Google ERA：把科学问题映射成可打分任务 — John Platt

Alessio

今天请来 John Platt——Google Fellow、Google Research 应用科学负责人。他自称「十亿级书呆」（giga nerd），什么都感兴趣。纠错欢迎：14 岁上大学，18 岁进 Caltech 读博，联合导师包括刚拿诺贝尔奖的 John Hopfield。

Pleasure to have John Platt with us today—Google Fellow and head of applied science at Google Research. He described himself as a giga nerd, excited about absolutely everything. Correct me if wrong: you started college at 14, PhD at 18 at Caltech, co-advised by John Hopfield, who just won a Nobel Prize a couple of years ago.

John Platt

对，是 Hopfield。

Oh yeah, yeah.

Alessio

教科书级算法：Platt scaling、SVM 训练的 SMO（scikit-learn 里仍在）；发现并命名两颗小行星；2006 年拿奥斯卡技术奖——看过 Pixar 就见过他的算法；Erdős–Bacon 数两边各三。跳过二十年职业生涯：在 Google 科学线做过聚变、量子计算、气候建模等——大致对吗？

Textbook algorithms—Platt scaling, sequential minimal optimization for SVMs, still in sklearn. Discovered and named two asteroids; Academy Award for technical achievement in 2006—if you’ve watched Pixar, you’ve seen his work. Erdős–Bacon number three and three. Skipping ~20 years of career: at Google Sciences, fusion, quantum computing, climate modeling—more or less right?

John Platt

对。还做过很多应用数学和信号处理之类有趣的事。维基上还有专利与 iPod 的故事——欢迎。

That’s right. I’ve also done lots of applied math and signal processing and all sorts of fun things. Wikipedia has a fun story about patents and the iPod. Thank you for having me.

Host

ERA 怎么读、和其他 Google 内外半相关项目比，有什么特别？

Can you tell us about ERA—how the acronym is pronounced—and what makes it special versus other semi-related work inside and outside Google?

John Platt

Google Research 做 AI for Science 已逾十年。约十年前多是「经典」机器学习——卷积网等，针对具体科学问题建专用模型。约两年前我们被通用大模型兴奋住了，想弄清能做什么。我们撞进一种映射：许多科学问题可写成「可打分任务」——希望有一段代码最大化某个分数；映射进这框架后，意外地能推进很多问题。

We’ve done AI for science in Google Research for more than 10 years. Around 10 years ago it was classical ML—conv nets and such—specific models for specific science problems. About two years ago we got excited about general LLMs and wondered what could be done. We stumbled into a mapping: many scientific problems can be phrased as scorable tasks—I want a piece of code that maximizes some score. Surprising how much progress you can make by mapping into that framework.

John Platt

科学家大量时间在建统计或物理模型。统计模型的打分就是拟合数据集；机器学习其实是可打分任务的子集。ERA 论文第一作者 Michael Brenner 很擅长——现在能一晚上用工具敲出一篇科学论文。例如渐近展开：问 ODE/PDE 在某参数 ε→0 时的行为——可变成经验任务：提出渐近正确的解，在 ε=10⁻⁴ 等处检查拟合，同时让底层 Gemini 做数学推理并最大化拟合。关键不是随机突变，而是本身聪明、有世界先验的 AI。

Scientists spend a lot of time building statistical or physically based models. For a statistical model the score is fitting the dataset—ML is a subset of scorable tasks. Lead author Michael Brenner is skilled—he likes to knock out a scientific paper in an evening with the tool. Asymptotic expansions: how does an ODE/PDE behave as epsilon goes to zero—turn it into an empirical task: propose asymptotically correct solutions, check fit at epsilon equals 1e-4, and ask Gemini to do mathematical reasoning while maximizing fit. The thing altering the code isn’t random—it’s an AI with huge prior knowledge. That’s the trick.

