🌐 双语
Archive

AI Builders
Digest

2026-11-22 12 builders · 23 tweets · 1 podcasts · 0 blogs

🔥 热点话题

AI驱动药物发现:Isomorphic Labs如何把药研搬上计算机Powering AI-Driven Drug Discovery with Isomorphic Labs President Max Jaderberg

The Takeaway:药物发现最大的机会,是从现实世界中缓慢的经验试错,转向能达到实验精度的预测计算模型,从而在进入临床试验前大幅提升成功率并减少迭代轮次。

Max Jaderberg是Isomorphic Labs总裁及创始团队成员。这家由Alphabet支持的公司目标是解决所有疾病。他此前在DeepMind工作多年,参与早期生成模型、强化学习和AlphaGo/AlphaStar等项目,并在AlphaFold证明深度学习能在生物学有意义空间泛化后,于2021年共同创立了Isomorphic。

他指出制药行业新药批准数量长期持平,尽管论文数量暴增、分子模态增多,但成本从数亿美元升至15亿美元、周期从8-9年拉长到15年。原因在于整个行业仍高度依赖经验过程:提出假设、合成分子、在真实实验中测试,循环往复。生物学和分子设计都是超高维空间,试错只能触及低垂果实,复杂度再往上就难以触及。

AI的转折在于能把疾病生物学理解、分子设计全部搬到计算机上,建立达到实验精度的世界模型。一旦模拟足够准,现实世界就从“学习迭代”变成“验证”,失败率(尤其是临床毒性)会显著下降。动物模型对人类临床毒性的预测力大约只有30%,他更看好体外终点、器官芯片、类器官与AI模型的结合,直接建模人类毒性。数据策略上,他们主动生成专有数据以追求泛化,而不是依赖有偏的历史仓库。

公司内部管线聚焦肿瘤、免疫与炎症,同时与药企合作攻克对方十年未破的难题。模型本身是疾病、靶点、模态无关的。他强调:“如果五年后疾病生物学的预测理解没有发生巨大阶跃变化,我会非常失望。”更精准、更低毒性的分子还将打开真正的多药联用可能,因为目前很多组合受限于药物相互作用和肝毒性。

患者安全始终第一,但风险偏好因适应症而异——晚期癌症与慢性炎症完全不同。随着设计理解加深,成功率与安全性统计上的差异有望改变监管对药物上市的看法。
The Takeaway: The biggest opportunity in drug discovery is moving from slow empirical trial-and-error in the real world to predictive computational models that reach experimental accuracy, so far more design happens in silico and success rates jump when molecules finally enter patients.

Max Jaderberg is President and a founding team member of Isomorphic Labs, the Alphabet-backed company whose stated goal is to solve all disease. He spent years at DeepMind working on early generative models, reinforcement learning and game-playing systems such as AlphaGo and AlphaStar. Once AlphaFold showed that deep learning could generalize in a meaningful biological space, the team spun out Isomorphic in 2021 to build a dedicated culture and move fast on the harder problem.

He notes that new drug approvals have stayed roughly flat for decades even as scientific papers exploded and new modalities proliferated. Costs rose from a few hundred million to $1.5 billion and timelines stretched from eight or nine years to fifteen. The root cause is that medicine-making remains largely empirical: form a hypothesis, make the molecule, test it in the physical world, repeat. Both biology and molecular design are combinatorially vast; trial-and-error simply cannot search the higher-complexity regions once the low-hanging fruit is gone.

The shift Isomorphic is driving is to pull the entire discovery loop onto computers. Build world models that achieve experimental accuracy so that the real world is used mainly for validation rather than iterative learning. Clinical failures, especially toxicity, are where most value is destroyed because problems only become visible in patients. Animal models predict human clinical toxicity only about 30 percent of the time. Better routes combine in-vitro endpoints, organoids and organs-on-chips with AI predictors aimed directly at human outcomes. On data, the team generates proprietary datasets deliberately rather than relying on biased historical repositories, and has already observed strong scaling laws.

