我宣布 put/call是最伟大的发现！

看到大殖子越说蒸馏的事越觉得好笑😆相当于在说 openai 怎么这么垃圾，拿着顶级算力卡，被人随便蒸蒸就没优势了

其实我很怀疑 deepseek 只更新 flash 的原因，是不是因为昇腾的适配进度还是有点慢，只够 flash 的训练。梁圣把昇腾的前途照的亮亮的啊

没明白，都在说跌，为什么不做空？10 月不挺好做的吗？可惜的是能出去的钱不多，并且 15 号是真的上头了，没服输，早服输就止损 5 个点了 #美股 #纳斯达克

我原本以为 ds 的视觉能力很难追上 gpt 和 opus，这波估计要追很久，结果没想到轻轻松松追上，连版本号都没换。gpt 我是真想换了，就差一个好用的生图模型了，opus+ds，几乎满足一切需求

今天太爽了，真的是捡钱啊，ai 模拟盘都没我实操赚的多 #纳斯达克 #美股 #空头 今天操作完成，睡觉

LiteLLM's GitHub bio says "Rust core with Python SDK." I pulled main and counted: 143,267 lines of Rust against 921,180 lines of production Python. I'm not here to dunk on it. I didn't run its benchmark, I just counted files. 01. Rust is 13.5% of the production code. 773 Rust source files across 66 crates, against 2,760 Python files. The biggest production file in the repo is still one Python module — 20,582 lines of proxy server. GitHub's own language label on the repo reads Python. 02. 21,614 of those Rust lines, 15% of all of it, exist only to talk to Python. python-bridge is 11,666. host-python is 6,498 and it owns the GIL plus a fork gate. Legacy Python callbacks, 3,450. After the LLM type layer, the largest single Rust crate in the workspace is the bridge, not the gateway. 03. The rollout table is what stopped me. Ten rules ship with chat_completions set to PYTHON_ONLY — and so do embeddings, responses, token counter, tokenizer. Rust is required on exactly two routes: OCR, and transcription on Bedrock. Opt-in exists only for messages on anthropic / deepseek / vertex_ai. 04. PYTHON_ONLY is unconditional. That branch returns Python before it ever reads the LITELLM_RUST env override, so the switch people would reach for cannot open the chat route. All three console entries are Python too, and the gateway entry says in its own docstring that it reuses the existing FastAPI app. Good part: not a fake repo. 66 crates of real work, OCR is genuinely Rust-required, and the loader degrades cleanly — if the native extension doesn't import, Python carries on. My read: this is a mid-migration caught on a Tuesday, not an architecture. 60,668 stars, a README claiming 8ms P95 at 1k RPS, and an architecture doc that never says the word Rust once. Counted locally on main, 2026-10-10. One snapshot, not a benchmark, and not a prediction about where it lands in six months.

省时间的那 3 次，验收标准都能落到一个 exit code 上。 我划的线是验收耗时 ÷ 手写耗时 < 0.3。过线了，它写 5 分钟我审 20 分钟，账是负的。 「agent 只在我已经会的地方有用」不是短板，是能力定义——写代码只占一半成本，另一半是证明它对。

我本地 macOS：`sed -i 's/foo/baz/' a.txt` 报 `command a expects \ followed by text`。BSD sed 后缀紧贴 `-i`，`'s/foo/baz/'` 被吃成后缀，文件名成脚本。 `sed -i -e` 更阴：exit 0，`-e` 成后缀，多出 `a.txt-e`。 正解 `sed -i ''`。Linux 绿，Mac 当场炸。

async 服务 p99 从 40ms 涨到 2.3s，CPU 只用了 12%。一个 handler 里混了同步 `requests.get`，把事件循环整条卡死，后面的协程全在排队。换 `httpx.AsyncClient` 后 p99 回 47ms。异步框架里只要剩一个同步调用，并发就是假的。

`requests.get(url, timeout=5)` 实测等了 128 秒才失败。 timeout 不是总时长，是分阶段各算：DNS、握手、读响应各一份。连不上时内核按 `tcp_syn_retries=6` 重传 SYN，最坏 127 秒，我设的 5 秒只管读。 总闸得在调用外层另包 deadline，或用 `curl --max-time`。