John Platt

非统计例子：遥感里重访频率、空间分辨率与光谱分辨率互相权衡。理想是每五分钟、高光谱、10 厘米——做不到。测大气 CO₂ 时，OCO-2/OCO-3（后者挂在国际空间站）给出准确、较高分辨率的窄带；GOES 等气象卫星约每五分钟一张红外图，但像元大、并非为 CO₂ 设计。于是用一边估计另一边，再塞进天气、长期反照率等——做成类似知情超分辨率的跨卫星模型。

A non-statistical example: remote sensing always trades revisit frequency, spatial resolution, and spectral resolution. Ideally constant monitoring every five minutes at hyperspectral, 10 cm—you can’t get that. For atmospheric CO₂, OCO-2 or OCO-3 on the ISS give accurate, fairly high-resolution strips; GOES has infrared every ~5 minutes but large pixels and wasn’t designed for CO₂. Ask one to estimate the other, shovel in weather and long-term albedo—a nice informed super-resolution from one satellite to another.

John Platt

产品形态是：你开口聊问题——映射往往不显然，虽 Brenner 这类专家会。我们写了个代理帮你定义可打分任务；它先像中间产物一样产出带打分函数的 Python notebook，再用 Gemini 聪明地突变 notebook，持续提出能抬高分数的代码。

In the product you just start talking—mapping is often non-obvious though experts like Brenner know how. We wrote an agent that helps define your scorable task. It produces a Python notebook with a scoring function, then mutates that notebook cleverly—Gemini proposing code to maximize the score.

Host

这和一般能优化 notebook 的智能体系统差在哪？

What’s different from a general agentic system that optimizes notebooks?

John Platt

用 2026 的话说，它是专用 harness，跑的算法对 ERA 是蒙特卡洛树搜索。维护成百上千份 notebook 候选；选一个，问 Gemini「怎么改更好」，生出新版本、测试，放回候选池。候选池是树状的。选择用强化学习里常见的 UCB（上置信界）——乐观估计突变后的约 95 分位结果，不一定总挑当前最好，有时是第五名；也试过把两个候选的想法砸成第三个。

In 2026 parlance it’s a specialized harness running Monte Carlo tree search for ERA. It keeps hundreds or thousands of notebook instances, selects one, asks Gemini what would make it better, makes a new one, tests it, puts it back. The pool is tree-structured. Selection uses UCB—optimistic, estimating roughly the 95th-percentile outcome of mutation. It doesn’t always pick the best notebook—sometimes the fifth best. We’ve also recombined ideas from two candidates into a third.

John Platt

初始候选：底下 Gemini 会写代码。你给问题的文本描述，有时塞五篇相关论文——它会去读，先试一版；可能有 bug、分数负无穷，再突变改进。可以给 starter code，但不必。每轮默认并行长约 10 片叶子——太多并行就学不到彼此：十路同时搜时，1 号看不到 2–10 在干什么；上千路则大量算力却少交叉学习。一批结束后共享历史（要修剪以免撑爆上下文），像共享上下文一起想，而不是一千条完全独立分支——确实在推 Gemini 的长上下文。

Seeding: Gemini is good at writing code from a textual problem description—optionally five papers it reads for a first stab. Bugs may return minus infinity; then it mutates. Starter code optional. Default about 10 parallel leaves—trade-off: too much parallel work and you can’t learn across attempts; with 10 searches, number one can’t see two through ten. After a batch you get pruned history of thinking and results—essentially one shared context, not a thousand independent branches. You’re pushing Gemini’s long-context abilities with careful management.

swyx

首要目标往往是定对要优化的分数——我未必总信代理独自完成，那更像人机协作。

The first key goal is identifying the score to optimize—I’m not sure I’d always trust an agent for that part; it seems more the human task in the loop.

John Platt

对，而且常要小心。你或代理一起写出打分函数后，迭代常会找到作弊漏洞——「我不是那个意思」——于是再绕一圈改指令。旧日（比如 2024）研究生写科学软件太费劲，只能试几种相关方案就停去写论文；这东西底下几乎不停试。使用者几乎把时间花在科学创造性层面：代价函数到底意味着什么——那才是科学问题的本质，而不是 CSV、数据库那些脏活。代理还能建议：「要不要 join 这个数据集？」

It is, and you have to be careful. You and the agent come up with a scoring function; iteration finds a way to cheat—“I didn’t mean that”—so you iterate instructions. In the old days, say 2024, writing scientific software took so much effort you tried a few related things then stopped for the paper. This is relentless underneath. Users now spend time almost at scientific creativity—what does the cost function mean?—the essence of the problem, not CSV imports and databases. The agent can even suggest datasets to join.