Internally they focus on oncology, immunology and inflammation while also partnering with pharma on the hardest problems those companies have worked for a decade. The underlying models are disease-, target- and modality-agnostic. Jaderberg’s benchmark: “I’d be really disappointed if we fast-forward even five years’ time [and] the predictive understanding of disease biology has not step-changed in a massive way.” Safer, more precise molecules should also unlock true multi-drug combinations that today’s toxicity and interaction profiles make impractical.

Patient safety remains non-negotiable, yet risk appetite correctly varies by indication. As design understanding improves and success rates become statistically different, the regulatory conversation itself is expected to evolve.
查看原文 →

Aaron Levie:AI正在彻底改变我们解决难题的方式Aaron Levie on Why the Riemann Hypothesis Moment Matters Beyond Math

Box CEO Aaron Levie指出,围绕Riemann假设的突破讨论之所以重要,不只是因为数学本身,而是因为它预示了问题解决方式的根本变化。人类数学家擅长有目的地构建理论,但没有人脑能独自创建一条需要25000页现有基础设施、与目标没有明显联系的“虫洞”。AI与人类的协作正在把这种能力变成可能。

过去,解决代码、生命科学、数学或工程问题最终受限于个人或团队能理解和连接的信息量。更多人或更多信息常常帮不上忙,因为协调和关联成本会上升。AI第一次把人类的上下文限制与我们发现和创造的能力解耦。企业里的每一份历史决策、每一篇无关论文、每一个过往突破,都可以被用来解决全新问题。解决方案可能不再那么优雅,但数量会多得多。我们才刚刚开始理解这意味着什么。
Box CEO Aaron Levie argues that the discussion around a Riemann Hypothesis breakthrough matters far beyond pure math. It signals a qualitative shift in how hard problems get solved. The best human mathematicians know how to build theory painstakingly toward a goal. No human brain on its own can create a “wormhole” that requires a 25,000-page tour through existing infrastructure with no obvious connection to the target. Collaboration between humans and AI is making that newly possible.

Until now, progress on any complex problem—code, life sciences, math, cyber, engineering—was ultimately bounded by how much information an individual or team could retain and connect. Adding more people or more data often failed because coordination and synthesis costs rose. AI decouples human context limits from our ability to drive discoveries. Every piece of prior art, past research or unrelated breakthrough can now be brought to bear on a new problem. Solutions may be less elegant than before, but there will be far more of them. We are only in the first months of understanding the implications.
查看原文 →

Amjad Masad:高中生用Agent Swarm重编译GTA6,开源新模式?Amjad Masad: High-School Kid Recompiling GTA6 with Agent Swarms

Replit CEO Amjad Masad关注到一名高中生正在把GTA6重新编译到PC上,更有趣的是他们用GitHub Issues来协调多个agent swarm的任务分工。他问这是否可能成为一种新的开源协作模式。同时他宣布由Replit Missions构建的Amuse已经开源。
Replit CEO Amjad Masad highlighted a high-school student recompiling GTA6 to run on PC, noting that the more interesting part is the use of GitHub Issues to coordinate tasks across agent swarms. He asked whether this points to a new model for open-source collaboration. He also announced that Amuse, built by Replit Missions, is now open source.
查看原文 →查看原文 →

Thariq:与模型协作本身就是一项深度技术技能Thariq: Collaborating with Models Is a Deeply Technical Skill

前Claude Code成员Thariq警告,我们太快把自己从流程中抽离,这对人类和模型都是损失。随着模型变强,与它们高效协作的技能只会越来越重要,目前还看不到收益递减。即使未来模型能一键证明Riemann假设或把JavaScript变成静态语言,推动前沿的人机协作本身仍是高度技术性的工作。我们还有大量问题需要解决。
Former Claude Code contributor Thariq cautioned that we are too quick to disintermediate ourselves, doing a disservice to both humans and models. As models grow more powerful, the skill of working with them becomes more important, and he has not yet seen diminishing returns on that skill. Even if future models one-shot “prove the Riemann hypothesis” or “make JavaScript a static language,” collaborating with models to push the frontier remains a deeply technical skill in its own right. There are still many problems left to solve.
查看原文 →查看原文 →查看原文 →

