What happened
TypeSafe AI released Jev on September 15, its first model after roughly two years of development; founder Diogo Almeida's announcement post drew over 37 million views on X by September 20.
Why it matters
Jev is presented as a new category called a "System One model" that outputs only structured decisions with probabilities and confidence scores — a design meant to slot into workflows as a kind of "smart if-statement" rather than a chat tool, potentially reshaping how automated sorting and routing tasks are handled.
What to watch
The 20–200x speed and 40–1,000x cost claims are TypeSafe AI's own; the real test is whether early-access users find the confidence scores reliable enough to trust without human review. Watch the waitlist access grant, which as of September 19 came with $5 in credits.
WHO IT HITSDevelopers and operations teams building automated decision pipelines — routing, triage, and output-checking — could use a model like Jev where an LLM would be slower and pricier, if the accuracy claims hold up.
Summaries like this, in your inbox every morning.
TypeSafe AI spent roughly two years developing Jev, and it is the company's first model. The name comes from economist William Stanley Jevons, whose paradox describes how efficiency gains in steam engines ended up increasing coal demand. TypeSafe AI expects AI to follow a similar pattern — as costs fall, new uses emerge.
Founder Diogo Almeida is a former OpenAI researcher and a co-author of the InstructGPT paper, the predecessor to ChatGPT. His X profile says he "co-invented ChatGPT," though the paper had roughly 20 co-authors and the RLHF method itself traces back to separate 2017 research. Almeida has framed Jev as an answer to the gap between chat models' abilities and the automation they have actually delivered, telling The Register in a statement that "humans should not be the only consumers of intelligence."
The model's design echoes psychologist Daniel Kahneman's "Thinking, Fast and Slow" — Jev is meant to be fast, intuitive System 1 thinking, while LLMs act as the slower, deliberative System 2. Whether that division of labor proves practical likely hinges on whether the confidence scores are calibrated well enough for developers to trust automatic routing at scale.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
OpenAI launched Astra, which reportedly uses "recurrent depth" to make reasoning more efficient by not spellin…
Microsoft and the University of Illinois built StudentSim, which mimics real students from as few as three ess…
A researcher at a Chinese frontier lab, who has read the Three Body Problem series since high school and grasp…
The University of Waterloo released ProgramAsWeights, which compiles an English description like "Classify urg…
Abeam Consulting and Notion are promoting an effort to shift companies to AI-driven operations and organizatio…
Zscaler introduced "Zscaler Agentic SOC," which embeds AI agents into security operations to support detection…