Host

有奖励黑客的搞笑翻车故事吗？

Any fun stories where agents comedically ran off the rails on reward hacking?

John Platt

细节一时想不起来，但确实会让人吃惊。像精灵许愿——小心你许的愿。另一方面 Gemini 知道很多；尤其指向论文时，会做一些相对正经的事。七十年代以来就有人爱进化式改代码，但随机在代码空间突变几乎没用——大多有害。这里底下循环本身是 AI，知道有趣的「梯度」方向，所以才管用；仍可能过拟合、有精灵问题，但也有一定 sanity。

I’m blanking on concrete comedy, but you get surprised. Like the old genie stories—careful what you wish for. On the other hand Gemini knows a lot; pointed at papers it does relatively sane things. Evolutionary coding has been loved since the ’70s, but random mutation in code space is nearly worthless—most changes harmful. Here the inner loop is an AI that knows interesting gradients to try—that’s why it works. It can overfit and have genie problems, but also some sanity from world knowledge.

Host

论文时代是 Gemini 2.5；内部有没有用起来几乎相位跃迁的感觉？

The paper used Gemini 2.5—have you seen almost a phase transition in how effective this tooling became over the last year or two?

John Platt

惊人。半个大版本都在变。在 Gemini 2.0 下我觉得这几乎不可能做成——我们试了很久不工作，然后突然开始工作，现在很惊人。大版本进步令人震惊。我还会跟科学家说：你一年前试过 2.0 不喜欢？那是永恒以前了。有篇预印本把 ERA 和 Anti-Gravity 结合——那个 harness 也很强，能拉很多论文写很多代码；Anti-Gravity 公开，ERA+Anti-Gravity 还没有。

Amazing—even half-versions. It would have been impossible under Gemini 2.0. We experimented for a while; things weren’t working, then started to work, now they’re amazing. Major-version progress has been stunning. I tell scientists: you tried 2.0 a year ago and didn’t like it? That was an eternity ago. One preprint combines ERA with Anti-Gravity—that harness pulls in lots of papers and writes lots of code. Anti-Gravity is public; ERA plus Anti-Gravity not yet.

swyx

代码突变从计算机科学早期就有；经典风险是过拟合与多重假设检验——状态空间爆炸。你怎么看？我甚至把开源版 ERA 接到 Claude 上跑着。

Code mutation has been around since the dawn of CS. The canonical risk is overfitting or multiple-hypothesis testing—an explosion of state space you’re exploring. Thoughts? Empirically I strapped an open-source ERA-like setup into Claude; it’s running now.

John Platt

问题里嵌了两层。谈多重假设时，要分预测模型与描述模型——做机器学习与统计久了会敏感。预测模型：有输入输出，写代码把某数据集误差压到最低——统计模型。描述模型才是科学要逼近的：能外推，因为里面抓住了物理或对现实的描述。牛顿想的是苹果与引力，但引力不是关于苹果；若 17 世纪「机器学习」只拟合「苹果会掉」，没有行星数据就不知道行星会怎样。

There are two questions embedded. With multiple-hypothesis testing there are predictive and descriptive models—you can tell I’ve done ML and statistics a long time. A predictive model: inputs and outputs, code with lowest error on a dataset—just a statistical model. A descriptive model is what science wants: it should extrapolate because it captures physics or a description of reality. Newton thought of apples and gravity, but gravity isn’t about apples. A 17th-century ML model—“apples fall”—with no planet data wouldn’t know what planets do.

