FULL STORY

Jev: The Text-Free Decision Model That Went Viral

Ex-OpenAI researcher Diogo Almeida's stealth startup TypeSafe AI launched Jev, a non-generative "System One" decision model claiming 200x speed. It went fully open, sparked heated debate over its positioning, and saw record-fast adoption in the agent community within days.

2026-09-20 ~ 2026-09-25 · 11 episodes · 84 posts

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)

Around September 20, ConvAI Innovations founder Nandakishor Mukkunnoth publicly disputed the "breakthrough" framing of Jev, a non-autoregressive fast decision engine released in September by frontier company TypeSafe AI (founded by ChatGPT co-inventor Diogo Almeida), claiming he had built and open-sourced similar work a year earlier. Several bloggers reshared and debated the claim.

Confirmed

  • According to Mukkunnoth, he published an arXiv paper on non-autoregressive decision models in March 2025 (arXiv:2503.23303), followed by a second paper (arXiv:2510.01237) on RL-guided related methods
  • He open-sourced model weights and a SaaS sales dialogue dataset on Hugging Face, and released a PyPI package for reproduction
  • He argues Jev is the same concept but captured attention and the "breakthrough" label due to the frontier lab's launch

Unconfirmed

  • The actual technical similarity between Jev and Mukkunnoth's model rests on his account alone; TypeSafe AI has not publicly responded
  • The posts do not give Jev's exact release date or the original promotional text

Why it matters

  • The episode renews debate over attribution and visibility asymmetries between frontier labs and independent researchers: the same concept draws little attention when published independently, yet is hailed as a breakthrough when a famous lab ships it
  • With papers, weights, dataset and PyPI package open-sourced, outsiders can independently verify the comparison

Episode 2 · ianand's Thought Experiment: An AI World Built on the Encoder (2026-09-20, 9 posts)

On September 20, ianand posted a long thread imagining Jev, an AI model from a parallel universe, using the thought experiment to revisit the encoder/decoder split of the Transformer and explore an alternative AI product form. His bottom line: Jev is not a ChatGPT replacement but another tool in the developer's toolbox, and the whole construct is fiction, not a real product or research.

Confirmed

  • Historical background: the 2017 Transformer originally did one thing — translate English to French — consisting of an Encoder (analyzing text) and a Decoder (producing output), the origin of the路线 split.
  • Key milestones in our universe: GPT-2 took only the decoder half and proved "all you really need" is four things — the decoder itself, one model handling many tasks, prompting replacing task fine-tuning (which ianand calls the biggest unlock), and scale. These breakthroughs led to today's ChatGPT, Claude and Gemini.
  • Setup of the experiment: in Jev's parallel universe there was no GPT-2 moment; all technical and product innovation went to the Encoder side, producing Jev — faster, more predictable, more reliable, but not generative: it cannot write code or chat and can only make decisions.
  • ianand describes Jev as giving the Encoder a ChatGPT-grade product form, complementary rather than a strict substitute.

Unconfirmed

  • ianand himself concedes we don't know whether Jev would truly be an Encoder underneath, and that decoder-style self-supervised training scales more easily — perhaps humanity was destined to build ChatGPT first. The entire Jev premise is speculative fiction.

Why it matters

  • The thread offers a lens on today's AI landscape: conversational generative AI is not the only possible path but the outcome of the decoder route plus prompting and scaling. Counterfactual reasoning helps separate essential breakthroughs from path dependence.

Episode 3 · Ex-OpenAI Founder Says RLHF Is a Detour, Human-AI Collaboration Era Ending (2026-09-20, 3 posts)

Jev founder and former OpenAI employee Diogo Almeida argues that RLHF-trained models are stuck in a human-in-the-loop "assistance era" that is ending. He claims Jev's automation-first approach is 200x faster and 400x cheaper.

Episode 4 · Ex-OpenAI researcher launches decision model Jev with bold speed claims (2026-09-21, 36 posts)

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'.

16 more related posts →

Episode 5 · Jev: A Decision-Only AI Model Returns Probabilistic Answers in 200ms (2026-09-21, 2 posts)

Jev is a new decision-only AI model that skips text generation, returning typed, probability-based answers to predefined questions in about 200 milliseconds at minimal cost.

Episode 6 · Decision Model Jev Opens to All, Developers Compare Buzz to GPT-3.5 Launch Week (2026-09-21, 2 posts)

TypeSafe AI's Jev, a text-free "System 1" semantic decision engine that outputs typed options with confidence, opened to everyone after removing its waitlist. Developers report GPT-3.5-launch-week levels of buzz, with claims of 200x speed and 400x cost savings.

Episode 7 · Jev model shines in tool use, lags in reasoning and retrieval (2026-09-22, 2 posts)

Benchmark results show the Jev model excels at tool calling, automation, and graduate-level knowledge, though its reasoning and retrieval remain weak spots.

