Jev does on-chain market making: six order book reads in 81ms
Jarrod Watts used Jev to quote live on an on-chain order book, spotting toxic flow and widening spreads, with risk parameters hard-coded so the model never votes on them.
Wall Street paid $300 million to shave 3 milliseconds off a trade, and a co-author of InstructGPT just priced a 24/7 trading brain at $10 a month... Citadel Securities and Jane Street still own the microsecond lane, but on-chain order books move at block speed, and nobody owns that lane yet. TypeSafe answered with Jev, a model that writes nothing and returns 6 calibrated reads of the order book in 81ms. Three days after launch Jarrod Watts had it quoting live on an on-chain order book, and the repo passed 1,000 stars. It spots toxic flow, widens the spread on a thin book and pulls quotes before the next block lands, all while you sleep. Drawdown, position size and stale data stay hard-coded, and the model never gets a vote on them. A full battery every block costs $10 to $25 a month, the same loop on a frontier LLM burns $50 to $150 an hour ↓
This post quotes another post: “Building an HFT trading bot with Jev: one buy/sell decision per block”