Research · Almeida · DCVC

Company & founder

Last updated: 2026-09-16

TypeSafe AI

❓ Who is the company, where does it sit, and what did it raise?

TypeSafe AI is a San Francisco lab. The team page says the company works in person five days a week near Embarcadero station.[36] DCVC describes it as founded by veterans of “the world’s top AI research labs” and dates its first investment to 2025 on the portfolio page.[15] The public coming-out was 15 September 2026, with Jev in early access the same day.[1][15]

FinSMEs and DCVC both report a $40 million seed led by DCVC.[15][22] Forbes, paywalled here on fetch, is quoted in search as putting post-money at $200 million “according to a person familiar with the deal.”[16] Treat $40 million as confirmed and $200 million as a single anonymous source.

The manifesto’s mission line is “pave the shortest path to an AI-based economic revolution by making intelligence composable.” A footnote’s favorite definition of that revolution is global total factor productivity (TFP — output not explained by extra labor or capital) reaching 3% growth within five years and holding for ten.[3] Closing slogan: “We’re building prod, not God.”[3] That is a refusal of AGI-as-product, not a claim that AGI is impossible.

Culture copy lists alumni of OpenAI, Google Brain, Meta/FAIR, Stripe, Airbnb, Plaid, and Docker. That is a hiring-page list, not a cap table.[36]

Diogo Almeida

❓ What did the founder actually publish, versus what the launch thread said?

Confirmed from primary paper. The InstructGPT paper, arXiv:2203.02155 (4 March 2022), lists Diogo Almeida among the starred primary authors, with Long Ouyang, Jeff Wu, Xu Jiang, Carroll L. Wainwright, Pamela Mishkin, and the team leads.[20][34] InstructGPT is the work that showed a 1.3B model fine-tuned with human feedback preferred over 175B GPT-3 on the API distribution. That paper is the origin of the public RLHF-for-instruction recipe at OpenAI.

Confirmed from the AI Engineer speaker page. Almeida is CEO of TypeSafe; previously Google Brain and OpenAI; coauthor of InstructGPT; acknowledged among ChatGPT contributors; coauthor of the GPT-4 technical report; also coauthor of research on learned optimizers.[25] He spoke at AI Engineer World’s Fair 2026 on “What’s next after RLHF?”[43]

Company page. “Diogo co-invented RLHF and InstructGPT, the methods that lead to ChatGPT and GPT4. Previously, he was at Google Brain.”[36] Forbes says he spent four years on ChatGPT’s responses at OpenAI and left in 2024.[16]

X. Handle @CompleteSkeptic. Bio on launch day: “co-created RLHF/ChatGPT @ openai now trying to right the wrong 🤭 (ceo @typesafeai).”[40]

The load-bearing correction: RLHF as a research program predates InstructGPT (Christiano et al. 2017; Stiennon et al. 2020 are cited in the InstructGPT paper itself).[34] Almeida is a primary author of the paper that productized that method for instruction-following GPT-3. “Co-invented ChatGPT” compresses a multi-author product launch into one name. The paper authorship is the fact that survives.

The same person who helped make chat models follow instructions is now arguing that those instructions are why the models still need a human in the loop. The 2026 talk title is the plot: “What’s next after RLHF?”[43]

Cofounders

❓ Who else is on the founding line, and what did they build before?

FinSMEs names the same three as the leadership line.[22]

Why announce from a personal account

❓ Does the launch channel change what is true about the model?

No. It changes the distribution. Doomers’ write-up of the 15 September 2026 thread notes the post went out from the founder, led with a verifiable credential, then a question, then two years of stealth, then three checkable claims.[37] That is launch craft. The model is still the API in 06-stack-and-api.

← Claims auditEvals & accuracy →