拆解那个用 Jev 的 24/7 做市 bot：HFT 标签站不住

Karmi @0xKarmi原推发布于 2026.10.16

在 X 上打开原推

Before you wire Jev into the 24/7 HFT bot from a 137,696-view post, look up what HFT means. HFT runs in microseconds. Jev's maker quotes 70-500 ms per decision. I read the post and broke the bot into steps. Under the HFT label sits a 24/7 market maker where code does the math and the model decides whether to quote. → STATE ENGINE: turns the raw order book into a causally timestamped snapshot, under 400 tokens per the author. → QUESTION BATTERY: one call asks Jev 6 questions, regime to inventory pressure, claimed at single-question latency. → CONFIDENCE GATE: nothing fires without calibrated confidence; positions are Kelly-sized from the model's probability. → QUOTE PRICER: plain code prices quotes with the Avellaneda-Stoikov reservation price, which shifts with inventory. → FALLBACK LADDER: a late answer holds, an outage drops to fixed rules, a drawdown breach flattens and kills. What I checked > claimed: "millisecond speed" HFT, an 8-page paper, complete code and a bot live since yesterday, per the author. > checked: Jev is real, a TypeSafe AI model out of stealth in September on a $40M round. > checked: Avellaneda-Stoikov is a published 2008 market-making model, openly implemented. > checked: RLCD is the vendor's calibration method; independent calibration data, none found. > checked: the HFT label fails. A colocation brochure cites about 100 microseconds; the vendor quotes 70-500 ms. > not public: the paper and the code. No link, no repo, only reposted copies. > not public: any PnL, statement or wallet behind "incredible so far". Even the vendor's 70 ms floor is 700x that colocation figure. Build it as slow market making and size the loop for 500 ms. Before trusting results like these, ask for the repo, backtest table and Brier score. I laid the bot's pipeline out below, beside a checked strip of what HFT means ↓

这条是引用着另一条写的：x.com/0xkarmi/status/2108207473379180586