让 GPT-6 Astra 自主交易：三天 200 美元变 1600
作者给 Astra 一个券商账户并设下「不赚钱就消失」的规则，它按计划唤醒、读上轮日志、并行子代理验证假设，三天把 200 美元做到 1600。
I gave GPT-6 Astra access to a brokerage account. You have no other job. Earn or you stop existing Three days later: $200 → $1,600 It doesn't sit online 24/7. It runs on a schedule, wakes up in set windows during the trading day, and every run starts with no memory of the last one So the first thing it does each cycle: reads its own log from the previous run. What it tested, what it killed, why It split the research work across parallel sub-agents. One scans tickers, another builds a hypothesis from that, a third runs it against historical data A hypothesis only survives if it clears a statistical threshold. Not "looks fine," actual numbers: win rate, drawdown, whether it's just luck on the sample Night one: one hypothesis passed the backtest clean and bled $60 on the live market in an hour. The log showed why, slippage the backtest never accounted for Night two: the same model had already filtered out similar setups on its own, because it read why the last one died I never wrote a new prompt between runs In three days it ran dozens of hypotheses through the cycle and only let a handful that survived actually trade The difference between it and a retail trader was never entry speed. It's that 90% of its own ideas get killed before they ever touch the account How this cycle, test → kill → repeat, actually works is broken down in the article below
这条是引用着另一条写的：x.com/qwinsi0x/status/2089777141508374848