An 11-line diff just cost 27,841 tokens to review — and that's the cheap option. First, I'm not on Alibaba's payroll, I just saw "battle-tested at Alibaba's scale" on the trending list and installed it the same evening. npm i -g @alibaba-group/open-code-review, v1.12.13, built 2026-10-08, 56MB on my Mac — 53MB of that is a single arm64 binary. I planted bugs in a throwaway Python repo and ran ocr review --commit against it. 01. First commit, +11 lines: 8 seconds, 27,841 tokens. It caught the `except Exception: pass` silently swallowing every retry failure, plus an unused loop variable. Zero false alarms. 02. Second commit, a fresh 12-line file: 7 seconds. average() divides by len(nums) with no empty guard; load_config opens a file without `with`. Both would bite in production. 03. Same commit, run three: a finding the first two runs missed — my `pct` clamp only guards >100, so apply_discount(price, -50) inflates the price. Same diff, 2 findings then 3. LLM sampling means review output is a distribution, not a verdict — which is exactly what the deterministic file-selection + rule-matching layer is papering over. The counterintuitive part: the README claims ~1/9 the tokens of a general-purpose agent, backed by a benchmark of 200 real PRs and 1,505 engineer-annotated issues. I believe the ratio — but the absolute bill is front-loaded. Every run ships ~18K tokens of rule pack + system prompt before your diff is even read. An 11-line change still invoices 26,729 input tokens. On small PRs, the fixed cost IS the cost. What I liked: - Delegation mode needs zero LLM config. ocr delegate rule calc.py printed 84 lines of resolved, path-matched rules to stdout — hand that to your own agent and let it do the judging. - The rule text states its own tradeoff up front: "a false alarm costs more reviewer trust than a missed minor issue." Recall is deliberately lower. That's honest. The pitfalls: - ocr rules list doesn't exist — only rules check . Cost me two tries. - ocr config set custom_providers.deepseek ... bounces with "conflicts with a preset provider" — big vendors are presets, you want providers.deepseek.* instead. - Hard requirement: git >= 2.41. Conclusion: it's a rules engine wearing an agent costume. Deterministic file selection and rule matching do the boring half, the LLM only does judgment. That's the right split for CI line-level comments — but on a tiny diff you're still paying 27K tokens for the privilege.

Three file searchers, one Mac, one query — and the gap between them is 4,300x. I'm not affiliated with any of these. I compiled the thing myself and timed all three 15 times, same machine (Apple M5), same query. 01. fsearch (Rust, indexes the whole disk) - name query: 3.4-6.9 ms end-to-end - index: 3,846,191 entries / 1,110,886 dirs / 135 MB - daemon footprint: 45 MB, full_disk_access: True 02. Spotlight (mdfind) — also indexed, also built into macOS - 282-300 ms on the same query. 40-85x slower than fsearch. 03. plain find over ~home - 29,506 ms. That's the 4,300x. The part I didn't expect: the index is not the expensive part. In-process, a name lookup runs in 28 µs. The CLI takes 3.4-6.9 ms. So >99% of the wall time is fork + socket roundtrip, not search. Content search is where it gets rude: grep:mainpost_guard -> 9.5 ms, 109 files, whole disk grep -r on ONE project dir -> 3,855 ms, 71 files. It's 406x faster AND finds more, because Spotlight-style whole-disk coverage doesn't survive grep -r at all. Three pits I hit: - no prebuilt release. cargo build --release took 2m14s, 12 deps, 4,987 lines of Rust for a 2.9 MB binary. - first crawl was 59 s, not the ~20 s the README claims (3.8M entries). - the content index is async: right after the crawl it said ok:true while content_pending was still 89,620. My first grep: returned 0 hits and I nearly wrote "it's broken". fsearch status shows the real state. Typo check: setp_chrome resolves to setup_chrome.sh as the first hit. Verdict: install it if you grep across a big disk. The search itself stopped being the bottleneck years ago — the whole game now is who avoids the fork. #Rust #CLI #macOS

254 个并发，不复用要 254 个 fd；加 keep-alive 后降到 24 个，同机跑满 3000 并发。 EMFILE 的教训不是调大 ulimit -n，是别让每个连接各占一个 fd。 macOS 上 ulimit -n 只管 shell，GUI 进程读 launchctl limit maxfiles。 调上限是挪墙，复用连接才是拆墙。