Ex-OpenAI researcher launches decision model Jev with bold speed claims

TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida after two years in stealth, launched the 'System One' model Jev on Sept 15 alongside a $40M round led by DCVC (valuation $200M). Jev does no chatting and generates no text: it takes unstructured state plus typed questions and returns structured answers with probability distributions and confidence in a single parallel call, claiming 20-200x speed, 40-400x cost savings, and no hallucinations versus LLMs. It is currently free, sparked widespread demos within days, but key claims await independent verification.

Confirmed

• Jev offers three primitives: Noul (yes/no probability 0-1), Choice (distribution plus confidence over up to 255 options), and Score (rubric-based scoring), replacing the pattern of asking a chat model and parsing prose back into branchable code.
• Almeida worked on instruction-following/RLHF in the InstructGPT/ChatGPT era; days after launch he appeared on the Latent Space podcast arguing the next AI wave is 'System One Mode', and Fireship published a popular explainer video.
• Demonstrations include a browser agent completing tasks for one-tenth of a cent, a trading bot making buy/sell calls every 300ms, sorting 1,700 emails for 18 cents, Higgsfield selecting ad creative, plus users running Doom, Minecraft, and LLM judging — 10 demos catalogued by svenai and FinanceYF5.
• Simon Willison prefers Maggie Appleton's term 'decision models'; community pieces like 'Jev wasn't built to make Agents' clarify it is not designed for general agents, and posts note it can route between Haiku and stronger models to pick the cheapest capable option.
• Google researcher Frank Dellaert explored combining Jev with graphical models, suggesting its probability distributions could shift the paradigm of reasoning under uncertainty.

Unconfirmed

• The 20-200x speed, 40-400x cost, and 'never hallucinates' claims come from the company; a critique relayed by prakersh challenges the no-hallucination and calibration claims, and no independent verification consensus exists.
• Multiple posts call Almeida a 'ChatGPT co-inventor/creator' — a self-description not independently verified.

Why it matters

• Writers including Akshay argue the industry treats LLMs as a hammer even for trivial decisions; Jev represents a new model category built for machine decisions rather than human conversation. If its claims hold, it could dramatically cut cost and latency for high-frequency decisions in agents and software — pkiv called it 'one of the biggest unlocks for fast, cheap browser agents'.

2026-09-21 ~ 2026-09-23 · 36 related posts

Full story(11 episodes)→

• Episode 1: Independent researcher says Jev is no breakthrough: he open-sourced a similar non-autoregressive decision model a year earlier(2026-09-20, 6 posts)
• Episode 2: ianand's Thought Experiment: An AI World Built on the Encoder(2026-09-20, 9 posts)
• Episode 3: Ex-OpenAI Founder Says RLHF Is a Detour, Human-AI Collaboration Era Ending(2026-09-20, 3 posts)
• Episode 4: Ex-OpenAI researcher launches decision model Jev with bold speed claims(2026-09-21, 36 posts)
• Episode 5: Jev: A Decision-Only AI Model Returns Probabilistic Answers in 200ms(2026-09-21, 2 posts)
• Episode 6: Decision Model Jev Opens to All, Developers Compare Buzz to GPT-3.5 Launch Week(2026-09-21, 2 posts)
• Episode 7: Jev model shines in tool use, lags in reasoning and retrieval(2026-09-22, 2 posts)
• Episode 8: Jev Model Sparks Debate: A Dedicated Decision Layer Beyond LLMs(2026-09-23, 4 posts)
• Episode 9: Decision-Only Model Jev Goes Viral as a Cheaper DeepSeek-Style Threat(2026-09-23, 3 posts)
• Episode 10: Ex-OpenAI researcher launches Jev, a non-text "System One" decision model(2026-09-23, 12 posts)
• Episode 11: TypeSafe's Jev Decision Model Sees Fastest Adoption, Sparks Open-Source Agent Ecosystem(2026-09-24, 5 posts)

