market fit startups should tolerate technical debt and broken code
“Tech debt is totally fine. You gotta get comfortable a little bit with that heat of your tech burning. Totally okay. You're gonna fear the right things, and that is towards getting you product market fit.”
Hu: AR core technology startups currently have better odds than app companies
“It is personally thinking, betting more on the installation phase kind of style of companies, which have to do more with our core technology that gets built out. Because a lot of those, there's tons to do. There's so many problems to solve, ah, before any of t…”
Diana Hu: Standard vibe coding tools fail to maintain design consistency
“When you try to create something like this with any other vibe coding tool, it's not consistent. It kind of remixes and forgets all the design consistency”
Hu: RT-X proved AI scaling laws apply to physical robotics
“That was a big paper, because it was the first that showed potential scaling laws that applied to robotics, because now you could start training all these models across multiple kinds of hardware, not just one, which has never been done in robotics ever before…”
Diana Hu: Opportunity exists for an agentic-first system of record
“And the system of record, there's opportunity for something that's kind of agentic first, because right now we're still kind of integrate very much with databases and SQL or NoSQL queries at a very low level. But imagine something that generates all the data t…”
Diana Hu: Getting Good Prompts Requires Test-Driven Development via Evals
“The way you get a good prompt is all test-driven, just like evals, right? In a sense, the test cases are your evals.”
Hu: Anthropic saw over 52% growth among YC applicant tech stacks
“They had this hockey stick with the growth over 52%.”
Hu: Cursor usage in YC batches jumped from single digits to 80%
“I remember in 2023, I think it was maybe single digit percentage of the batch we used cursor, then 2024, it was like 80%.”
Diana Hu: XML Prompts Produce Better Output Due to Post-Training RLHF
“We found that it makes it a lot easier for LLMs to follow, because a lot of elements were post-trained in RLHF with kind of XML type of input, and it turns out to produce better results.”
Hu: Even top LLMs struggle to detect N+1 query bugs alone
“It's actually hard for today for even, like, the best LLMs to find those”
Hu: Example-based prompting is the LLM version of test-driven development
“I think this pattern of sometimes when it's too hard to even kind of write a prose around it, let's just give you an example that turns out to work really well, because it helps LLMs to Reason around complicated tasks and steer it better because you can't quit…”
Hu: Claude is naturally human-steerable while Llama requires heavy prompting
“One of the things that's known a lot is Claude is sort of the more happy and more human steerable model, and the other one is Lama. Four is one that needs a lot more steering. It's almost like talking to a developer, and part of it could be an artifact of not …”
Hu: Cursor is the undisputed leading AI code editor among YC founders
“The vibe started to shift back in summer 24 when cursor was being used by a big portion of the batch. And now, by far, is the leader.”
Hu: YC survey shows AI tools struggle with systems engineering
“Based on our survey, current tools are not good at that low-level systems engineering.”
Hu: Effective vibe coding still requires engineering taste and debugging knowledge
“In order to do good vibe coding, you still need to have the taste, and you still need
That kind of classical, maybe not necessarily classical train, but enough knowledge to judge what's good versus bad.
And you only become good with enough practice.”
Hu: Top AI Products Use Multi-Model Pipelines Splitting Extraction and Reasoning
“I think this is actually a very common pattern that we're seeing a lot of the interesting products built with AI. You use different kinds of models. So, yes, four O mini is for PDF extraction, and then O one for the reasoning, because it's actually very hard t…”
Hu: OpenAI o1 Combines Next-Token LLMs with RL Reward Functions
“GPT is all generative based on predicting the next token and patterns and then getting those results to check that they're correct. So I think a lot of it is you had to have a lot of data that was factually correct and Fed into probably the model and the train…”
Hu: OpenAI o1 unlocks startups in physical engineering and biology
“To connect to Sam's essay is a lot of things that are going to make the atom world, physical world better because it's really good at math and physics. So any startup that's working around mechanical engineering, electrical engineering, chemical engineering, b…”
Hu: Leading frontier AI models operate around 500 billion parameters
“Yeah, for a bit of context right now, the frontier models, I mean, they're not public, exactly how many parameters they have. But they're roughly in the 500 of billions-ish, like Lama-III, four or five billion. Anthropic is speculated to be five hundred billio…”
Hu: A YC startup built a Replit Agent clone in hours using o1
“They basically got a version of Replit Agent working with the product. All they had to prompt O-one with was all their developer, actually, some of their developer documentation and some of their code, and they could just prompt it, build me a web app that wri…”
Hu: YC startup DryMerge reached near-100% accuracy by adopting OpenAI's o1
“There's a company Dry Merch. That you work with, and they went from 80% accuracy to pretty much a 99 or for intents and purposes, a hundred percent using a one and they unlocked a bunch of things.”
Hu: OpenAI's o1 massively increases GPU compute requirements for inference
“I think the other thing that's interesting about one is that it makes a lot of the GPU needs even bigger because it's moving a lot of the computation needs a lot higher for inference because it's taking a lot more time to do a lot of the inference. So I think …”
Diana Hu: Top startup founders are quirky outliers, not well-rounded
“There's this aspect of the best people at the top of their careers. They're actually very not well-rounded. They're very quirky people. For good reason. And that's what makes them outlier. Like, by definition, if you're average, Then it's like, okay, you're no…”
Hu: Big-company multi-year secret launch playbook does not apply to startups
“And then founders think, oh, I did this at a big company. Of course, this is how my product needs to be done, because when I was an engineer at Apple, I had to work for, let's say, the Vision Pro project. I was like, I don't know. Eight years, a decade of work…”