John Platt

物理学家用直觉——其实是一堆已知事实——保证模型与已知一致。ERA 里目前主要是「LLM 直觉」：尤其指向已有论文时，会偏向建「像样」的模型；也可显式要求并入某些约束。指向论文时特别擅长改编文中方法，甚至常能复现/逆向工程论文——Brenner 爱干这个。有次建议他看一位 MIT 教授屋顶固定面积最大化日采光的设计；我相信他甚至没装仿真器，代码代理直接从论文复现了代码。

A physicist uses intuition—really a pile of known facts—to keep models consistent with what’s known. In ERA today it’s essentially LLM intuition: pointed at papers it builds saner models; you can bias toward certain choices. Given papers it adapts methods amazingly—often recreate or reverse-engineer a paper; Brenner likes that. I suggested an MIT professor’s rooftop solar-maximization design widget; I believe he didn’t install the simulator—the coding agent reproduced the code from the paper.

John Platt

若你完全不知道某种物理形式，它未必从零发现全新物理；但若告诉已知约束，会演化进去。人类未来一两年仍有空间：Co-scientist 可帮假设生成，我在地球化学问题上用过、学到岩浆里一种离子——但创造、哲学与严谨仍需人。我信 jagged frontier：编码、蒐集知识、找相关工作已有大刺；严谨、哲学、创造仍弱。暂时仍需要人类。

It won’t necessarily discover completely new physical models from scratch if you didn’t know something existed—but given known constraints it will evolve them in. Totally room for humans: Co-scientist helps hypothesis generation—I used it on a geochemistry problem and learned about an ion in magma—but creativity, philosophy, and careful rigor still need humans. I’m a believer in the jagged frontier: big spikes in coding and gathering related knowledge; less so in rigor, philosophy, creativity. So far you need the humans.

John Platt

多重假设检验：当你声称描述世界的模型，拿十亿支飞镖扔数据，就要管虚假发现率。ERA 目前更像强力工具帮你做新科学，而非自己发现全新物理——科学家仍要保证说法是描述性的。至于普通过拟合：要极其小心，像电锯能切手——必须更严谨，藏好绝不偷看的 hold-out。

Multiple-hypothesis testing: when you claim a descriptive model and throw a billion darts at data, watch false discovery rate. We haven’t made a system that discovers completely new physics—this is a power tool to help scientists do that; the scientist ensures it’s descriptive. Plain overfitting: be very careful—it can slice your fingers off. Be excruciatingly rigorous about hidden hold-out sets you don’t look at.

John Platt

品味与严谨都要——软件演进里，有人锚定创造、有人锚定「别崩、能扩展、别错」；不必同一人，但两边都要。讽刺的是 ERA 起源于人们未必知道的事实：Kaggle 是 Google 的一部分，项目曾叫 Auto-Kaggle——做出能赢竞赛的系统，所以形状像排行榜。Kaggle 上过拟合猖獗；我们办过尾迹相关竞赛，有人赢我们，是因为发现标注有半个像素误差——旋转增广时 0,0 在像素中心还是左下——Goodhart：一旦指标成目标就不再是好指标。AI 领域不断耗尽排行榜，也是同一律。

You need both taste and rigor—as software evolves, some people anchor on creativity, others on “doesn’t crash, scales, isn’t wrong.” They needn’t be the same person. Ironically ERA started because Kaggle is part of Google—the Auto-Kaggle problem: win competitions—that’s why it has that shape. Overfitting on Kaggle is rampant. We had a contrail competition; people beat us by finding a half-pixel label error—center versus lower-left of the pixel when rotating synthetic data. Goodhart’s law: any metric that becomes a target ceases to be a good metric. AI keeps exhausting leaderboards for the same reason.

