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 1,065 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
Dan Fu: DeepSeek-V3 was trained on ~2,000 H800s with 20% MFU
“If you look at the deep seek model, for instance, this is one of the best open source models we have out there today. It was trained at the end of 2024. On last generation, kind of nerfed GPUs, H 800 instead of H 100, the 800 is nerfed by all sorts of ways fro…”
Izmailov: Duration of tasks AI can robustly automate doubles every six months
“There is this famous meter plot, which shows how long of a task AI is capable of robustly automating, and it's been kind of consistently doubling at that time every half a year, I think and it's now in like some hours so maybe a couple hours.”
Kaiser: Translation industry grew and translators earn more post-Transformers
“The translation industry has grown considerably since then. It has not shrunk. There's more translations to be done. Translators are paid more.”
Kant: Poolside first compute online late 2025, 250MW by Q1 2027
“So at the beginning of Q four, end of Q three, so next year, we've got the first compute coming online. And then we're finalizing in Q one, 20, 27, the first you know, 250 megawatts, but already next summer, the next 250 megawatts starts construction. It's a s…”
Kant: Poolside's Malibu agent matches initial Gemini 2.5 Pro SWE-bench scores
“But if you think about the Malibu agent as a coding agent, for instance, right now, it sits at a level of like sweet bench verified, for instance, where Gemini two and a half pro was when it came out.”
50% of Hugging Face model derivatives are now based on Qwen
“I think 50% of all model derivatives being downloaded from Hugging Face or Quen base now.”
AI search traffic converts at higher rates than direct e-commerce traffic
“Traffic that's generated through conversations and AI search onto a commerce platform converts high to higher level than direct traffic.”
Corporate AI subscription retention rose from 50% to 80% since 2022
“Across a segment of 43,000 or so US customers, we work with ramp to show that retention of subscriptions on AI products across this customer set has really improved markedly since 2022. On 2022 is around the 50% after 12 months. And now in 25, it's hitting aro…”
Average corporate AI spend hits $500,000 and will reach $1M
“The total spend on AI products per customer kind of went up from 35,000 dollars or so maybe two years ago. Now it's around half a million dollars, and it's predicted to hit a million dollars next year.”
NVIDIA stock outperformed private AI chip startup investments 12x to 2x
“If you bought NVIDIA stock on the day of the announcement of all the, like, private rounds in these companies, what would the value of your stock be in NVIDIA versus these companies? And if I recall correctly, it's basically 12 X in NVIDIA versus two X in in t…”
NVIDIA's sovereign AI business generates around $20 billion in revenue
“Nvidia has even started marketing this as like a new kind of product line basically for its business that currently generates, I think, around twenty billion dollars worth so it's real money”
OpenAI holds AMD stock warrants tied to a $600 share price
“There's like another financial sweetener in the deal where OpenAI has warrants in AMD if the stock price hits 600.”
ProFluent developed the first AI language model-created CRISPR genome editor
“ProFluent, which is kind of leading the charge for these language models in protein design, developing the first, like CRISPR genome editor that an AI has created.”
Reinforcement learning scaling yields returns on compute similar to pre-training
“If you look at all the RL literature over time, we see very similar returns on compute in pre-training and in RL, where we can invest exponentially more compute in RL and keep getting benefits.”
Douglas: Anthropic models autonomously replicated the Claude.ai website in hours
“And in this case, the model replicated Claude.ai with artifacts, with everything else I can't quite remember how long that one took. Maybe a couple hours to do.”
Douglas: AI progress on METR evaluations plots as a straight line
“Like on the meter eval, if you look at progress over the last two years, you can plot it with a straight line, right?”
Douglas: Cognition rebuilt Devin's architecture around Anthropic's highly useful Claude Sonnet
“The Cognition folks in Devon found the model so useful, they had to, like, rebuild their architecture around it.”
Turck: The Information reports Claude Code hit $400M ARR in six months
“There was an article in the information just a couple of days ago that was saying that Cloud Code, which really came out, what, at the end of February of this year, so five, six months ago, is already generating four hundred million in annualized revenue. That…”
Misha Laskin: Engineers spend 70% of their time searching information, not coding
“If you actually hover over a shoulder of an engineer at any large organization, or even a startup with a really sizable code base, and you look at what they do, you'll find that a minority of their time is actually spent coding. Like, 70% of the time they're D…”
Rauch: Anthropic's Claude Code uses Vercel to deploy applications
“But nowadays when you ask Claude Code Anthropics agent to deploy, They use for sale because we had, you could argue accidentally created the perfect tool for an agent to deploy.”
Rauch: Multiple security incidents have already occurred with MCP servers
“In fact, there's already been quite a few incidents with MCP servers where depending on what chat message was sent, the MCP tool that was called was made to disbehave.”
Rauch: Vercel AI SDK is second largest AI module behind OpenAI
“It's the second largest AI module behind OpenAI.”
Canva established a $200 million fund to compensate AI training contributors
“We do have a creators fund I think it's two hundred million dollars and it's designed to reward creators.”
Dohmke: Most developers have adopted AI tools, unlike other knowledge workers
“Most developers have adopted some form of AI tooling into the workflow already, while other knowledge workers, other white collar workers are still early in that adoption journey.”
