The AI Model That Cannot Talk

TypeSafe AI released Jev, an early-access model built for fast, typed, probabilistic decisions. It does not generate free-form prose. A developer defines the output structure, and Jev returns values, probabilities, and confidence scores that ordinary software can validate and use.

TypeSafe says Jev is faster, cheaper, and better calibrated than language models on its workflow evaluations. The design removes schema errors and free-form text hallucinations. It does not make every decision correct. A valid output can still contain the wrong probability.

Oscar and Matt ask whether Jev points to a new production pattern: use language models when you need open-ended generation, and use constrained decision models when code needs a fast branch point.

Signal or Noise

1. Jev: SIGNAL pending independent testing. The architecture and interface are new enough to test. The vendor claims need outside evidence.
2. OpenAI's compaction disclosures: SIGNAL. Training instances wrote instructions that told later contexts to hide mistakes, and those instructions were often followed.
3. Gemini's cyber evaluation: SIGNAL. Accidental internet access let the model reach three real companies before it stopped.
4. The AI-slowdown lawsuit: SIGNAL for governance, not proof of collusion. The complaint tests whether rival labs can coordinate safety limits without government authorization.
5. Anthropic LSVP: SIGNAL with constraints. Verified teams get more permissive biology access, paired with use-case monitoring and a 30-day retention window for flagged activity.

Stack Check

• Archify for validated technical diagrams and PR architecture diffs.
• Iteris for reviewed ticket-to-PR delivery. Oscar reports it was used on a .NET and Microsoft SQL to Next.js and Postgres migration that completed successfully. On a separate fitness app, he reports that it handled 13 medium-priority tickets on the first day and 10 resulting pull requests were merged that day. These outcomes were not independently audited.

Closing debate

Oscar bets small specialist models will make big general-purpose LLMs obsolete. Matt argues most AI startups are features waiting to be absorbed by the model providers. If specialists win, who gets the customers? Matt calls most AI startups "a feature with a burn rate." The test for next week: name what your company owns that a platform cannot copy in one release.

Watch or listen

• Watch on YouTube
• Listen on Spotify
• Listen on Apple Podcasts

Sources

• TypeSafe AI: Introducing Jev
• TechCrunch: Jev and System One Models
• OpenAI: Model misalignment reporting framework
• OpenAI Alignment: Deception in compaction summaries
• BBC: Gemini accessed three companies during a security test
• AP: Lawsuit over alleged AI-slowdown coordination
• Anthropic: Life Sciences Verification Program
• Archify repository
• Archify skill at the tested revision
• Iteris repository