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 19 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
Wu: OpenAI deployed o3 on an air-gapped Los Alamos supercomputer
“We actually did a custom on-prem deployment with them onto one of their supercomputers called Venado. And so this actually involves a bunch of, you know very bespoke work with some FDs also with a lot of our developer team. To actually bring one of our reasoni…”
Wu: GPT-5 hallucinations dropped to near zero on certain benchmark evaluations
“I think there was an eval that showed that hallucinations basically went to zero for a lot of this.”
Wu: Short on the entire AI tooling startup category
“I'm short on the entire category of like tooling
around AI AI products.”
Wu: Short on reinforcement learning environment startups
“RL environments I think are really big right now as well.
Unfortunately, I'm very short on those.
not really I don't really see a lot of potential there.
See a lot of potential and reinforcement learning and applying it, but I think the startup space around R…”
Wu: Los Alamos OpenAI supercomputer deployment is shared with Lawrence Livermore and Sandia
“The other cool thing is it's actually being shared between Los Alamos and some of the other labs Lawrence Livermore Sandia as well because it, it's the supercomputer setup where they can all kind of connect with it remotely.”
Wu: Failed Enterprise AI Deployments Usually Lack Data Scaffolding
“My hunch is some of the enterprise deployments that don't actually work out likely don't have the scaffolding or infrastructure for these agents to interact with as well. A lot of the, like, really successful deployments that we've made, a lot of what our FDs …”
Wu: GPT-5 Pro solves previously unsolved problems but takes ten minutes
“These like unsolved problems that none of the other models could handle, you throw out a GPT-V Pro, and it just like one shots it is pretty crazy, but the trade-off here is you're waiting for 10 minutes.”
Wu: GPT-5 solves coding problems no other AI model can solve
“Especially for like coding use cases, especially at the, you know, at the, when it thinks for a while, it'll usually solve problems that no other models can solve.”
Wu: GPT-5 enables single-generation code completions in the Codex CLI
“The second thing, honestly, is, is GPT-V. Like, I just think GPT-V really allows the product to shine. It's, you know, at the end of the day, this is kind of a, this is a product that really is dependent on the model, underlying model. And when you have to, yo…”
Wu: OpenAI is currently operating in a GPU crunch
“We are in a GPU crunch, so we'll see how, you know, how long that goes.”
Wu: ChatGPT is roughly the fifth largest website in the world
“ChatGPT obviously is really, really, really big now.
[233] Sherwin Wu: It's I think like the fifth largest website in the world.”
Wu: The majority of the startup ecosystem builds on OpenAI's API
“The biggest product that we have is obviously our developer platform, which is our API.
[271] Sherwin Wu: You know, many developers, you know, the majority of the startup ecosystem builds on top of this, as well as a lot of digital natives, Fortune 500 enterpr…”
Wu: Enterprise AI Evals Must Be Built Bottom-Up by Operators
“And evals also, oftentimes, need to come up bottom up. Right? Because all of these things are kind of in people's heads, in the actual operator's heads. Like, it's actually very hard to have a top-down mandate of, like, you got, like, this is how the evals sho…”
Wu: Product users often prefer instant substandard AI answers over 10-minute waits
“As a product builder, there's a latency, there's a real latency trade-off that you have to deal with where, you know, your user might not be happy waiting 10 minutes for, like, the best answer in the world. It might be more okay with the substandard answer and…”
Wu: Reinforcement fine-tuning is an order of magnitude more powerful than SFT
“Reinforcement fine tuning introduces like RL or reinforcement learning to this loop. Way more complex, way more finicky, but an order of magnitude more powerful.”
Wu: GPT-4 already existed internally at OpenAI by September 2022
“The first one was right when I joined the company in September, 20, 22. We, it was pre-TiGPT. But at the time, GPT-IV already existed internally.”
Wu: OpenAI's original product was its developer API, not ChatGPT
“When I joined OpenAI around three years ago to work on the API, it was actually the only product that we had.
[178] Sherwin Wu: So I think a lot of people actually forget this, where the original product for, from OpenAI actually was not ChatGPT.
[183] Sherwin…”
Wu: OpenAI used T-Mobile deployment learnings to improve core Realtime models
“And a lot of the improvements that we actually got into the model came out of, you know, the learnings that we have from T-Mobile. It brings in a lot of other change from other customers, but because we were so deeply embedded into T-Mobile and we were able to…”
Wu: Accordance achieved SOTA results on Tax Bench via OpenAI RFT
“There's another startup called Accordance that's doing this in the tax space. I think they've been targeting an eval called Tax Bench, which looks at, you know, CPA style tasks as well. And because they, because, you know, they're able to turn it into a very g…”