A ChatGPT Co-Inventor Is Betting 40 Million That AI's Future Isn't Conversation

A co-inventor of the training method behind ChatGPT says the technology he helped build has hit a ceiling, and he is betting roughly 40 million dollars that the fix is to stop teaching machines to talk. Diogo Almeida, who co-invented RLHF and InstructGPT at OpenAI, came out of stealth on September 15, 2026 as chief executive of TypeSafe AI, a San Francisco lab whose first model does not chat, does not write, and is not designed to. The company says it raised about 40 million dollars in a seed round led by DCVC. Dealroom lists 25.9 million dollars as the seed portion, so the larger figure is company-reported.

A Co-Inventor Breaks With the Road He Paved

Almeida's credentials are not in dispute. He co-invented RLHF, or reinforcement learning from human feedback, and InstructGPT, the training methods that led to ChatGPT and GPT-4. He previously worked at Google Brain. Few people are better positioned to argue that the approach has limits, and fewer still have staked a company on that argument.

His public thesis is blunt. RLHF, he argues, structurally limits artificial intelligence to being an assistant. It cannot get you to automation. The reason is confidence: you cannot let a system act on its own if it will not tell you how sure it is. A model that always answers in fluent prose has no reliable way to say "I do not know."

That is the intellectual core of TypeSafe. Almeida contends calibrated confidence is the unlock, not better conversation. The company's name signals the claim: types, in the programming sense, as a guarantee. Its first product is called Jev, named after Jevons paradox, the economic observation that when something becomes more efficient we consume more of it, not less.

What Jev Actually Does, and What It Refuses To Do

Jev is what TypeSafe calls a "System One" model, a phrase the company uses to name a new class of model rather than a chatbot. It does not chat and does not write. The company calls it a decision primitive. You send it a piece of context plus typed questions with defined criteria, and it returns typed answers with a calibrated confidence score.

Because the range of questions is constrained, TypeSafe argues the model cannot invent facts the way a text generator can. The company says it returns decisions in under 100 milliseconds. That latency claim matters commercially: a system that answers in a tenth of a second can sit inside a workflow, not beside it.

The training method is called RLCD, Reinforcement Learning for Calibrated Decisions, which TypeSafe describes as a new architecture and a new sampler. The company claims type errors are mathematically impossible and has invited outsiders to falsify that. It also cites performance figures such as 193.6 times faster and 444.6 times cheaper than comparison models, while explicitly describing those as being on the higher end of real-world gains.

The Team Behind the Bet

Almeida is not running this alone. CTO Erik Gafni is a repeat founder who built Ravel, a multimodal AI company for DNA sequencing, and was an early employee at two unicorns, Invitae and Freenome. That is a diagnostics and infrastructure background, not a consumer chat background.

COO Sasha Sheng is a former research engineer at Meta and FAIR who worked on News Feed and AI Experiences, and has published at NeurIPS and ECCV. The three-person leadership mix, research pedigree plus sequencing infrastructure plus large-scale product systems, reads as a deliberate bet on deployment rather than demos.

The company was founded in 2024 and spent roughly two years quiet before the September 15, 2026 announcement. DCVC led the round. TypeSafe states about 40 million dollars total; Dealroom's 25.9 million dollar seed figure suggests the two numbers measure different things, and the gap is worth noting for anyone tracking how AI seed rounds are described.

The Price, and the Admission Attached To It

TypeSafe's pricing is aggressive. Input tokens cost 0.042 dollars per million, which the company describes as 42 dollars per billion. Output tokens are free. On a decision carrying roughly ten thousand tokens of context, that works out to about 0.0004 dollars per decision. One million decisions would cost roughly 400 dollars.

Then comes the line that separates this launch from most. Their own blog says: "We can't prove it isn't subsidized; we'll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up)."

That is an unusual admission from a funded AI lab. It concedes the economics are unproven and that the current price may not reflect true cost. It also commits to a direction, cheaper rather than more expensive, which is a promise customers and competitors can hold them to.

What This Means for the AI Industry and Automation

The industry has spent three years optimizing conversation. TypeSafe is arguing that conversation is the wrong interface for the work that pays. If Almeida is right, the next wave of AI value is not in assistants that draft emails but in systems that make small, typed, auditable decisions at volume, inside claims processing, logistics, compliance, and diagnostics.

The stakes for ordinary workers are concrete. An assistant that suggests is a tool a person supervises. A decision primitive that returns a calibrated confidence score is a component a business can wire into a process and let run. That is the difference between augmenting a job and replacing a step in it, and it is why the confidence score, not the fluency, is the politically and economically loaded part.

For businesses, the appeal is cost and auditability. Four hundred dollars per million decisions is a number a finance team can model. A confidence score is a number a risk team can threshold. Neither is available from a chat model that answers everything with equal fluency and no stated uncertainty.

What Is Verified, and What Is a Company Claim

Verifiable: Almeida's role in RLHF and InstructGPT, his prior work at Google Brain, the September 15, 2026 stealth exit, DCVC leading the round, the leadership backgrounds at Invitae, Freenome, Meta and FAIR, and the published token price of 0.042 dollars per million input tokens with free output.

Company claims: the 193.6x and 444.6x performance figures, the sub-100-millisecond latency, the assertion that type errors are mathematically impossible, and the roughly 40 million dollar raise, which conflicts with Dealroom's 25.9 million dollar seed figure. TypeSafe itself flags the performance numbers as the high end of real-world gains, which is a credit to them.

Credit is also due on the subsidy question. Most labs at this stage do not volunteer that their pricing may not be sustainable. TypeSafe did, in writing, on launch day. That does not make the claims true. It makes them falsifiable, which is more than the sector usually offers.

Availability, Limits, and the Open Question

The product is officially in early access and is waitlisted. It is hosted only, with no public model weights, meaning it cannot be run on a customer's own systems. For regulated industries that cannot send data to a third party, that is a real constraint, not a footnote.

So the honest summary is this: a man who helped teach AI to talk is now selling AI that refuses to. The argument is coherent, the credentials are real, the pricing is startling, and the company has admitted it cannot yet prove the pricing works. The model is not generally available.

What to watch is narrow and testable. Whether independent users reproduce the latency and cost claims. Whether the calibrated confidence scores hold up when someone tries to falsify them. And whether the price goes down, as promised, or up, as the subsidy admission quietly allows. Almeida has made his bet public. The long term will grade it.

This article was produced with AI-assisted research and editorial support. Reporting is based on the source material cited below. Sources: TypeSafe AI (typesafe.ai); DCVC; Dealroom; Global1.News (22 September 2026).

By Jessica Ali, Staff Writer