wenchangyue.bsky.social @wenchangyue.bsky.social · 03/08/2026One week of AI-assisted development produced what I estimate would take one conventional programmer more than six months. I did not run the computer 24/7. I shut it down when I went home each evening. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 19/07/20261/ On July 17, 2026 the FDA approved Lipfendra (enlicitide), the first PCSK9 inhibitor that comes as a daily pill instead of a shot. It lowers "bad" LDL cholesterol in adults, including the inherited form. Made by Merck. Here is the logic. 🧵 130
wenchangyue.bsky.social @wenchangyue.bsky.social · 15/07/2026Your body picks its fuel every minute: fat, glucose, or protein. One hormone makes the call. Insulin high after a meal means burn and store glucose. Insulin low in fasting flips everything toward fat, and the liver starts making ketones. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 10/07/2026A small controlled test on the thing I care about: can a model catch real problems in a biomedical review manuscript? Same manuscript, same prompt, same offline sandbox, same ~88.5k-token budget. Only the model changed: GPT-5.5 vs the newer GPT-5.6 (codex build). 110
wenchangyue.bsky.social @wenchangyue.bsky.social · 05/07/2026On July 1, 2026, the FDA expanded the CRISPR gene therapy Casgevy (exa-cel) to children as young as 2 with sickle cell disease or transfusion-dependent β-thalassemia. First gene therapy for sickle cell down to age 2 (was 12+). 🧵 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 03/07/2026Anthropic shipped Claude Science this week. The useful thing to know first: it's not a smarter model. It runs the same Claude models everyone already has (Opus 4.8), no special access, no gating. What you get is the wiring, not more intelligence. 102
wenchangyue.bsky.social @wenchangyue.bsky.social · 02/07/2026On June 10, 2026, the FDA added an acute kidney injury warning to alli (orlistat) 60 mg, the only OTC weight-loss drug in the US. Flagged as a rare side effect. The mechanism is the interesting part. 🧵 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 29/06/2026Everyone says Codex is more "durable" than Claude, so I got the $100 Codex plan too and started switching. It went the other way. My honest take: Claude lasts longer — same intensity, but Codex's weekly quota drops faster, and almost every week I'm waiting on the reset. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 27/06/2026This is one of my two personal Macs, and the other one isn't even counted here. On just this one: 5 working days of AI coding (Claude Code + a Codex agent) = ~1 billion tokens, about $1,400 at API list price. More than a US postdoc earns in a week. 130
wenchangyue.bsky.social @wenchangyue.bsky.social · 26/06/2026A high hCG doesn't always mean pregnancy. The same hormone can also be a tumor marker. A short thread on how one molecule shows up in two very different settings. 🧵 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 24/06/2026Codex was writing to my SSD much more than it should. I checked because of reports about Codex TRACE logging. My machine showed the same pattern. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 24/06/2026One workflow has been quietly changing how I check scientific writing. I keep the manuscript, figure legends, scripts, and source tables open in VS Code, then let Codex and Claude Code read the same files from different angles. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 23/06/2026Claude Code and Codex both read "skills," reusable instruction packs you write once and reuse. The trap: each keeps its own copy. Edit one, the other quietly goes stale. Ours drifted two weeks before anyone noticed. What fixed it 👇 300
wenchangyue.bsky.social @wenchangyue.bsky.social · 22/06/2026Sharing a case. 31-year-old man: a month of cough, chest discomfort, exertional dyspnea, plus 4.5 kg weight loss. CT: large anterior mediastinal mass. Labs: AFP and β-hCG both up. Answer: nonseminomatous germ cell tumor. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 21/06/2026A lot of what we do in our neuroendocrine prostate cancer (NEPC) work runs on public data. Sharing the databases we actually use day to day, in case it saves someone some digging. 🙏🧵 120
