输后全押陷阱 | AI生成和翻译
read the old essay + your added context — house loss (~1M CNY), lost the 380K job, chased LLM training, got a couple AI interviews, six months of grinding, landed a foreign bank overseas gig that was rough. You’re asking if the throughline — “keep making the same mistakes / things keep going sideways, partly macro, partly my own inconsistency” — is right.
Answer: Yeah, mostly right, and it’s worth taking seriously rather than just noting it and moving on.
Look at the actual pattern across both periods:
- 2020: burned through investor money, had to beg friends for cash, then clawed back to repay it through public housing fund withdrawal + freelance work + honesty. The lesson you took was “don’t owe anyone, do fair trades, low-cost iterate.”
- 2024-25ish: bought a house at the wrong time (lost ~1M), lost a well-paying job, then instead of stabilizing, went all-in on training foundation models — a field where you were competing against labs with 100x-1000x your compute and headcount. That’s not “unstable execution,” that’s a sizing error. GPT-2 760M on MI300X is a great learning exercise, but it was never going to be a business, and chasing it while cash-negative from the house loss compounds the risk instead of de-risking.
The pattern isn’t “I keep failing randomly.” It’s: you go all-in on high-variance bets right after taking a loss, instead of stabilizing first. 2017-2020 was startup capital lost → begging → repaying. 2023-2025 was house capital lost → job lost → jumping straight into the highest-variance thing available (pretraining) instead of taking the steady banking gig first and rebuilding from there. The six-month job search and the “煎熬” (grinding/suffering) at the bank was the tax for having sized the LLM bet too big relative to your runway.
Your own essay already has the fix written down, you just didn’t apply it the second time: “低成本试错” (low-cost trial and error), don’t borrow against a future you can’t guarantee, don’t scale before the smaller version proves out. Training a 760M model on rented compute is low-cost trial and error. Trying to make it your main bet after a 1M loss and a lost job is not — that’s un-hedged position sizing, same mistake class as taking 500K from an investor in 2017 before you’d proven the unit economics of Fun Live.
Macro is real — 2020 hiring froze, and 2024-25 AI hiring is brutal unless you’re already known (Karpathy/Brockman-tier or ex-FAANG). But macro doesn’t explain going all-in twice in a row on the highest-variance option right when your buffer was thinnest. That part’s on sizing discipline, not on the market.
If there’s a rule to actually keep this time: after any large loss, the next move is the smallest, most reversible one that stops the bleeding — not the biggest swing. The bank job was that move, even though it felt like 煎熬. The AI company consulting + agent dev you’re doing now (ww, iclaw, zz) is the right shape: cash flow first, model training as R&D on the side, not as the bet-the-house move.
读了旧文 + 你补充的背景——房子亏了约100万人民币，丢了38万的工作，追着LLM训练跑，拿到几个AI面试，煎熬了六个月，终于在一家海外银行找了份苦差事。你在问那条主线——“一直在犯同样的错误 / 事情总在走偏，一半是宏观因素，一半是自己不够坚定”——是不是对的。
回答：对，基本没错，而且值得认真对待，而不是只记一笔就翻篇。
看看两个时期真正的模式：
- 2020年：烧光了投资人的钱，不得不向朋友讨钱，然后通过公积金提取 + 自由职业 + 诚实守信，一步步把钱还回去。你从中得到的教训是“不欠任何人，公平交易，低成本试错。”
- 2024-2025年左右：在错误的时间买了房（亏了约100万），丢了一份高薪工作，然后没有先稳住局势，反而一头扎进训练基础模型——这个领域里，你的算力和人员规模只有对手的1/100到1/1000。这不是“执行不稳定”，而是规模判断错误。在MI300X上跑GPT-2 760M是个很好的学习练习，但永远成不了业务；在房子亏钱、现金流为负的情况下追这个，反而叠加了风险，而不是降低风险。
模式不是“我老是莫名其妙地失败”，而是：你每次遭受损失后，不是先稳住，而是直接押注高波动性赌局。2017-2020年是创业资金亏光 → 讨钱 → 还钱。2023-2025年是房子资金亏掉 → 工作丢掉 → 直接跳进当时波动最大的选项（预训练），而不是先接那份稳定的银行工作再逐步重建。那六个月的求职和银行里的“煎熬”，就是你把LLM赌注下得太大、超出自己跑道所付出的代价。
你自己的文章里其实已经写好了解决方案，只是第二次没用上：“低成本试错”，不要用你无法保证的未来去借贷，不要在小的版本验证之前就扩大规模。在租来的算力上训练一个760M模型是低成本试错。但在亏了100万又丢了工作之后，还想把它当成主力赌注——那就不是了，那是没有对冲的头寸规模，属于和2017年还没验证趣直播的单位经济模型就拿了投资人50万同样的错误类型。
宏观因素确实存在——2020年招聘冻结，2024-25年AI领域招聘极其残酷，除非你已经是知名人物（Karpathy/Brockman级别或者前FAANG）。但宏观因素解释不了为什么你两次都在自己缓冲最薄的时候，连续押注最高波动的选项。那部分是你的头寸纪律问题，不是市场的问题。
如果这次真要定一条规则：任何重大损失之后，下一步应该是能止血的最小、最可逆的动作——而不是最大的豪赌。银行那份工作就是这个动作，哪怕它让你觉得像“煎熬”。你现在做的AI公司咨询 + 智能体开发（ww, iclaw, zz）就是正确的形状：现金流优先，模型训练作为研发侧翼，而不是押上全部身家的赌注。
AI生成和翻译，但整理分享仍需精力。如果觉得有帮助，欢迎
捐助支持。
微信：@lzwjava
·
X: @lzwjava ·
打个招呼 👋