John Platt

在 playground 竞赛上 ERA 很好；也参加过很多。骄傲的一例：CDC 预测美国各州下周一周 COVID/流感病例——ERA 表现非常好。有的竞赛差一口气，因为挤最后 0.001 需要大量 TLC，没人去刮最后几十分名。外环很有人：跑几小时回来给你例子，你塞论文、指出它在做坏事——像不眠的超级积极研究生。设计就是人在环；和许多「一枪」工具不同。算力上：内部调 Gemini，token 预算我不详；但打分本身可能含蒙特卡洛或仿真，CPU/GPU 也能烧很多——它太执着，比研究生更快撞上预算墙。

It did very well on playground competitions. One we’re proud of: CDC competition predicting next week’s COVID and flu cases per US state/territory—ERA did super well. Elsewhere we got close but didn’t shave the last 0.001—that needs TLC. The outer loop is the fun creative part: it runs a few hours, comes back; you say look at this paper, you’re doing something bad—like a hyper-eager grad student who doesn’t sleep. Designed human-in-the-loop, not one-shot. Token budget via internal Gemini I don’t actually know; scoring may include Monte Carlo or simulators—fair CPU/GPU. Being relentless, it hits the budget wall faster than a grad student.

Host

你们优化过尾迹——能讲讲吗？

One of the things you optimized was contrails—can you talk about that?

John Platt

尾迹，不是阴谋论里的 chemtrails。喷气机后的白云是凝结尾迹。估算上，尾迹约占全部人为全球变暖约 1%。有时划过即散，影响小；持久尾迹可铺成「华夫饼」似的天空。薄白云白天反光，但地球约 300 K 黑体辐射在约 10 μm 远红外，尾迹那里反照率极低、近乎黑，吸收并再辐射、像毯子困热——全天候发生，故净增温。欧洲等航线密集处，额外尾迹卷云可盖住百分之几天空，局地强迫约 1 W/m²；全球人为平均约 3 W/m²，故高交通区局地很可观。

Contrails, not chemtrails. Those white clouds behind jets are condensation trails. Estimates say about 1% of all anthropogenic global warming is from contrails. Some streak and fade; persistent ones can waffle the sky. Thin white clouds reflect sunlight by day, but Earth at ~300 K emits far-IR around 10 microns where contrails have very low albedo—nearly black—so they absorb outgoing IR and re-emit, trapping heat like a blanket 24/7—hence net warming. Over Europe, extra contrail cirrus can cover a few percent of sky, adding ~1 W/m² locally; all anthropogenic warming averages ~3 W/m² globally—so in high-traffic places, a lot locally.

John Platt

成因是大气中冰过饱和区，像冰糖结晶：区域呈薄饼状，仅数百米厚。飞过时尾气湿气成滴再冻；在坏区里每克水/冰/烟尘可「吸」出约 10 kg 水——约一万比一。这些区不可见；告诉飞机从下方绕过，通常只需降约两个飞行高度层，多耗一点油但不多。我们用卫星连续监测尾迹，并因数值天气模式不够准，用卷积/U-Net 等定制模型预测冰过饱和位置，交给飞行规划软件低成本躲闪。这些区可由对流层顶附近暖湿空气注入，可持续很久甚至数天，但仍需持续预测。

They’re caused by ice-supersaturated regions—like rock candy—pancake-shaped, only a few hundred meters tall. Fly through and exhaust moisture freezes; in a bad region, ~10 kg of water per gram of water/ice/soot—roughly 10,000:1. Regions are invisible; tell the plane to go underneath—drop about two flight levels; costs a little fuel. We monitor satellites for contrails and, because weather models aren’t accurate enough, build a custom conv net/U-Net to predict ice supersaturation for flight-planning software. Regions persist—possibly days—from warm moist air near the tropopause, but you keep predicting where they are.

John Platt

还想知道单条尾迹贡献了多少增温——反事实：没有这条尾迹时红外会怎样？测长波射出辐射的反事实模型还行，反射阳光更难，我们卡了约两年。ERA 搜遍混杂因素，在人造注入尾迹的测试集上——我们自己的尝试甚至过不了自测——ERA 找到了管用的简单模型：若干我们没试过的混杂组合，事后看很显然。长波已有论文；反射阳光那篇在 EGU 讲过、尚未投。

You also want how much warming a contrail added—a counterfactual. We had an okay counterfactual for outgoing longwave radiation but were stuck ~2 years on reflected sunlight. ERA searched confounders; on artificial injected-contrail tests our own attempts failed our tests, but ERA found a simple model with a combination of confounders we hadn’t tried—obvious in retrospect. We have a paper on longwave; the sunlight one was talked about at EGU, not submitted yet.

