Lesson 01 · Jev 101

What is Jev?

Jev is TypeSafe AI’s first public System One model—an AI built to return typed decisions software can use directly, not chat replies for people.

The problem (ChatGPT-native framing)

Models got great at chat. Teams then asked chat models to run products: classify tickets, route agents, score risk, judge other model outputs. That works in demos—and often fails in production because:

  • Outputs are strings you must parse and validate
  • Models can be overconfident or inconsistent when you need automation
  • Latency and cost hurt when decisions sit on a hot path

TypeSafe’s public thesis: chat optimization (RLHF) created inherent issues for reliable autonomy—so they built a different model class for decisions.

“Yet RLHF creates inherent issues such as mode dropping, overconfidence, and lack of reliability. These flaws mean that LLMs require humans-in-the-loop.”

TypeSafe AI homepage · fetched Sep 20, 2026 · Source

“We built a new class of models, System One Models, to be natively used by machines. We're building with a new architecture, a new sampler, and a new training algorithm: Reinforcement Learning for Calibrated Decisions (RLCD).”

TypeSafe AI homepage · fetched Sep 20, 2026 · Source

System One (without the textbook)

Inspired by Kahneman’s fast vs slow thinking as a naming metaphor: System One Models target fast, structured decisions inside software. Jev is the first public model in that class (company). Official launch FAQ: they believe System One Models can be more reliable than alternatives; they have promised a future explanation. This 101 does not treat Jev as “human System 1.”

“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”

Diogo Almeida / TypeSafe launch blog · Sep 15, 2026 · Source

What Jev is / isn’t

Jev isJev is not
A decision model for softwareA ChatGPT competitor for chatting
Choice / Score / Noul over your stateA writing or coding assistant
Probabilities + confidence for thresholdsA guarantee of perfect accuracy
Complementary to LLMs in workflows“Just a smaller LLM with JSON mode” (company rejects this framing)

Who makes it

TypeSafe AI — San Francisco frontier lab; emerged aroundSeptember 15, 2026. Public materials cite an in-person office near Embarcadero (five days a week). Founders on the official team page: Diogo Almeida (CEO; company/press describe InstructGPT / RLHF lineage), Erik Gafni (CTO), Sasha Sheng (COO). Confirm bios againsttypesafe.ai/teambefore publishing quotes.

Press (Business Wire): about $40M seed led by DCVC. Treat that as press, not a product metric.

“We're building prod, not God.”

TypeSafe manifesto · fetched Sep 20, 2026 · Source

“the bottleneck isn't raw intelligence. It's that today's intelligence is hard to build on.”

TypeSafe manifesto · fetched Sep 20, 2026 · Source

Names, on purpose

  • System One ← Kahneman, Thinking, Fast and Slow (official launch FAQ)
  • Jev ← William Stanley Jevons / Jevons paradox (official launch FAQ)
  • Not Cyera SafeType, and not “Safetype AI”

What this guide will not claim

  • Independent verification of speed / cost multiples
  • A free tier, open weights, or SLAs (not clearly established for marketing here)
  • That you should delete ChatGPT from your stack
  • Architecture rumors (transformer base, open-weight lineage) as fact

Coming from chat UIs? OptionalWhat’s the difference?. Next in the 101:How Jev works.