用 Jev 替换 1140 条退款 if-else 规则

starmex @starmexxx原推发布于 2026.10.06

在 X 上打开原推

ANDREJ KARPATHY CALLED THIS IN 2017. NINE YEARS LATER IT FINALLY HAPPENED TO THE MOST BORING CODE YOU OWN: YOUR IF STATEMENTS his software 2.0 essay said neural networks would slowly eat hand-written code. everyone assumed they'd start with the hard stuff they started with the boring stuff this is what deciding a refund looked like in my codebase last month: if "refund" in msg and order.age_days < 30: if not user.flagged and http://order.total < 500: if "damaged" in msg or "never arrived" in msg: approve() elif user.refunds_this_year < 3: escalate() ...and 212 more lines of exceptions to the exceptions this is what it looks like now: decision = jev.decide(ticket, schema=RefundDecision) jev is the model typesafe shipped last month that never writes text. you give it the messy input and the shape of the answer, it gives back a typed decision with a real confidence score in about 70 milliseconds what happened when i swapped it in: → 1,140 hand-written rules across 9 services, gone → 9,300 lines deleted in one week → the edge cases my rules kept missing, like "it arrived but the box was wet", now get handled because nobody had to predict them → every low-confidence call still goes to a human, so nothing risky runs on a guess vibe coding was the 2025 version of karpathy's idea. 2026 is quieter and stranger: deleting the code you were proudest of and replacing it with one line that understands the problem better than your rules did what's the ugliest if-else chain in your codebase right now? the full migration is in the article below ↓

这条是引用着另一条写的：x.com/rewind02/status/2107063404691132663