John Platt

CO₂ 有趣，因我们其实不太清楚生物圈碳通量。海洋、陆地都在吸，但误差条不小；到 2100 更大——模型对升温与 CO₂ 下生物圈反应不确定，仅吸收不确定就可差约 300 ppm；而今约 440–450 ppm，不确定性巨大。浓度估计是朝减少该不确定迈出的一步。

Estimating atmospheric CO₂ is interesting because we don’t really know carbon flux in and out of the biosphere. Ocean and land absorb a lot, but error bars are moderately large—and huge by 2100; uncertainty in how the biosphere reacts can be ~300 ppm of CO₂ versus ~440–450 ppm today. Concentration estimates are one step toward shrinking that.

John Platt

人们常把气候与天气糊在一起——大气物理相同，但完整地球系统模型复杂得多。天气大约到 15 天：大气混沌，蝴蝶效应，两三周后轨迹全不同。天气预测近些年有巨大进步，多靠 2018 前后那代机器学习加大数据大算力，不全是新 LLM；Google 做了很多新天气模型，热带气旋路径可提前多天更准——飓风是把洋面热转成大尺度运动的热机，经典模式常漏掉增强。气候难得多：你不预测 2070 西雅图是否下雨，而要平均态；且非平稳——植物、冰行为都在变。没有三十五十年后的标签，又不能等——多重假设与描述模型问题在气候极尖锐。AI 尚未革命气候，正因低数据、外推残酷。

People collapse climate and weather—same atmospheric physics, but a full Earth-system climate model is far bigger. Weather is up to ~15 days; the atmosphere is chaotic—butterfly effect, totally different in 2–3 weeks. Weather has had a huge step change from ~2018-era ML, lots of data and compute—not even new LLMs. Google built new weather models; we can predict tropical-cyclone tracks more accurately many days ahead—hurricanes are heat engines converting ocean heat to motion; classic models often miss intensification. Climate is harder: not rain in Seattle in 2070, but averages; non-stationary—plants and ice behave differently. No data from 30–50 years out; you can’t wait. Multiple-hypothesis and descriptive-model issues are incredibly severe. AI hasn’t revolutionized climate yet—resistant low-data problem.

John Platt

Lorenz 的简单 ODE：天气是吸引子上你在哪，气候是吸引子的统计；我们还在改变吸引子本身，还有 tipping points。跑仿真时，不稳定是模型糟还是真物理失稳，极难分辨。过程模型把气候拆成上千片，每片找专家拟合——例如云中冰微物理决定尾迹能活多久、下落快慢取决于未知冰晶形状与混合，而冰微物理对气候模式含义很强。希望是：未来工具不仅拟合数据，还能读远超个人的论文，把已审慎评估的知识与数据合成——今天还没有，但我已在 ERA 里看到一点苗头。LLM+搜索这条路也像 ERA；Google 十蓝链在 LLM 之前就是一种「AI」。对生物与气候，我们更信可拆解、可检查的拼图，而非不可检的巨型黑箱——黑箱要外推到极端与黑天鹅很难，自动驾驶难也在边角案例。

Lorenz: weather is where you are on the attractor; climate is the attractor’s statistics—and we’re moving the attractor, with tipping points. In a simulator, is instability bad modeling or real physics? Extremely hard to tell. Process models boil climate into a thousand pieces with expert fits—ice microphysics in clouds governs how long contrails last; crystal shape and mixing are poorly known, yet matter hugely for climate models. Hope: a future tool that fits data and reads far more papers than any person, synthesizing carefully evaluated knowledge—we don’t have that today, but I see little indications inside ERA. Combining LLMs with search is amazingly like ERA; Google’s ten blue links were a form of AI before LLMs. For biology and climate we wouldn’t trust a giant black box we can’t examine—extrapolation to extremes and black swans is hard; self-driving is all corner cases.