Episode 8 · Jev Model Sparks Debate: A Dedicated Decision Layer Beyond LLMs (2026-09-23, 4 posts)

The community debates Jev, a model that outputs only structured judgments, positioning it as a missing decision layer between deterministic code and LLMs, while users clarify misconceptions such as its non-determinism despite type safety.

Episode 9 · Decision-Only Model Jev Goes Viral as a Cheaper DeepSeek-Style Threat (2026-09-23, 3 posts)

TypeSafeAI's Jev, an OpenAI-alumni project, outputs only structured decisions with calibrated probabilities rather than text. Commentators argue its cheap, API-ready decisions make it more disruptive than DeepSeek.

Episode 10 · Ex-OpenAI researcher launches Jev, a non-text "System One" decision model (2026-09-23, 12 posts)

TypeSafe AI, co-founded by ex-OpenAI researcher and InstructGPT co-author Diogo Almeida, emerged from two years of stealth to launch Jev, a proprietary "System One" model (early access) with a $40M seed round led by DCVC. Unlike LLMs, Jev generates no text: given a state and typed question, it returns schema-constrained answers—choices, scores, or typed numeric values with probability estimates and confidence scores—for direct consumption by other software.

Confirmed

  • Response time of 70–500 ms; $0.042 per million input tokens, output free. Per Kuailiyu, up to 193x faster than large models, with per-decision cost around $0.0004.
  • Founder Diogo Almeida claims he helped invent ChatGPT (AxSaucedo calls him co-inventor of RLHF) and spent two years researching an alternative training method.
  • A Reddit explainer relayed by techtechnical says Jev uses RLCD (reinforcement learning from critique) for calibrated decisions; the "System One" name borrows from fast/slow thinking in psychology.
  • Per Kuailiyu, 13% of teams paid within 24 hours of launch.
  • One user (relayed by PrajwalTomar) tested Jev by hiding read counts of 468 tweets and asking it to predict them.
  • Dhravya Shah (relayed by julianweisser and blaizedsouza) argues Jev is reshaping AI memory and context engineering, with benchmarks and use cases in computer use, browser use, and data labeling.

Unconfirmed

  • The "193x faster" and "13% paying teams in 24 hours" figures come from secondary sources (Kuailiyu); the comparison baseline and methodology are not officially disclosed, and RLCD details come from community posts.

Why it matters

  • Jev is positioned as a new primitive beyond LLMs: it cannot write essays, explain itself, or hallucinate much, making it suited for massive volumes of fast, cheap, calibrated bounded decisions inside agent runs (hardimanjames notes agents spend most time on small decisions, not "reasoning"), rather than conversation.
  • Authors like danielmac8 argue its speed and pricing make it a judgment component that other systems can call at scale—an inverse exploration of the text-centric AI paradigm.

Episode 11 · TypeSafe's Jev Decision Model Sees Fastest Adoption, Sparks Open-Source Agent Ecosystem (2026-09-24, 5 posts)

TypeSafe AI's Jev decision model saw massive adoption within three days of release, being called the fastest-adopted model in AI history, and quickly spawned a set of very low-cost open-source agent projects along with teardowns and practice writeups from developers and veteran engineers.

Confirmed

  • Jev is positioned as a "System One" decision model: it takes unstructured state as input and outputs typed decisions with probabilities, generates no text, and runs at millisecond-level inference. This comes from Victor Dibia's (@vykthur) long-form teardown, based on official materials and hands-on experiments including a personal reimplementation with Qwen.
  • Within three days of release the community shipped multiple open-source projects, such as jev-ultrafast, a browser agent using a "dynamic indexed action space" design where Jev itself picks operations and elements.
  • @TrainerIntelligent shared their practice of embedding Jev into the open-source agent framework JarvisCore, landing it within one week in four places—RAG classification, subagent routing, model routing, and agent decisioning—and foreseeing more embedding scenarios like memory assembly, tool recovery, and human-in-the-loop.
  • An engineer with 20 years of ML experience (@NoFriendship4813) spent a week mapping Jev and its open-source ecosystem (Kev, SemIf, Laya, Nimble, OpenJev, decider, Winnow, etc.), focusing on call-count problems in agents and arguing that agent architecture is moving toward heterogeneous inference.
  • Arindam (@Arindam1729) cautioned developers not to use Jev as a general LLM (e.g., writing emails or long answers); the correct usage is to let the model make decisions—provide context, define the decision set, and let application code execute on the result.

Why it matters

  • Jev demonstrates a new division of labor: "small decision model + large generative model." Millisecond-level, low-cost decision calls can be embedded throughout agent frameworks instead of relying on general LLMs for text generation.
  • The speed from release to ecosystem formation in three days, plus consistent hands-on feedback from independent developers, suggests strong real-world appeal and that heterogeneous inference may be the next direction for agent architecture.