Fireship · 2026-09-21 · notable
Fireship — 'An ex-OpenAI researcher just deleted language from the LLM...'
Fireship's 21 September 2026 video covers Jev, the model from TypeSafe AI that returns a typed value with a confidence score instead of generated text. TypeSafe was founded by Diogo Almeida, a former OpenAI researcher.
Fireship walks through Jev, the TypeSafe AI model that answers with a typed decision and a confidence score rather than text.
What is it?
'An ex-OpenAI researcher just deleted language from the LLM...' went up on the Fireship channel on 21 September 2026. Its subject is Jev, which TypeSafe AI calls its first 'System One Model' — built for structured decisions inside software automation rather than for chat. TypeSafe AI was founded by Diogo Almeida, a former OpenAI researcher who worked on ChatGPT's instruction following.
How does it work?
TypeSafe sums Jev up as 'unstructured state in, typed probabilistic decisions out'. Rather than generating tokens one at a time, the model samples all outputs in parallel in a single query and returns a type-safe value carrying a confidence score. Because the answer must be a member of a fixed type rather than free text, TypeSafe claims Jev 'can't hallucinate'. Fireship compresses a story into a few minutes, which is often where a new model first reaches working developers.
Why does it matter?
Classification, routing, and scoring make up a large share of production LLM calls, and paying frontier prices plus seconds of waiting for a yes/no answer is wasteful. TypeSafe lists Jev at 70–500ms per response and $0.042 per million input tokens with output tokens free. Those are the company's own figures, and Fireship's video is commentary on them rather than independent testing.
Who is it for?
developers wiring LLM calls into automation