Host

物理里模型常是几条方程；AlphaFold 迫使人们重估「什么是科学」——极强却未必给物理学家想要的直觉。Box：所有模型都错，有些有用。何时数据驱动够用，何时要可解释过程模型？

In physics a model was often a few equations. AlphaFold forced people to re-evaluate what science is—incredibly powerful tools that often don’t give intuition physicists wanted. Box: all models are wrong, some are useful. When are data-driven models enough versus interpretable process models?

John Platt

几乎像天气对气候。数据丰富——天气，或蛋白质因 PDB——可安心用统计模型如 AlphaFold；他们把 PDB 跑遍并公开，非常酷。气候开放、非平稳、大外推就要谨慎。生物很多题即使数据多，也极抗「简单基线之上的任何东西」——我总说先拟合线性回归，或 SVM。闭环、数据盖得住就信统计；否则要过程约束。我团队关心干预：尾迹妙在成本相对低，且偏局地——一国治理本国上空主要改善自己。增温大致全时全球，冷却只在太阳角度合适时；两 sigma 下很少「确定在冷却所以我想要更多」的尾迹——极地夏季可能，但南极几乎无航班。气候没有银弹，是「悲伤饼图」；聚变若够便宜或能干掉一大块，但还不知道成不成。

It boils down almost to weather versus climate. In a data-rich regime—weather, or proteins via PDB—you’re comfortable with statistical models like AlphaFold; running it exhaustively on PDB and publishing was amazing. Climate is open, non-stationary, big extrapolations—be cautious. In biology many problems resist anything beyond simple baselines even with lots of data—I say always fit linear regression, or SVMs. Closed problems data fully covers: statistical OK; otherwise process constraints. My team cares about interventions: contrails are relatively low-cost and local—a country fixing its own mostly helps itself. Warming is roughly constant/global; cooling only at good sun angles; few contrails where we’re sure at two sigma they’re cooling—polar summer, but almost no Antarctic flights. No silver bullet—pie chart of sadness. Fusion might knock out a lot if cheap enough; we don’t know yet if it’ll work.

Host

玩笑总是聚变还有三十年；现在是不是不到三十年？

The joke was fusion is always 30 years away—is it actually less than 30 now?

John Platt

有相当概率本十年末就有人做出商业相关聚变——更像三年量级，不是三十年。DeepMind 等在做托卡马克控制系统，防 disruption：能量打成一束砸真空室，很惨。我们曾与 TAE 等合作推荐实验；场反向位形（FRC）是自洽「橄榄球」等离子体，内外磁场相反，有分界面；MHD 假设下理论不稳、实践可稳，Helion 等也在用，相对皮实，但仍可能打穿真空室。

There’s a definite probability someone makes commercially relevant fusion by end of this decade—more like three years away, not thirty. Google DeepMind has worked on tokamak control so they don’t disrupt—all energy into one beam hitting the vacuum chamber; very sad. We worked with TAE recommending experiments. Field-reversed configurations are self-contained football plasmas with opposite inside/outside fields and a separatrix—unstable under MHD assumptions but stable in practice; Helion and others use FRCs; relatively robust, but discharges can still punch vacuum chambers.

John Platt

聚变归结为 Lawson 判据：密度 × 温度 × 约束时间（能量按 1/e 衰减时间）须大于某常数。乘积解释了难：每种路线总有一个数不够大。新闻常只吹其中一个数——要三个都够才行。计算因路线而异：托卡马克主要是控制系统；FRC 有 Z 不稳定性，可用 PID 把「橄榄球」稳住。能源干预必须和经济现有方案竞争；有人甚至炒过往堆里注汞、用中子通量嬗变成金——投机、未必可行。若能电气化一切，用电量或增约五倍；可再生加电池盖最后 10–20% 会极贵，需要基荷——聚变可能是一块，但世界与全球经济太复杂，没有万能干预。

Fusion boils down to the Lawson criterion: density × temperature × confinement time (1/e energy decay) must exceed a constant. The product explains hardness—every approach has an Achilles heel on one number. Breathless news often cites one number; you need all three. Compute depends on approach: tokamaks are largely control systems; FRCs have a Z instability a PID can keep centered. Energy interventions must compete economically; someone floated injecting mercury to transmute into gold via neutron flux—clever, might not work. Electrify everything and electricity might grow ~5×; renewables plus batteries get ever more expensive covering the last 10–20%—need baseload; fusion might help. No silver bullet; the global economy is complicated.

