The story of the week: Jev

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, launched Jev — a model built not to chat, but to decide.

You give it messy input and a fixed question (yes/no, a score, a choice). It returns structured answers with confidence scores in roughly 70–500 milliseconds. Pricing sits around $0.042 per million input tokens. Early users are plugging it into agent loops, trading bots, email routing, and flight search — the kind of fast structured calls that usually waste a big language model.

People are also using it against AI sludge: detectors that flag clickbait and low-effort posts on X in milliseconds, and compaction plugins that shrink million-token coding sessions down to tens of thousands without the usual fuzzy summary tax.

Vercel already had Jev on AI Gateway during the week, which helped adoption land fast.

AGI check: Jev is System-1 judgment for machines — cheap, fast, narrow. It does not settle the AGI debate. It does show where money is going this week: not every useful step needs a chatty generalist.

Elsewhere this week

DeepMind opened an institute on AGI's social impact. Demis Hassabis and Shane Legg framed current systems as near important thresholds, while still weak on coherence and creativity. First essays cover economic policy for AGI, governance, and human values. Studying the aftermath before the arrival is either prudent or premature — your call.

"Potemkin understanding" is back in the feed. A 2025 paper (Mancoridis, Mullainathan, and others) that tested GPT-4o, Claude-3.5, and Llama-3.3 is circulating again: models can define concepts cleanly and then fail to apply them. Fluency is not comprehension. Keep that next to every AGI timeline slide.

A DeepSeek kernel engineer wrote about displacement. Shengyu Liu compared watching AI take over his specialty to a master weaver watching the loom. He sketched two futures: open AI that lifts living standards, or closed systems that concentrate power. The essay landed alongside DeepSeek's efficient V4.1-Flash release — capability and labor politics in the same week.

Real-world AI use still trails the hype. The expert feed talks agents and command lines; most people still treat chatbots like a smarter search bar. The gap between demos and daily habit is part of the story too.

Closing

This week's pattern: decisions get cheaper, specialized agents get stickier in workflows, and "understanding" still looks thinner than the demos suggest.

Jev is the cleanest signal from the week ending Sep 19. We are not waiting for one model that does everything. We are wiring specialized reflexes into software that already runs our work. That can look like intelligence from the outside — especially when an agent plans, acts, and loops without you — and still leave the classic AGI questions open.

Same newsletter next Friday. Bring receipts.