Rob Shocks breaks down Vercel's 6x speedup after switching to Jev
In a launch-week breakdown, Rob Shocks reported Jev at about $42 per billion tokens and roughly 100ms per answer, with a smart home demo returning an executable choice in 185ms.
Rob Shocks says Vercel pulled a cheap classifier, dropped in Jev, and got 6x faster on the same job - the reason it's cheaper than anything he's used shows up halfway through. this is his launch-week breakdown, and most of it happens live in the playground, not on slides. $42 per billion tokens → about 100ms per answer → 6x faster at Vercel → "it basically saturated the eval" "If you want chat, this is not the model for you." - Rob, in the first two minutes the part that got me is the smart home demo. "turn off all of the lights", and the mock house goes dark in 185 milliseconds. no text comes back, just a choice the code can run. Jev is TypeSafe's System One model, led by GPT-4 co-author Diogo Almeida. he also shows it playing Doom for an hour for around $7. bookmark & watch today ↓
This post quotes another post: “Long read: how to build a Jev-powered agent stack”