OpenAI RLHF co-inventor’s TypeSafe AI exits stealth with $40 million for machine-native models

The researcher who helped teach chatbots to please people is now raising money for models that prefer not to chat at all. TypeSafe AI, a San Francisco frontier lab founded by former OpenAI researcher Diogo Almeida with Erik Gafni and Sasha Sheng, emerged from stealth with about $40 million in seed funding led by DCVC, according to a 15 September 2026 Business Wire release.

Machine-native intelligence instead of another chatbot

TypeSafe’s claim is that most useful intelligence should live inside software, not in a dialog box. Almeida, described in the company release as a co-inventor of RLHF and ChatGPT, says he spent years making models better at talking to humans and then concluded that human conversation cannot be the only interface for intelligence. TypeSafe is building what it calls machine-native, composable AI: models designed as primitives developers can wire directly into systems that need semantic judgment and decision-making.

That framing is the odd hook. While the industry races to make agents write emails and argue in natural language, TypeSafe is selling the opposite specialty. Its models are meant to classify, route, score, and decide with typed outputs software can consume immediately, rather than generate free-form text that another layer must parse and distrust.

Jev, System One, and RLCD

The first public model is Jev, named with a nod to Jevons Paradox and currently offered in early access with a waitlist at typesafe.ai. Business Wire says Jev aims for frontier-level intelligence at less than 100 milliseconds of latency and claims it can be up to 100 times faster and less expensive than other frontier models on the company’s comparison. The model can process hundreds of outputs in parallel from a single prompt and attaches calibrated confidence scores so software can decide when to act alone and when to defer.

FinSMEs’ 16 September summary adds product vocabulary that matches TypeSafe’s own launch materials. Jev is described as a System One model built with Reinforcement Learning for Calibrated Decisions, or RLCD, delivering typed outputs with calibrated confidence at low latency. Where RLHF optimized for human preference in chat, RLCD is pitched as optimizing for decisions whose stated probabilities match real-world accuracy. Founded in 2024 and headquartered in San Francisco, TypeSafe presents Jev as the first of a class rather than a one-off demo.

Company materials also emphasize what Jev will not do. It is not positioned as a replacement for GPT-style generators. It is positioned as a fast decision layer for tool choice, risk scoring, routing, and guardrails, the boring switches that make autonomous software either trustworthy or reckless.

What $40 million is buying

A DCVC-led seed of that size is large for a stealth exit and signals that investors are willing to fund an architecture bet, not only another fine-tune of an existing chat stack. Almeida’s OpenAI pedigree gives the story oxygen. The harder test is whether calibrated, typed decisions at machine latency become a standard software primitive or remain a niche API beside generative models that already try to do everything.

For now the public facts are clean enough. Roughly $40 million seed, DCVC leading, three founders including an RLHF co-inventor, a 2024 founding date, and a waitlisted model that would rather return a confident choice than write a paragraph. If TypeSafe is right, the next wave of AI value may look less like conversation and more like a very smart switch that knows when it is unsure.