Readings and reflections on recent literature and AI for Science.
Arc、CZI、DeepMind 都在押注 AI 虚拟细胞。但把尺子对准最硬的那个结果：Arc 自己的挑战赛承认模型『还没能稳定超过朴素基线』，一篇 Genome Biology 评测里基础模型甚至输给『挑高变基因』。
Arc, CZI, and DeepMind are all betting on the AI virtual cell. But hold the ruler to the hardest result: Arc's own challenge admits models 'do not yet consistently beat naive baselines,' and a Genome Biology evaluation shows the foundation models lose to simply picking highly variable genes.
Doudna 团队用 AI 造出编辑效率反超天然酶的 SynTnpB，但论文标题里的 guided 才是最诚实的信号：AI 压缩的是酶设计这段可廉价验证的前端，通往体内安全那道贵门还没碰。
Doudna's lab built SynTnpB, an AI-designed nuclease that out-edits its natural template. But the honest signal is the word in the paper's title — guided: AI compressed the cheap front end of enzyme design, and never touched the expensive gate of in-vivo safety.
通用 agentic 科研平台会把“会用 AI 做科研”变成基础能力，真正的护城河会转向问题定义、证据判断和验证设计。
General agentic science platforms will turn AI-assisted research into baseline literacy; the moat shifts to question quality, evidence judgment, and validation design.
儿童 exa-cel 证明编辑细胞可以很干净地起效, 也暴露出血液基因治疗真正难以规模化的部分: 不是编辑, 而是为了让编辑细胞住进骨髓所付出的清髓代价。
Pediatric exa-cel shows that edited cells can work cleanly, while exposing the hardest scaling problem in blood gene therapy: the price paid to open marrow niches.
礼来的 AI-bio 大单不是现金胜利，而是用 milestone 把临床前平台期权提前定价。
Lilly's AI-bio deals are not proof of clinical victory; they are milestone-heavy options on preclinical platforms.
AI 在科学里快在能被便宜验证的地方,慢在验证昂贵又主观的地方;而最汹涌的资本,正押在验证的前面、跑向终点。
在基因与细胞治疗里,编辑本身已经基本不是难题;真正决定成败的,是递送、清髓与制造,以及一个几乎无人系统在做的交叉口——HSC 体内特异编辑。