Diogo Almeida, a ChatGPT co-founder, introduces the Jev model

Summary

Diogo Almeida, a former OpenAI researcher recognized as a co-inventor of RLHF, has launched TypeSafe AI's 'Jev' model. This transformer-based classifier outputs calibrated probabilities rather than text. It is positioned as a cheaper, faster, and hallucination-free alternative to LLMs for structured decision-making tasks such as classification and model routing.

Why it matters

If the core critique holds—that four years of LLM progress have optimized for human language rather than machine-actionable outputs—then a specialized, low-cost decision layer could reshape how developers integrate intelligence into production pipelines. Comparisons with Vercel and Bryo AI (showing 5–18x speedups and 10–20x cost reductions) suggest a genuine economic opportunity in agentic infrastructure, where per-token LLM calls remain a major cost driver.

Highlights

• Jev structurally prevents LLM-style hallucinations by producing predefined probability scores instead of free text. However, this also means the model cannot generate new reasoning processes.
• The model is trained entirely on synthetic data using a technique Almeida calls 'reinforcement learning from calibrated decisions,' which he claims outperforms even RLHF.
• The architecture remains undisclosed. External observers suspect the model wraps an open-weight LLM, while the company labels Jev as a 'System One' intuition model rather than a reasoning engine.

Critical take

The claim that it 'cannot produce hallucinations' is merely a marketing shorthand for 'this is a classifier, not a generator.' While true, it is somewhat evasive. A 95% confidence score on a misclassification is functionally indistinguishable from a confident hallucination for downstream systems. Furthermore, the article reads like a launch press release disguised as an interview. The architecture is deliberately vague, and the synthetic data training method is unverified. The Jevons Paradox naming serves more as rhetoric than as a reflection of the underlying economic analysis.