Long post argues Jev is not just function calling with extra steps
The author rebuts the claim that Jev is just function calling with extra steps, explaining that it only outputs a choice, score or gate with a confidence value, and shares speed and cost figures plus how to use it as an agent approval gate.
someone in my feed called jev "just function calling with extra steps." i don't think that's right, and the reason is annoying to explain, which is usually a sign it's actually interesting. typesafe ai came out of stealth sept 15, two years quiet, $40m seed. founder's been through chatgpt / instructgpt. their model is jev. it does not write text. it cannot. you give it program state plus typed questions, it gives you back a choice, a score, a gate — with a confidence number attached, in one parallel pass. here's what got me. every agent loop today makes the same mistake: it asks the same model to write the plan AND decide which branch to take. one job needs slow deliberate reasoning. the other needs a fast, calibrated yes/no under load, thousands of times a second. typesafe's bet is that these were never the same job — llm writes, jev decides, code acts. numbers, their own benchmarks: → ~193x faster than comparable llm calls on classification → ~445x cheaper per decision → most calls resolve in ~100ms → input $0.042/m tokens, output free (there's nothing to generate) fine print, because a model that "cannot lie in prose" is still a marketing line. this is new — independent benchmarks are thin, it's their own demo numbers so far. it only works when the possible answers are known upfront, so it's a router/classifier/gatekeeper, not a reasoner. couple your approval logic to a proprietary decision model and you've got a new kind of vendor lock-in, just a quieter one. next up, people are already wiring this into coding agents as the approval gate before a tool call fires. someone made it play doom using only typed decisions, zero generated text. that's either a stunt or a preview, hard to tell which yet. the llm was never the slow part because it was dumb. it was slow because you were asking it to decide when you only needed it to guess.