wenchangyue.bsky.social @wenchangyue.bsky.social · 21/06/2026Two terminal coding agents, two philosophies for the same job: Claude Code vs OpenAI Codex, by how each keeps its config on disk. Both live in a home dotfolder (~/.claude vs ~/.codex) + per-project files. From there they diverge 🧵 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 21/06/2026Where Claude Code keeps its config: a few files in two folders, across four scopes. See the map once and setup stops being guesswork. Two homes: • ~/.claude/ → you, every project • ./.claude/ (+ repo root) → this project, shared via git 🧵 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 19/06/2026Want your AI coding session to last far longer before it chokes on its own context? It is less about compressing the chat and more about what you let into the window in the first place. Three moves that actually work: 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 19/06/2026Sync all your Claude Code skills across machines with one git repo and one symlink. Setup: 1. Make ~/.claude/skills a git repo and push it to GitHub. 2. Move your global ~/.claude/CLAUDE.md into that repo, then symlink it back. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 19/06/2026Mendelian randomization (MR) is having a moment in biomedicine, pitched as a "natural randomized trial" you can run from summary stats. It's powerful. It's also routinely over-interpreted. A thread on what MR CAN and CANNOT tell you. 🧵 110
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/202610 colour palettes I reach for in research charts — curated, and most are colourblind-safe. Categorical + sequential: Okabe-Ito, Paul Tol, Tableau 10, viridis, cividis, magma. Skip default tab10 and rainbow/jet. Which do you swear by? Anything I'm missing? #DataViz #ColorPalette #SciComm 34914
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/2026Fed held at 3.50-3.75% today, unanimous. Warsh's first FOMC as chair, no cut signal. Yet semis are outperforming: $AMAT and $LRCX green while the S&P and Nasdaq slip. AI capex over the Fed, so far. Rotation for real, or a one-day fade? What's your read? #FOMC #Fed #StockMarket #Semiconductors 010
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/2026AI-written English has tells. I keep a style pass that strips them without changing meaning, numbers, or citations. The loudest one: em-dash pile-ups. The model wraps asides in "— like this —". Replace with commas, a colon, or two sentences. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/2026AI 写的中文是有破绽的。我整理了一套只改语言、不碰事实的方法，专门抹掉这些破绽。 最显眼的是破折号塞长插入语：「状态最差时——累、烦、刚失败——还能…」。改法是拆成正常句子，用逗号或句号断开，别让一段里堆好几个。 第二个也最普遍：「不是 X，而是 Y」这种对称句。模型特别爱用，一段能塞好几处。能直说就直说，只有当它是真正的对比、又是最清楚的说法时才留。 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/2026Messi, still bending time. A World Cup hat-trick for Argentina in a 3-0 win over Algeria. At 38, he keeps making the impossible feel routine. Congratulations, Leo. Archive photo: Kirill Venediktov/soccer.ru, CC BY-SA 3.0, via Wikimedia Commons. 000
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/2026I audited my own genomics pipeline this week. 86% of it is R. 126 R scripts, 14 Python, 7 shell. That's roughly the honest state of bioinformatics right now. R does what it's best at: statistics, Bioconductor, bulk omics, survival models, publication figures. 2101
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/2026My screener hit a perfect 4/4 market score today and named 4 fresh breakout setups: $AMAT $VCYT $LION $GEO. I bought zero of them. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 17/06/2026My screener hit a perfect 4/4 market score today and named 4 fresh breakout setups: #AMAT #VCYT #LION #GEO. I bought zero of them. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 16/06/2026Computers can't make truly random numbers. They run a fixed formula that produces a sequence which only looks random. A seed sets where that sequence starts. Use the same seed and you get the same numbers every run. 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 16/06/2026做生信和临床数据分析,最贵的错误大多不在代码,在分析逻辑。而且这几类,训练有素的人也会反复搞混。下面是我常给自己提的醒👇 100
wenchangyue.bsky.social @wenchangyue.bsky.social · 16/06/2026系统今天给市场打了满分,推荐 4 只票。我一只没买。 先认错:上周三系统把 LRCX 排第一,我说"等美联储议息再买",结果它 4 天涨了 7%,我踏空了。 但今天我还是没追。说说为什么 👇 220