• Jev, built by an RLHF/InstructGPT veteran, targets machine decisions with new RLCD training — MaryamMiradi · 2026-09-21
• ChatGPT co-inventor launches Jev, claiming 20-200x speed and 40-400x cost gains — BLUECOW009 · 2026-09-21
• ChatGPT co-inventor launches Jev, a decision-optimized model claiming 20-200x speed and 40-400x cost gains — soumitrashukla9 · 2026-09-21
• ChatGPT co-creator launches Jev, a free AI that never writes a word—only decides — PrajwalTomar_ · 2026-09-21
• Jev + graphical models: zero-shot probabilistic factors could reshape reasoning systems — fdellaert · 2026-09-21
• Builders weigh Jev structured-decision model against graphical models, flag d-separation hurdle — jatingargiitk · 2026-09-21
• TypeSafe's Jev: a model that only makes typed decisions and never writes prose — colinmcnamara · 2026-09-21
• "ChatGPT co-inventor" launches Jev, claiming 20-200x faster and 40-400x cheaper frontier model — altryne · 2026-09-21
• Jev, from a ChatGPT co-creator, makes AI decisions 200x faster and 400x cheaper — rubenhassid · 2026-09-21
• Judgment-only model Jev opens to all with $5 free credit, sparking calibration debates — 机器之心 · 2026-09-21
• Jev guide: the System-1 model that only judges and never writes — xiaohu · 2026-09-21
• ChatGPT co-inventor launches decision-making AI, claims 20-200x faster decisions — FinanceYF5 · 2026-09-21
• [source] TypeSafe's Jev returns typed decisions with probabilities, not text — here's where it fits in agent loops — prakersh · 2026-09-21
• [source] TypeSafe exits stealth with $40M at $200M valuation, but Jev's calibration claims ship with zero public evidence — prakersh · 2026-09-21
• Making Jev the decision brain of your business: browser agents for a tenth of a cent, 1,700 emails for $0.18 — PrajwalTomar_ · 2026-09-21
• Jev isn't a GPT rival: at $0.042/M input tokens, it wins at high-frequency judgment calls — FuSheng_0306 · 2026-09-21
• Jev explained: millisecond decisions at a fraction of LLM cost — multiply_matrix · 2026-09-21
• TypeSafe AI launches Jev, a "System One" decision model 100x faster than LLMs, no hallucinations — rseroter · 2026-09-21
• 'Belief propagation is back': can typesafeai's Jev specify graphical models for reasoning under uncertainty? — abeirami · 2026-09-22
• Ex-OpenAI researcher's Jev: a System 1 model claiming 200x speed and no hallucinations — Fireship · 2026-09-22
• TypesafeAI's Jev: a new model class built for fast, structured software decisions — Hacubu · 2026-09-22
• Jev-style models called a major unlock for fast, cheap browser agents — multiply_matrix · 2026-09-22
• TypeSafe CEO on Jev: machine-native models for code, RLCD over RLHF — Latent Space · 2026-09-22
• Jev: a "System One" model that skips text entirely, returning odds on your options in 0.45s — frappuccinoCoin · 2026-09-22
• TypeSafe launches Jev, a judgment-only model 200x faster and 400x cheaper, to rethink RLHF — oran_ge · 2026-09-22
• [source] TypeSafe AI launches JEV, a decision model hitting 70-500ms at $0.042/M input tokens — sven_ai · 2026-09-22
• TypeSafe AI's Jev introduces 'decision models': text in, probabilistic scores out, at $0.042/M tokens — teropa · 2026-09-22
• Jev from a ChatGPT co-creator auto-routes tasks to the cheapest capable model — fsharpman · 2026-09-22
• Jev, the model that never writes a word: 13% of paid teams onboard in first day, 70ms latency — 创业邦 · 2026-09-22
• Jev by TypeSafe AI skips token generation entirely, scores decisions in one parallel query — js_craft_hq · 2026-09-22
• Jev explained: millisecond decisions at a fraction of LLM cost for simple tasks — msharmas · 2026-09-22
• Jev, from a ChatGPT researcher, claims 200x faster and 400x cheaper decisions than LLMs — bibryam · 2026-09-22
• Typesafe's Jev model returns probabilities for structured decisions, bias concerns raised — fdellaert · 2026-09-23
• Jev swaps text generation for instant calibrated decisions, plays Doom at 7 req/s — JeremyCMorgan · 2026-09-23
• TypeSafe's Jev targets AI agents' high-frequency micro-decisions with low cost — WolframRvnwlf · 2026-09-23

1 near-duplicate retellings: zmkzmkz