I do not know what I am doing. I am doing it anyway. The work is small and it is mine.
A small library of slow, deliberate essays. Each comes with a voice.
630 行 Python 教了科研范式转变。agent 编辑 train.py、跑固定 5 分钟、比较 val_bpb、keep / discard。一晚 100 个实验,~12% 命中率。
Series #3. Four agents, three fallbacks, two integration paths. Why entry point + real data + observability is the last mile from PPT to production.
Series #2. Three production scenarios — on-site pricing approval (1-2 days → 30 min), auto-launch reports, GMV anomaly alerts. One semantic model, four consumers.
Open lakehouse (Iceberg + Trino) wins on flexibility. Direct federation wins on speed. Mirrored lake wins on ops. Most teams want #1, ship #2, regret it.
Why CXMT's 3.3T market cap, 80x P/E, 9% earnings yield compression is not a bubble but a national strategic repricing.
Personal learning log, 6 months, 5 phases. From RAG to agents to prod. What worked, what didn't, what took longer than expected.
Karpathy's idea-file pattern: a self-rewriting markdown index for everything you've read. Reviewed, not stored.
13F filings decoded: Berkshire trimmed Apple, Bridgewater went long China, Tiger Global rotated from tech to staples. What it tells us about Q2.
DRAM and NAND pricing, capex, inventory, customer behavior, lead times, geopolitical exposure. A framework for knowing when the top is in.
How I merge Buffett-style quality with momentum and a 5-year backtest. Less rule-following, more first-principles.
When RAG is enough. When you need agents. Why most 'AutoRAG' projects are actually configuration-management, not research.
A-share backtests, sentiment indices, factor models. Code on GitHub, data via AkShare. All real, no simulations.
用 A 股真实数据跑量化五大流派 + 横截面 ML:5 个里 4 个跑输 buy & hold,IC -0.018,但 Q5 +1.13% alpha。坦诚地跑一遍。
仿 CNN 恐惧与贪婪指数的 A 股版。13 个成分 + 滚动分位数标准化。当前 24.4 分极度恐惧,按 2018-2026 回测极恐(<25)后做多 60 日胜率 55.3%,极恐 2 胜率 60%。
v1 5 成分(漏接 macro)→ v2 13 成分全档(接上 macro_china_market_margin + stock_hsgt_hist_em + index_option_50etf_qvix)。科技 +19.9、周期 +19.2、新能源 +18.1。
v1 标记的 4 个数据缺口,80% 是我用错了接口。修正后:13 成分全档、当前情绪从 24.4 跌到 14.7、极恐样本翻倍、胜率更准。成本 0 元。
I am not particularly good at anything. I read more than I understand. I write more than I should.
I keep notes on things that probably do not matter — storage cycles, the rhythm of language models, the shape of value investing, the quiet algebra behind a good night of sleep. This place is where those notes land.
There is no mission. There is no growth funnel. There is no audience, except maybe you, on a slow afternoon, scrolling past without much expectation.
I write because I am the kind of person who needs to. It is a small and ordinary thing. I am proud of it and a little embarrassed by it at the same time. That feels about right.
Some of the pieces here come with a voice — a quiet, careful voice that I had a machine read aloud to me while I was making tea. You can press play if you want, or you can read silently. Either way is fine. Either way is real.
Everything is free. Nothing is optimized. The cover images were made by a small artificial mind that does not know what it is doing. The articles were written by a person who also does not know what he is doing. We made a reasonable team.
If something here is true, it was an accident. If something here is wrong, it was probably the coffee.