Gemini 4 Argon 审查 AI 交易机器人代码，抓出 12 个 bug
11 分钟视频记录三个机器人 18 天亏 213 美元后，作者让 Gemini 4 Argon 逐笔复盘、重写代码并设每日 3 美元上限，末尾与 Opus 5.5 做 3D 打印对比。
a guy gave 3 AI trading bots $1,000 of real money. in 18 days they lost $213 and burned $31 in AI fees doing it, so he handed the whole mess to google's new model and it called his setup "statistical malpractice" this 11 minute video is the most honest AI trading test i've seen in a while, because he shows the losses first and then lets gemini 4 argon tear his own code apart 0:38 the bots: down 21% in 18 days 1:29 argon reads every single trade 2:08 the verdict, bot by bot 3:51 it rewrites the code and catches 12 bugs 4:50 back to trading with real money 5:33 results after the takeover 6:12 argon vs opus 5.5 on a 3d printing test 7:49 the printed figures side by side 9:17 which model won my take: steal the director prompt. he gave it 18 days of trades, every rule and all the decision code, then one line: you're the director now, only what's good can stay. it cut the rule behind 64% of entries and benched the AI that rewrote the bots 29 times without a single rollback the bugs are the real lesson. his cost counter used wrong token prices and there was no pause between paid calls on a 2 second loop, so the bill could run away while the dashboard said everything was fine. he now caps argon at $3 a day and plain code takes over when it hits the cap skip the "+$5 since argon took over" part, that's a few hours of data. the repo runs on paper trading by default, keep it there would you let an AI trade your money with a $3 a day cap
这条是引用着另一条写的：x.com/sonicdr1p/status/2111145320348299292