🛠️ 开发者工具与技巧

ChatGPT / Codex桌面与移动端更新反馈征集ChatGPT/Codex Desktop and Mobile App Updates

OpenAI Codex与ChatGPT团队成员Thibault Sottiaux在昨日桌面应用大更新后立即征集反馈,询问哪些地方仍需改进,并透露团队已在准备下一轮大更新。Peter Yang补充了实用技巧:新版ChatGPT移动应用非常干净,如果想念Codex线程,可以到最右侧标签长按,把Codex固定到底部导航——这个入口并不明显。他还建议左侧导航的聊天列表应同时显示本地最近的Codex线程。
OpenAI Codex and ChatGPT team member Thibault Sottiaux asked for rapid feedback after yesterday’s desktop app update, wanting to know what still needs attention, while noting the team is already preparing another major release. Peter Yang shared a practical tip: the new ChatGPT mobile app is very clean; if you miss your Codex threads, go to the rightmost tab and long-press to pin Codex to the bottom navigation—an action that is not obvious. He also suggested that the left-nav chat list should include local recent Codex threads, not only regular chat threads.
查看原文 →查看原文 →查看原文 →

Garry Tan:工作自动化已变得极其容易Garry Tan: Automating Your Work Is Now Extremely Easy

Y Combinator总裁兼CEO Garry Tan表示,用AI自动化自己的工作已经变得非常简单。他分享了自己在软件工厂中使用某工具的体验,称这是目前桌面、移动和云端综合体验最好的一次。同时他预测,鉴于AI能力和对GTA VI的普遍吐槽,一个由爱好者用vibe coding做出的、去掉所有被诟病设计的PC移植版,可能只需一两个月就能出现。
Y Combinator President and CEO Garry Tan observed that automating one’s own work with AI has become extremely easy. He shared that a particular tool has given him the best experience so far across desktop, mobile and cloud for roughly half of his software-factory work. Separately he predicted that, given AI capabilities and widespread frustration with GTA VI design choices, a hobbyist vibe-coded PC port that removes the hated flaws could realistically appear in a month or two.
查看原文 →查看原文 →查看原文 →

🌍 其他动态

Swyx质疑提供商秘密回传数据Swyx Questions Secret Phone-Home Behavior

AI工程师与播客主持人Swyx转发相关讨论并追问:这些提供商秘密回传数据是否存在任何善意解释?他的态度是“wtf”。
AI engineer and podcast host Swyx quoted a discussion about providers secretly phoning home and asked whether any benign explanation is possible, summing up his reaction as “wtf.”
查看原文 →

Dan Shipper:基准测试上升并不总意味着你以为的那样Dan Shipper on Benchmarks and Bad Tech Takes

Every CEO Dan Shipper指出,基准测试必须选择测量什么、保持什么不变,数字上升并不总是表面看起来的含义。他推荐通过观察真正AI原生团队如何把模型能力落地,来获得更真实的窗口。同时他将某个预测列入“史上最差技术判断”名人堂。
Every CEO Dan Shipper noted that benchmarks must choose what to measure and what to hold constant; rising numbers do not always mean what people assume. He pointed to detailed looks at how AI-native teams actually put model capabilities to work as a better window. He also filed one recent prediction in the hall of fame of bad technology takes.
查看原文 →查看原文 →

其他简短动态Other Brief Notes

Nan Yu(OpenAI Codex产品)简单写道:“Forget second brain. Third brain.” Guillermo Rauch(Vercel CEO)呼吁需要一个类似平台来记录AI壮举。Zara Zhang强调“Build for yourself”。Matt Turck调侃人们在完全不懂Riemann假设的情况下兴奋发推,并分享了Slush大会的观察:学生组织、视觉风格独特、非科技中心的会议反而让人更专注。
Nan Yu (product, Codex at OpenAI) simply wrote “Forget second brain. Third brain.” Vercel CEO Guillermo Rauch called for an equivalent of a well-known platform dedicated to AI feats. Zara Zhang reiterated “Build for yourself.” Matt Turck joked about the art of excitedly tweeting that AI solved the Riemann Hypothesis while having no idea what it is, and shared observations from Slush: student-organized, visually striking, and non-tech-hub conferences work well because attendees stay put.
查看原文 →查看原文 →查看原文 →查看原文 →查看原文 →