An ex-OpenAI researcher just deleted language from the LLM...

Summary

Jeff Delaney presents an episode of The Code Report on Fireship discussing the release of Jev, a “System One” frontier decision model by TypeSafe AI founded by ex-OpenAI researcher Diogo Almeida. The video covers how Jev replaces conversational output with typed probabilities and structured decisions, contrasting it with traditional reasoning LLMs and examining community reactions and open-source alternatives.

What is shown

• [00:15] Blog announcement of “Introducing System One Models & Jev” from TypeSafe AI by founder Diogo Almeida.
• [00:34] Comparison of a traditional reasoning model generating extensive chain-of-thought tokens for a boolean question versus Jev’s output format.
• [00:48] Comparative table of LLMs versus System One models across speed, cost, and output tokens.
• [01:34] TypeSafe AI Playground UI demonstrating input state and strongly typed schema queries (isHorse, primitive types like Noul, Score, Choice).
• [02:07] Side-by-side terminal comparison benchmark running 27 queries on TypeSafe API (System One) completed in 0.114s vs. autoregressive OpenAI API (GPT-5.6 Terra) waiting for output.
• [02:20] Demonstration of a hypothetical app “Horse Tinder” using Jev for automated moderation of donkey accounts.
• [02:50] X (Twitter) community clips showing Jev applied to Flappy Bird gameplay and an AI-first calculator.
• [04:07] GitHub repository and architecture diagrams of open-source reproductions, including “SemIf” (formerly OpenJev) running on a local GPU using Qwen 4B logits in a single forward pass.
• [04:25] Sponsored demo of Mux API, dashboard, and AI video directives.

Claims & numbers

• TypeSafe AI raised $40 million in funding (presenter statement) [01:29].
• End-to-end response time for Jev ranges from 70ms to 500ms, claimed to be 40x to 200x faster than traditional frontier models (shown in TypeSafe announcement chart) [00:50].
• Input token cost is stated as $0.042 per million tokens ($42 per billion tokens), while output tokens are free, compared to standard LLM costs of $0.20 to $10 per million tokens [00:52, 02:46].
• Claims 444x cheaper pricing than top-tier models and zero hallucinations due to returning typed probability distributions rather than generated text [00:54, 02:48].
• Within 24 hours of launch on AI Gateway, Jev reached nearly 13% share of paid teams, exceeding GPT-5.6 and Claude Fable 5.1 adoption rates [00:06].
• Uses a technique called RLCD (Reinforcement Learning for Calibrated Decisions) to produce calibrated confidence scores [01:00, 03:28].
• Community open-source project SemIf reproduces the mechanism via a single forward pass reading logits of declared options from an unmodified Qwen 4B model on a consumer RTX 3090 [04:12].

Notable quotes

• “It deleted language from the large language model.” — [00:46]
• “A System One model is fast and goes from gut instinct, while a System Two model is slow and deliberate...” — [02:07]
• “Type safety is not factual correctness, and it’s not even deterministic.” — [03:04]

Assessment

This is a comedic commentary and news review video by Fireship analyzing the product release of Jev, complete with third-party social media clips and playground demos. While the playground and terminal races illustrate functional API mechanics, the benchmarks cited originate directly from TypeSafe AI’s promotional announcements and trust-me-bro self-reported data.

Described by gemini-3.8-flash on 2026-10-09 from the video’s audio and frames.