Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI

What it claims

Jev is TypeSafe AI's first "System 1" / "large programmable model". Its outputs are meant to be consumed by code, not read by people. The name comes from Jevons Paradox, and the model family is positioned on the frontier of intelligence per dollar rather than raw capability. TypeSafe publishes no benchmark numbers. The only traction figure Almeida gives is that usage passed a trillion tokens per day within the first week. He says the traffic continues overnight, which he reads as machines calling the model in loops rather than people trying it out. The launch video was cited at ~36–40M views.

How it works

• RLCD is a training target, not an algorithm. It is unpublished, and the "no paper yet" status is confirmed on the pod. Almeida frames RLHF as "the task of instruction following" independent of PPO, in the same way DPO still counts as RLHF. RLCD is likewise a new north star: calibrated decisions with "programs in the loop."
• The calibration argument. He rebuts LeCun's "error compounds with sequence length" slide by pointing to mode collapse. RLHF-style models stay accurate over long strings because they become extremely conservative and drop minority modes, much as GANs did compared with blurry mode-covering models. That makes their token probabilities "total poison" for decision-making. Jev targets mode-covering, calibrated probabilities on discrete decisions instead.
• Data. TypeSafe calls itself a "data lab." All training data is synthetic and task-shaped. Staff look for the model's specific weaknesses ("jaggedness") and fix them so the fix generalises. Almeida says TypeSafe does not want to train on user data even if permitted, because real-world query power laws cause overfitting.

How to use it

The page's guidance, with the detailed API section falling in the truncated part:

• Decompose workflows into small, measurable decisions. Use structured state instead of giant prompts or system messages. Use queries where you would otherwise write non-trivial branching logic.
• Evaluate on your own workflow. Almeida is "extremely anti-public benchmarks" and "medium" on private proxy benchmarks. A terms-of-use clause from the preview period that appeared to restrict benchmarking is being removed.
• No refusals at the API layer, by policy. Almeida treats refusals as a "type error" for dependencies and rejects use-case vetting ("we are an API… it's none of my business").
• Use cases with official material:
• an official coding-agent guide, covering linting and compacting tool calls (with Theo's pushback and Almeida's "Tyranny of the KV Cache" note as counterpoint)
• speed-bound games and computer use, including the Doom demo
• dark-data pipelines
• Jev as a judge
• analytics replay
• Plan around rate limits, not signup access. Rate limits were the real bottleneck at launch.

Limits

• No seed and no determinism guarantee. The host notes the API has no seed parameter, and Almeida says reliability matters more than strict determinism. The full answer is cut off.
• Reliability is not yet at the target. By Almeida's own account, Jev is "not anywhere close" to database-query reliability and needs "so many more nines."
• No published method or benchmark evidence. RLCD is unpublished and no benchmarks are released, so the claims rest on vibes, demos and usage volume.
• Content controls are your responsibility. The no-refusal policy means any filtering has to happen in your own application.
• Only part of the page was reviewed. This covers the first ~60k of ~186k characters of the transcript. Sections on model versioning/LTS, the API primitives, build patterns, coding agents, safety and RLVR limits, and the multi-agent future are not reflected here.