Jev (TypeSafe AI)
by TypeSafe AI (founder Diogo Almeida, ex-OpenAI)
“Intelligence beyond chat.” TypeSafe’s first System One Model — a frontier model that gives up free-text generation entirely to specialize in fast, typed, probabilistic decisions: unstructured state in, calibrated structured output out, with a 0% schema-error rate and 70–500ms end-to-end latency. Released 2026-09-15 in early access.
See https://typesafe.ai/blog/introducing-system-one-models-and-jev
System One Models — the category
Named after Daniel Kahneman’s System 1 (fast, intuitive judgment) vs. System 2 (slow,
deliberate reasoning). TypeSafe’s framing: existing LLMs are trained and optimized for
chat/conversation, but most real automation doesn’t need prose — it needs a fast, reliable
decision a program can act on directly. Jev is pitched as “similar levels of intelligence
on System One tasks compared to existing LLMs, while being two orders of magnitude faster
and more efficient” — a frontier-intelligence function call, not a chatbot.
RLCD — Reinforcement Learning for Calibrated Decisions
RLCD is the training method behind Jev, and the direct answer to “how is this different
from RLHF.” Where models like GPT-5.6 or Opus 5 are trained with RLHF to produce text
humans prefer, RLCD optimizes for calibrated probabilities on structured decisions —
the goal is that a stated 90% confidence is actually right about 90% of the time, not just
a plausible-sounding number.
- Uses verifiable ground-truth data and proper scoring rules (e.g. the Brier score) to
compute reward, instead of human preference ratings - Produces “epistemically honest” probabilities: higher stated confidence should
correlate with higher real-world accuracy — essential for a system that hands off to a
human or a slower model only when it’s actually uncertain - Trains the model to answer multiple structured questions in parallel against the
same state/context in a single forward pass, reading answers from hidden states at
specific token positions rather than generating text token-by-token
Output types: Choice / Score / Noul
Three structured decision types Jev is trained to emit:
| Type | Answers | Returns |
|---|---|---|
| Choice | Select one option from a provided set (up to 255) | chosen option, full probability distribution, confidence score |
| Score | A position on an ordered scale | expected score = Σ p_i · i (can fall between discrete levels), probabilities, confidence |
| Noul | Binary yes/no | answer + P(yes) probability, no separate confidence metric |
Performance claims
- Latency: 70–500ms end-to-end, reported as 40x–200x faster than frontier LLMs on
comparable tasks - Cost: priced per input token, output tokens free (“too cheap to meter”),
far below typical frontier LLM API rates - Reliability: 0% structured-output/schema error rate (schema-guaranteed by
construction, not prompted) - Workflow-level claims: up to 193.6x faster and 444.6x cheaper on production-style
automation workflows in TypeSafe’s own (higher-end) benchmarks
Use cases
Conditional logic inside AI workflows, real-time/sub-second decision paths, map-reduce
operations across large datasets, verifying/scoring/guardrailing the output of other
(text-generating) models, and high-cardinality classification/routing decisions — the
class of problem where a chat-tuned LLM is both slower and less calibrated than needed.
Trade-off
Jev cannot generate free text — it is not a chatbot replacement, and the entire pitch
depends on pairing it with a conventional LLM for anything requiring prose, explanation,
or open-ended generation. It’s positioned as the fast, cheap decision layer in front of or
alongside a slower model, not a general-purpose substitute for one.
Company
TypeSafe AI, founded 2024 by CEO Diogo Almeida
(previously at Google Brain and OpenAI, co-inventor of RLHF/InstructGPT and a GPT-4/
ChatGPT contributor) with co-founders Erik Gafni and Sasha Sheng. Exited two years of
stealth development on 2026-09-15 with a 200M
valuation, launching Jev the same day. Company framing: “Models have been superhuman at
chat for years, so where is all the automation?”