ThursdAI: TypeSafe's Jev decision model hits 70-500ms at $42/1B input tokens, outputs free
thursdai_pod · x · 2026-09-18
This 136-minute ThursdAI episode covers:
- TypeSafe (out of stealth) launched Jev, a non-LLM "System 1" decision model from RLHF co-creator Diogo Almeida: calibrated probabilistic decisions in 70-500ms, priced at $42 per billion input tokens with free outputs, 32K context.
- On the Cua computer-use test, Jev completed 71/80 decisions at 0.1s median latency vs 4-5s for Astra — TypeSafe claims 133x faster and 444x cheaper than competitive LLMs.
- Francesco Bonacci on computer use: OSWorld was 10% two years ago; accessibility-tree control is harder to flag than CDP browser hooks; CuaDriver fetches OS-specific skills at runtime.
- Assistant startup Instinct (incorporated in April) is reportedly in talks at a $10B valuation for a free product, per The Information.
- David Pawlan ran 273 tests across 23 agents in week one for the Assistant Benchmark (116 assistants, 16 dimensions).
- Three CoreWeave-linked themes: AI decisions 400x faster, AI assistants going mainstream, and slowing frontier AI pacing.
Related event: TypeSafe Launches Jev, a Decision-Only Model That Never Writes a Word(44 posts)→
More from coding & agent
- Agent-driven DaVinci Resolve editing: dev urges Blackmagic to fully open its API — joshwhiton · 2026-09-18
- Pydantic creator builds a one-tool MCP server that drafts his emails, free to run — samuelcolvin · 2026-09-18
- Agentic AI is just arranging the world around a next-token guesser, says Microsoft's Seth Juarez — sethjuarez · 2026-09-18
- Cloudflare is building a new CI/CD system you customize with TypeScript — dinasaur_404 · 2026-09-18
- ProgramAsWeights compiles plain-English specs into tiny neural programs that run locally — yuntiandeng · 2026-09-18
- MonidHQ launches free agent web search, taking aim at Exa and Tavily's $7/1k pricing — gnukeith · 2026-09-18