Patrick Cheng

reminds me of some startup advice i heard recently: start early because the energy decline after 40 is real and noticeable

Talked to a senior AI researcher about the panic Astra is causing in robotics research circles. His read: Astra isn't closing the sim-to-real gap. It's speeding up real-to-sim. The gap comes down to how accurate the underlying simulator is, and you could always get there by throwing enough human hours at it. Astra makes building the sim 10-100x faster.

Fun thought experiment: once Anthropic and OpenAI IPO, many early researchers retire, and the rest are liquid enough to go anywhere. Does that talent finally push big-co catch-up labs into the frontier? Or does it flow to neolabs betting on weirder architectures?

Yes. Mech interp data has been a focus for a lot of labs. I know several data teams in this category that reached tens of millions in run rate quickly. Goodfire has also published a lot of excellent research here.

Start of the year I genuinely could not understand why so many neolabs were raising at multi-billion valuations. Now with the Anthropic IPO actually happening and the Yasmin Razavi success story being so real, if you're hunting for the next trillion-dollar company, neolab is still the most likely category. In 2023 Anthropic was a neolab too. So were Inflection, Adept, Character... "Pick the best neolab" printed a buttload of alpha.

Of the MTS I've talked to this month, more than ever expect to be jobless in 12-18 months. Some of the best say coding agents are closing the gap between them and the "geniuses" they used to admire.

Talked to the data leader at a frontier lab today (top 10 globally by data spend). One thing he couldn't square: if a data vendor's core value is building RL envs and tasks, and that's literally part of post-training, why can't labs just bring it in-house? My answer: incentives. (1) Money. A great team on the outside does $$M in revenue at 70%+ contribution margin. Very hard to pay that inside a lab. (2) Prestige. At OpenAI and peers, the data team is not where top MTS go. So you lose on status too.

Two loops on AI capability rn: Loop 1: autonomously doing complex, high-value white-collar work inside real companies. Loop 2: raw intelligence. Even this camp can’t agree on what that means. Natural science, or a model that trains itself? AI4S-pilled vs RSI-pilled.

Stream of thoughts on the data line in Anthropic's S-1: (1) Roughly $1B in 2026, and it's data + evals. - The line item is literally "vendor costs related to data quality and performance evaluation services." - They never disclose the level. Only YoY deltas: +$165M in 2024, +$165M in 2025, +$251M in H1'26. - In 2023 the entire opex line it sits in was $62M, which caps the starting point. -> Back out the levels and 2026 lands around $900M-$1B. My estimate, not a disclosed number. (2) Post-training is where the demand is. - In their words, post-training has "evolved from a minor adjustment step into one of the largest areas of our research investment." - It's built on "complete, realistic tasks whose complexity reflects the demands of real professional environments." - What matters in that task portfolio: breadth and realism. (3) The target is long-horizon work. - Capability focus: coding, agentic workflows, computer use, complex reasoning. - Goal: models that run autonomously on tasks "lasting hours, days, or weeks." (4) It's going physical. - They built a molecular biology lab, do robotics research, and run their own lab facilities.

If you're selling chip design or FPGA data, dm me. I can connect you with several frontier labs.

Claude Duet is an impressive launch, and not only because of the product. The opsec was elite. Every MTS I talked to over the past few weeks was convinced Anthropic had ZERO interest in proactive agents / personal assistants, because "that has nothing to do with frontier intelligence." Then they ship one, and build the entire value prop around proactive. Right call. That's the real differentiator of this whole product wave. Biggest lesson from my time at ClickUp: the users who matter are the ones who never imagined using your product to get more productive. They will never open a chat box and ask. Proactive is how you reach them.

Stream of thoughts on the data biz: it's going anti-scale. Last gen was about landing giant lab contracts and running massive labeling ops. This gen looks like 3-4 person teams taking small, sharp orders. (1) The core skill changed. - Scale / Mercor era: can the founder land the big lab PO, then recruit and manage an army of labelers and experts? An ops + sales game. - RL env / vertical era: do you understand one domain well enough to build precise tasks and clean data for it? An expertise game. → Headcount used to be the moat. Now it's mostly overhead. (2) Post-training keeps getting harder and more stratified. - The bar for RL data goes up every quarter, and it splinters into narrow specialties. - A small team living inside one task does it better AND faster. A 1000-person vendor can't turn on a dime. (3) Researchers control the budget now. - Spend isn't one big centralized procurement decision anymore. Individual researchers can route budget to the vendor they like and buy exactly the data they want. → Your buyer is a researcher with opinions, not a procurement team. (4) AI made the labor optional, so the incentives flipped. - Synthetic data got way better. You need far fewer humans in the loop. - So it starts to look like boutique advisory: 3-4 people, very real $$. Anyone good does the math on what they'd make at a big shop and leaves to do it themselves. Put plainly: data used to reward scale. Now it rewards domain depth, and that doesn't scale the same way. Expect a lot more tiny data shops generating both $$$ and great value to the frontier intelligence.

Super impressive work from @NunchuxAI team! Batch video gen is a cost game. Real-time video gen is a product game. Once 5s renders in 1.3s and you can steer the prompt mid-stream, video stops being a render queue and becomes an interface. Good to see MiniMax-H3 get there on AMD rather than only on the newest NVIDIA parts.

The offline trace editing setting is the real contribution here. It turns a game into a verifier in the loop optimization problem, same shape as autoresearch or kernel tuning, and it drops the vision requirement so open weight models can finally be compared on the same footing.