TypeSafe AI exits stealth with $40M and a Doom-playing decision model

TypeSafe AI emerged from stealth with $40 million in seed funding and a first model that refuses to chat. The San Francisco startup, founded by former OpenAI researcher Diogo Almeida, released Jev, a system built to return typed probabilistic decisions for other software rather than natural-language answers for people.

Almeida is one of the co-inventors of reinforcement learning from human feedback and early ChatGPT work. DCVC led the round. Co-founders include Erik Gafni and Sasha Sheng. The company says it was founded in 2024 to pursue what it calls machine-native AI.

What Jev actually returns

Jev does not write paragraphs. Developers feed it a stThe Register describes a customer-service example that might return department weights such as billing 0.08, technical 0.85, and sales 0.07, with a confidence score. That output is awkward for a human asking a question. It is useful for a workflow that must route tickets, verify tool calls, or gate automation.

TypeSafe’s public demo shows Jev playing Doom when fed structured game-state data. The stunt is a proof of constrained control loops, not a product pitch for gamers.

Why the company says chat models are the wrong tool

Almeida argues that models optimized for people cannot be the only consumers of intelligence if AI is going to change how work runs. In a statement quoted by The Register, he said TypeSafe is building for intelligence that lives inside software.

The company trains Jev with an architecture it calls Reinforcement Learning for Calibrated Decisions, or RLCD, and brands the result a System One model. Unlike token-by-token chat models, TypeSafe says the architecture returns all outputs for a query at once through parallel processing.

ate value, such as a JSON object or a short string like a customer complaint, then ask through primitives that TypeSafe names Choice, Score, and Noul. The model answers with structured values and probabilities that another program can consume without parsing free text.