On September 17, 2026, AI research laboratory TypeSafe AI introduced 'Jev', fundamentally challenging the prevailing dogma that all AI interactions must take the form of conversational natural language generation. Jev is the industry's first purpose-built 'System One' foundation model, engineered specifically to provide machine-native, structured, and probabilistic decisions for software agents.
The concept draws directly from psychologist Daniel Kahneman's seminal dual-process cognitive framework. In human cognition, 'System One' operates automatically and quickly with little effort, while 'System Two' allocates attention to effortful, complex mental operations. In modern AI development, however, engineers have been forced to use heavyweight System Two reasoning models (like Claude Sonnet or GPT-4o) to handle lightweight System One operational decisions—such as whether a bash command succeeded or which tool to invoke next—wasting billions of dollars in redundant compute.
Jev completely strips out the autoregressive text vocoder and generative decoder. Instead of outputting free-form Markdown or messy JSON text that requires fragile regex parsing, Jev accepts unstructured program state (such as error logs, HTML DOM trees, or database records) alongside a typed question schema. The model projects the input through a specialized transformer backbone trained via Reinforcement Learning for Calibrated Decisions (RLCD), outputting calibrated probabilities directly into strictly defined types.
Jev operates on three core primitives: 'Choice' (selecting one item from an exhaustive list of enum strings), 'Score' (assigning a calibrated ordinal floating-point value), and 'Noul' (a strict binary judgment of true or false). Because valid response coordinates are compiled into the model's output layer in advance, Jev is mathematically incapable of hallucinating invalid JSON keys or committing schema type mismatches.
The performance metrics are staggering. In benchmarking conducted against OpenAI and Anthropic models across 100,000 classification and routing tasks, Jev operated up to 200 times faster, recording an 8.2ms median response time compared to 1,600ms for conversational models. Priced at just $0.05 per million input tokens, Jev is over 400 times more economical than traditional frontier LLMs.
The release immediately ignited the developer community. Over 700 engineers on Hacker News mobilized around 'OpenJev', an open-source initiative designed to replicate Jev's single-pass logit projection techniques on open-weights foundation models such as Qwen and Gemma. Production agent frameworks including Claude Code, Stagehand, and browser automation suites are already integrating Jev endpoints to handle inner-loop triage.
By recognizing that software agents need fast, reliable decisions rather than verbose polite text, TypeSafe AI has introduced a vital new architectural primitive that will dramatically accelerate the transition to fully autonomous agentic workflows.