# almeida-system-one-models-jev-2026-09-15

## Veille

Announcement by **Diogo Almeida**, founder of **TypeSafe AI** (formerly of OpenAI, where he worked on instruction-following training methods), published on **September 15, 2026** on the company blog (~2,500 words). The text introduces a new class of models, **System One Models**, and their first model, **Jev**, in early access.

## Titre Article

Introducing System One Models & Jev

## Date

2026-09-15

## URL

https://typesafe.ai/blog/introducing-system-one-models-and-jev

## Keywords

System One, decision model, Jev, TypeSafe AI, RLCD, typed output, calibrated probabilities, parallel sampling, function call, workflow, workflow evaluation, hallucination, latency, cost per call, Kahneman, Jevons, classification, automation, Pareto

## Authors

Diogo Almeida — fondateur de TypeSafe AI, ancien chercheur chez OpenAI (billet du blog d'entreprise).

## Ton

A first-person launch announcement, openly enthusiastic ("beyond excited"), but built around evidence: an "Evidence" section separates verifiable claims (speed, cost, absence of type errors) from the bolder ones, and each demonstration is followed by a "Nuance" box listing its biases (simplified queries, evaluations written by the team, a reference favorable to OpenAI and Anthropic, OpenRouter figures). The humor of the demonstrations (Doom, Wikiracing) sits alongside a dry comparison table between LLMs and System One. The central argument is framed as a question: "Models have been superhuman at chat for years, so where is all the automation?"

## Pense-betes

- **Definition**: a System One Model produces **typed values** whose structure is fixed in advance; it therefore cannot make a type error. The name comes from Kahneman's distinction (fast System 1 / slow System 2); "Jev" comes from William Stanley Jevons (each drop in the cost of intelligence opens up new uses).
- **Stated comparison with LLMs**: optimization via RLHF/RLVR versus RLCD; character-string output versus typed values with probabilities; sequential versus parallel sampling; output billed at roughly 5× the input versus free output.
- **Targeted use cases**: "smart if-statements" (classifying, routing, scoring, extracting, branching), large-volume processing, real-time applications, verification (guardrails, jailbreak detection on LLM outputs). Excluded: conversation, coding agents, proofs.
- **Workflow evaluation**: a computation graph assumed correct, reference probabilities drawn from the largest models (average of GPT-6 Astra and Fable 5.1), the same workflow for all. Jev occupies the Pareto frontier across nearly two orders of magnitude; four workflows published.
- **Caveat raised in the article**: the absence of type errors is guaranteed by construction (a non-empirical "0%"); the sustainability of the pricing cannot be demonstrated at this stage.
- **Demos**: an agent playing Doom (about 10 requests per second, ~$7/hour) on structured state, not on images; Wikiracing, with a two-stage system beyond 255 choices.

## RésuméDe400mots

**Diogo Almeida**, founder of **TypeSafe AI**, announces on **September 15, 2026** the first **System One Model**, named **Jev**, available in early access. His starting question: given that models have been superhuman at conversation for years, where is the automation? His diagnosis is that LLMs, optimized for human preference (RLHF) or for verifiable rewards (RLVR), produce character strings; for software to use them, they must be parsed and validated, with the constant risk that they fall outside the expected format.

TypeSafe offers a stack entirely oriented toward automation: a new model architecture, a parallel sampler, and a training method, **RLCD** (reinforcement learning for calibrated decisions), which optimizes responses accompanied by honest probabilities. Jev takes unstructured data as input, preferably a program state, and returns **typed** values, whose shape is defined in advance, with probabilities and confidence scores. It forgoes text generation, which, according to the author, prevents it from hallucinating or making type errors. It describes itself as a frontier-intelligence function call.

The announced figures: input at $0.042 per million tokens, free output, latency of 70 to 500 ms versus 3 to 329 seconds for frontier models, a factor of 40 to 200 on System One-type queries. Sampling is parallel: all outputs are produced in a single request.

The article separates verifiable claims (speed, cost, absence of type errors) from the bolder ones. It presents a side-by-side demonstration with GPT-5.6 Terra, whose only disagreement concerns a starting risk level judged genuinely ambiguous, followed by **workflow evaluations**: a correct computation graph is assumed, and each model's probabilities are compared against the average of GPT-6 Astra and Fable 5.1. Jev dominates the Pareto frontier across nearly two orders of magnitude; the company's homepage cites **193.6× faster and 444.6× cheaper**, which the author places at the high end of the real gains. Each demonstration is followed by caveats: LLM figures sourced from OpenRouter, workflows written by the team, a reference favorable to OpenAI's and Anthropic's models.

Use cases: fuzzy decisions embedded in code (classifying, routing, scoring, extracting), bulk data processing, real-time applications, verification of LLM outputs. Two demonstrations, Doom and Wikiracing, illustrate the speed and the absence of hallucination in high-cardinality choices (up to 255 options).

The name refers to Kahneman (System 1) and to Jevons: each order-of-magnitude drop in the cost of intelligence would open up just as many new uses. TypeSafe is opening early access and asking developers which decisions they want to automate.

## GrapheDeConnaissance

- Diogo Almeida —travaille_chez→ TypeSafe AI (ORGANISATION, 0.96)
- Diogo Almeida —publie→ article Introducing System One Models & Jev (DOCUMENT, 0.97)
- TypeSafe AI —publie→ Jev (TECHNOLOGIE, 0.98)
- Jev —est_instance_de→ System One Models (CONCEPT, 0.97)
- System One Models —s_inspire_de→ Thinking, Fast and Slow (DOCUMENT, 0.9)
- Jev —utilise→ RLCD (METHODOLOGIE, 0.97)
- Jev —utilise→ échantillonnage parallèle (CONCEPT, 0.95)
- Jev —résout→ hallucination (CONCEPT, 0.85)
- Jev —permet→ décisions typées à probabilités calibrées (CONCEPT, 0.95)
- Jev —surpasse→ LLM de frontière (TECHNOLOGIE, 0.8)
- Jev —mesure→ 193,6× plus rapide et 444,6× moins cher sur les évaluations de workflow (MESURE, 0.9)
- Jev —mesure→ latence de 70 à 500 ms, 0,042 $/MTok en entrée (MESURE, 0.93)
- évaluation de workflow —est_basé_sur→ moyenne de GPT-6 Astra et Fable 5.1 (AFFIRMATION, 0.88)
- Jev —s_applique_à→ smart if-statements (CONCEPT, 0.92)
- Jev —concurrence→ LLM de frontière (TECHNOLOGIE, 0.8)

---
Canonical: https://www.thekb.eu/en/fiches/almeida-system-one-models-jev-2026-09-15/