有人让 GPT-6 Astra 跑 1024 个交易策略基因组 67 小时

Valentin @martynov014原推发布于 2026.09.18

在 X 上打开原推

whoever let this run unattended for three days has more nerve than sense someone handed GPT-6 Astra a thousand trading genomes and one instruction: survive fees, drawdown, and data you have never seen. sixty seven hours later one of them was up 389.5 on a single position and the other 1,023 were in the archive with their reasons written down. the book went from 1.9042 ETH to 4.8871 every strategy that ever worked left a trail in four places: the entry, the exit, the fee, and the worst moment in between. almost nobody reads all four together this is GPT-6 Astra, the layer that runs a whole population instead of one prompt, $0 on top of the plan you already pay for: - give every genome the same $1,000 and the same tape. 1,024 of them, 141 days of history, identical conditions, no hand-tuning between rounds - make the veto a separate agent. 2,318 handoffs went down the chain in 67 hours and RISK killed 54 of them, each with the reason written into the ledger - settle the exit before the entry. STOP and TAKE agree on one number before any token is named, which removes the argument that eats the whole run - measure the tail, not the average. 15,770 sessions, 6.14% tail mass, implied multiple ×52.3. the average is where strategies go to look fine - lock the winner before it sees the test. the holdout is opened last, once, and nothing gets funded on a lucky run none of this was clever. it was just kept the catch is slow money. it is a simulation, the fills are modelled, and a population this size costs real tokens to run for three days. nothing here has touched size a losing strategy is a story told one trade at a time every week you tune one prompt by hand, someone else is running a thousand and reading the archive bookmark this, the full architecture of the troop is in the post below ↓

这条是引用着另一条写的：x.com/tonyprediction/status/2100235209618722955