Jev: The Schema-Safe AI That Could Change Automation Forever!

節目時間軸

1. 0:00

   Typesafe AI launches Jev, a non-LLM schema-safe model that abandons text generation for typed outputs.

   • Jev is not an LLM and Typesafe rejects calling it a distilled or smaller language model, categorizing it instead as a 'system one' model inspired by Kahneman's fast-intuitive versus slow-deliberate distinction.
   • Instead of predicting the next token, Jev takes unstructured state like raw text, code traces, or program logs plus a strictly predefined output schema and returns typed values, discrete decision paths, and calibrated probability scores in a single parallel pass.
   • Latency sits between 70 and 500 milliseconds end-to-end, but the trade-off is stark: Jev cannot write essays, code, or arbitrary strings, and direct choice cardinality caps at 255 options.
   • Typesafe claims schema violations are mathematically impossible because the model samples in parallel directly into allowed types, reporting a 0% type error rate in internal telemetry.
2. 1:36

   Jev's calibrated probabilities come from a proprietary training method called RLCD, but the architecture and weights remain secret.

   • Typesafe built reinforcement learning for calibrated decisions (RLCD) so the network's stated confidence matches actual statistical accuracy, arguing that RLVR on unverified business logic makes models brittle and overconfident.
   • Unlike RLHF which trains for human conversational preference or RLVR which optimizes against deterministic verifiers like math proofs, RLCD targets fuzzy business logic such as moderation, routing, and risk scoring that lacks cheap programmatic verification.
   • Jev cannot hallucinate an invalid schema or illegal enum, but it can still make incorrect decisions, which is why it outputs a calibrated probability alongside every decision.
   • Typesafe keeps the RLCD loss formulation, network architecture, parameter scale, and compute budget completely secret, with training data described only as proprietary and synthetic.
3. 3:41

   Aggressive pricing and performance claims rest on vendor-designed evaluations run under favorable conditions.

   • Published pricing lists input tokens at 4 cents per million ($42 per billion) and output tokens as free, since Typesafe claims the parallel sampler makes decision extraction too cheap to meter.
   • In highlighted workflow evaluations, Jev reportedly ran 193.6 times faster and 44.6 times cheaper than frontier baselines like GPT 5.6 Terror, GPT6 Astra, and Fable 5.1.
   • Those numbers were measured from laptops on the US West Coast pinging servers physically close to the author, comparing Jev against LLMs wrapped in a custom open-source probability adapter.
   • The host asks viewers whether Jev could replace existing LLM routing layers given the 255-option cap and strict schema requirements, or whether the lack of free-form generation makes pipelines too rigid.

關鍵概念

• Jev— The new schema-safe AI model from Typesafe AI that abandons text generation entirely.
• schema-safe AI— The core innovation: the model physically cannot output values outside a predefined schema.
• Typesafe AI— The company founded by ex-OpenAI researcher Diego Almeida that launched Jev.

精選金句

• Their thesis is that the fundamental bottleneck in software automation is the string itself and their solution is ditching text generation entirely.

  💡

  — It reframes the entire problem of AI automation — the bottleneck isn't model quality but the string format itself, which is a radical departure from mainstream thinking.

  跳至 0:21

• Jev cannot write an essay. It cannot write code. It cannot generate arbitrary strings and direct choice cardinality caps at 255 options. If you need free form text, Jev is completely useless.

  💡

  — A company deliberately shipping an AI that cannot do the most basic thing every other model does — this level of constraint is almost unheard of in the industry.

  跳至 1:15

可執行的洞察