XNews

TypeSafe AI Launches Jev, a System One Model for Fast Structured Decisions

TypeSafe AI Launches Jev, a System One Model for Fast Structured Decisions

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, released Jev, a new category of AI model called System One. Jev outputs typed structured results with calibrated probabilities instead of free text, achieving 70-500ms response times and costing $0.042 per million input tokens with free output. It is designed for software integration as an 'intelligent if statement' for classification, routing, and extraction tasks.

Key insights

Jev Abandons Text Generation to Eliminate Hallucination

Jev does not generate free text; it outputs predefined structured decisions with epistemically honest probabilities. This design eliminates hallucination because the model cannot produce arbitrary strings. TypeSafe CEO Diogo Almeida argued that RLHF inherently produces hallucination by rewarding models for pleasing humans, and Jev's approach bypasses that by constraining output to type-safe structures.

Extreme Speed and Cost Advantages Over Frontier LLMs

Jev responds in 70-500 milliseconds, compared to 3-329 seconds for frontier LLMs. Input token price is $0.042 per million tokens, and output tokens are free. TypeSafe claims Jev is 20-200x faster and 40-400x cheaper than models of equivalent intelligence. These figures come from workflow evals using GPT-6 Astra and Fable 5.1 as references, with the caveat that extreme multiples represent best-case scenarios.

Parallel Sampling Enables Speed, at Cost of Text Generation

Jev uses parallel sampling, computing all outputs at once, unlike autoregressive LLMs that generate tokens sequentially. This architectural difference is similar to how Transformers replaced RNNs. The trade-off is that Jev cannot generate text; it only outputs predefined structured results. Type error rate is mathematically zero because output structure is enforced at the architecture level.

Target Use Cases: Intelligent If-Statements in Software

Jev is designed for embedding AI decisions into code workflows: classification, routing, scoring, extraction, branching, and guardrails. It serves as an 'intelligent if statement' for tasks where rule engines are too brittle and LLMs are too slow or expensive. Example demos include real-time DOOM AI (10 calls/second, ~$7/hour) and Wikipedia racing, outperforming non-reasoning LLMs in speed and steps.

RLCD Training Method Replaces RLHF for Calibrated Decisions

Jev was trained using RLCD (Reinforcement Learning for Calibrated Decisions), a new method developed by TypeSafe. Unlike RLHF, which Almeida argues produces overconfident and inconsistent confidence scores, RLCD trains the model to output well-calibrated probabilities. Developers can set confidence thresholds: high-confidence decisions execute autonomously, low-confidence ones escalate to human review.

Jev Named After Jevons Paradox: Efficiency Drives Demand

The name 'Jev' references William Stanley Jevons, who observed that steam engine efficiency gains led to increased coal consumption. TypeSafe bets that lowering AI decision cost will massively expand usage. The company is already profitable and plans to continue lowering prices, with output tokens free and input tokens cheap. The economic logic is that cheaper intelligence per decision will unlock new applications.

Developer Demos Show Practical Integrations

Third-party demos include semantic search in web pages (Ctrl+F with meaning), a PostgreSQL function for semantic retrieval (e.g., 'find people who can work remotely'), and analyzing millions of user event logs to identify bugs and generate pull requests. These illustrate Jev's potential as a drop-in component for structured decision-making in existing systems.

Source posts