AI High-Frequency Trading? This 100ms Claim Doesn’t Hold Up

0:00Most people don’t realize that you can use Jev

0:01to build a 24/7 high frequency trading system

0:03that makes trading decisions in under 100 milliseconds

0:04while normal LLMs take three to 10 seconds.

0:06Let's fact check this video.

0:08First thing I wanna say is,

0:09you're not building a high frequency trading system, period.

0:13You cannot build one on regular home internet.

0:18In his video, he explained that you can use Jev

0:20to respond in less than 100 milliseconds,

0:25even if that's the case.

0:27High frequency trading systems respond

0:30in the order of microseconds.

0:33Even nanoseconds, you can't respond in 100 milliseconds

0:38for a high frequency trading system.

0:41You just simply can't.

0:43Here's how you set it up in three steps.

0:44Step one, connect your live exchange feed.

0:45Instead of dumping raw charts into AI,

0:46your code summarizes buyer demand.

0:47Order depth and price data into a tiny data snapshot.

0:48Two, send that snapshot over to Jev

0:50using the type-safe SDK.

0:50Jev then evaluates multiple typed questions in parallel,

0:52flagging whether incoming market flow is informed or noise.

0:53This part of the video is pretty interesting

0:55because I can see a universe where it might actually work,

1:00but he does a horrible job of proving it.

1:02He has no evidence that his set of rules,

1:05his set of questions, actually lead to better returns.

1:09Step three, execute with deterministic pricing.

1:10Your code then computes the reservation speed

1:11using the Avellaneda–Stoikov formula,

1:12verifies account limits and automatically cancels

1:14or posts limit orders.

1:14The best part is Jev is trained with RLCD

1:15instead of human preference.

1:16This means that high confidence scores

1:17actually correlate with real market probability.

1:18That's a lot of claims for absolutely no evidence.

1:22Like if he had something,

1:24even if it were back test results out of sample,

1:28then I might give him a little bit of credit.

1:30Sounds like a decent idea on paper,

1:33but he has absolutely no evidence

1:35that the reinforcement learning

1:37for calibrated decision making

1:38leads to better decision making

1:40than I don't know, a specially trained model.

1:43And the fact that he's presenting the video in such a way

1:46is extremely irresponsible.

1:49Someone might actually try to use Jev for this,

1:52trade on it and then lose a bunch of money.

1:56And I think the worst part about this video

1:58is that he knows he's lying.

2:01He's an ex-SWE at Meta,

2:04and he knows what he's telling you

2:06doesn't make any actual sense.

2:09But he doesn't care because the likes, comments,

2:12and follows he's getting from this video are reward enough.

2:17Do better.