Gomez: Google quickly integrated Transformer architecture into Search and Translate
“No, they jumped all over it. So it went to production inside of search, inside of translate, like the existing product suite. And so they, to say that they didn't adopt the transformer architecture would not be correct.”
Evans: Google and Meta delayed LLMs in 2022 due to high error rates
“This is why Google and Meta didn't launch their own LLMs in twenty-twenty-two when they had them as well, because they looked at them and said, well, they're wrong too much.”
Amazon approached InfluxData for a partnership due to Timestream customer demand
“Last year we did a very unique deal with Amazon. Where they have Amazon's time stream, but their customers kept asking them for influx. And so they approached us.”
Chollet: MindAI dropped out of ARC Prize over open-source requirements
“They ended up dropping out because they did not want to open source their solution. And of course, that meant that they were not eligible for the prize.”
Kiela: FAISS was the first vector database
“In the initial paper, we used a vector database or a face. So the words vector database didn't exist at the time. But so face was the first vector database.”
Gaurav Misra: TikTok's US growth was heavily fueled by Snap ads
“TikTok spread in the US mostly through Snapchat, and this is like not well known out there today, but a lot of it was they were spending a very, very large amount of money on advertising on Snap, right?”
Mike Murchison: LLM model costs plummeted 90% in 12 months
“And as you know, well, you know, we're seeing model costs plummet. They've plummeted 90% in the last 12 months.”
Replit Agent was the first AI programming agent on the market
“So we got to September, we launched it in early access, and at the time, there was no other programming agent on the market, so it was really the first.”
Replit trained a 3B model that outperformed Google PaLM in coding
“We trained this three billion parameter model. It was the first three billion parameter model that was state of the art in coding and open source coding. It did better than all the other open source models. It actually did better than Palm on coding.”
Masad: Mark Zuckerberg is reversing identity-based hiring at Meta
“You see Zuck, you know sort of bringing fast Facebook back, or meta back to the hacker culture and reversing a lot of the sort of identity-based hiring and identity-based sort of promotions and all of that stuff.”
Borgman: Presto is mostly Facebook internal, while Trino is the mainstream branch
“They started as identical copies, but the code bases have diverged quite a bit. And today Presto is really just used by Facebook. So it's sort of like their own private branch in a way used by a small number of people. And Trino has become the mainstream commu…”
Vector RAG requires full re-indexing for single-file updates, Habib claims
“And when that policy, even one policy gets updated, you're throwing away the entire embedding store and starting over.”
Enterprise AI app 12-month retention jumped from 43% to 65%
“Just directionally, it was like retention after one year in 20, 22 cohort was like 43%, I think. And then the cohort from 20, 23 after 12 months was like 65%. So that was pretty material.”
AI inference costs plummeted from $60 to $0.06 per million tokens
“The most expensive models like a year ago to now, I don't know, the most expensive one is maybe like 600 dollars, sorry, 60 dollars per million tokens, and now the cheapest one is probably like 0.06 dollars per million tokens. Not the same level of intelligenc…”
Nvidia chips appear in 11x more AI research papers than all competitors
“If you sum all the papers that use NVIDIA chips versus all the papers that use TPUs, FPGAs, ASICs, Apple, Huawei, the sum of all NVIDIA papers versus the sum of everybody else, the delta is like 11 times.”
OpenAI's $6.6B round required a $250M minimum ticket
“In the OpenAI round that we're talking about, you know, the 6.6 billion dollar round. The minimum ticket size was two hundred and fifty million.”
Accenture reported $3B in annual generative AI consulting fees
“Famously Accenture reported three billion dollars in in annual fees which You know, made the headlines and all the things.”
OpenAI's $6.6B round was the largest VC round ever
“We just had the biggest venture capital round of all time with OpenAI, which raised 6.6 billion at one hundred and fifty seven billion post-money valuation.”
Safe Superintelligence raised a $1B seed round pre-product
“There's also kind of the biggest seed round ever, right, that we saw. Safe superintelligence helmed by Former OpenAI chief scientist and co-founder Ilya Sutskover raising roughly a billion dollars in cash at a five billion dollar valuation pre-basically anythi…”
Google's $2.7B Character.AI deal was the largest acquihire ever
“And, you know, we also had the biggest acquihire of all time. So for those of you who don't know, 2.7 billion dollar acquihire by Google of the Character AI founders who, funny enough, were already at Google before they went to start Character AI.”
xAI built its Memphis Colossus data center in 122 days
“Colossus, which is based in Memphis, which by the way apparently was built in a 122 days. So the fastest you know, time ever to build a data center.”
Big tech and AI companies are driving a U.S. nuclear revival
“But yeah, in 24, it turns out that big tech companies and AI companies are the ones driving this, Revival of nuclear power with, you know, Microsoft entering into a deal with Constellation to revive Three Mile Island, which was the site of the last nuclear acc…”
AI startups reach $30M ARR five times faster than legacy SaaS
“The AI startups in the cohort that Stripe measured took 11 months to reach one million in ARR after their first sales on Stripe versus 15 for the previous generation of SaaS businesses, and then maybe more impressively, you know, scaling to more than thirty mi…”