Host

LA 大片烧过；除减缓外，有没有韧性、风险与早期干预？

Large parts of LA burned. Beyond fixing the underlying problem—risk assessment, hardening, preparation for what’s coming?

John Platt

Google 有气候危机韧性方向。FireSat：野火若房间大小时尚易扑灭，到一英亩难很多，且可指数长大。若有全球低轨星座用中波红外（火的黑体温度）探测，大约 50–80 颗，可在行星任意处约 15–20 分钟内发现约 5 米边长量级火点，再决定扑灭或受控烧燃料。与 Earth Fire Alliance 等合作，Muon Space 造星，已发射原型。也做野火边界检测，经 Android/搜索告知公众；与美国林务局做传播模型——Rothermell 七十年代过程模型的神经网络代理以便极快运行。WHO 估计全球野火烟尘每年约 30 万超额死亡。气候像重病：症状（适应）与病灶（减缓）都要治；天气与气旋预测也挂在韧性上。

Google has a big climate-crisis-resilience effort. FireSat: a room-sized wildfire is easy to put out; an acre is much harder, and they can grow exponentially. A global LEO constellation detecting mid-wave IR—fire blackbody temperature—roughly 50–80 satellites could find ~5 m fires anywhere within ~15–20 minutes, then intervene or let safe fuel burn. With Earth Fire Alliance and Muon Space we’ve launched a prototype. We also do wildfire boundary detection via Android/Search, and neural proxies of Rothermell-style USFS spread models to run very fast. WHO estimates ~300,000 excess deaths/year from wildfire smoke worldwide. Climate is like a serious disease—treat symptoms and the underlying cause; weather and cyclone prediction hang together with resilience.

John Platt

中波红外需要制冷，不是随便搭微服务的超小载荷；星不算巨大。分辨率硬件约 50×50 m，靠多光谱与超分等 AI 叠到约 5×5 m；还有扫帚式光谱展开，不是拍一张拍立得那么简单。

Mid-wave IR needs refrigeration—not tiny hitchhikers. Satellites aren’t huge monsters. Hardware resolution ~50×50 m; multispectral super-resolution and AI reach ~5×5 m. It’s a broom sensor spreading the spectrum—not a simple Polaroid.

Host

AI×科学变化很快；年轻科学家该怎么走？

AI for science is evolving quickly—what should young scientists and practitioners do?

John Platt

过去 12–18 个月有相位变化：以前专模逐题求解，现在通用 AI，社区还在摸工具链。我 21 岁儿子同时钻 RNA 领域、又 vibe coding——我认为对：领域专长不会消失，品味从哪来若没有苦功仍是开放题；同时玩现有工具，因连「聪明老人」也在实验。实验是真值与瓶颈；lab-in-the-loop、通用「什么都能做」的实验室仍很开放。别怕，试狂野想法——我们带学生研究员时那最开心。

There’s been a phase change in the last 12–18 months—from specialized models per problem to general AI; the community is still feeling out the toolchain. My son just turned 21, into AI, coding, and chemistry—domain expert on RNA while vibe-coding. That’s right: domain expertise isn’t going away; how you get taste without grunt work is open. Play with tools—smart old people are experimenting too. Experiments are ground truth and a bottleneck; lab-in-the-loop and a general lab that does everything are very open. Don’t be scared—try wild crazy things; student researchers doing that are delightful.

John Platt

像登山：能开车上去，偶尔仍值得徒步练肌肉。八九十年代没有 sklearn，我自己写数值库、boosting 写过四门语言——现在才「懂」boosting。运动员有比赛也有训练；局部不用 LLM 看似低效，可能换长期能力。世界爱挤干一切，但我在组里仍强护 20% 时间——学东西、试疯狂想法甚至不必告诉我；超优化生产力本身也是一种过拟合。LLM 犯错方式与人不一样，不能当空心西装；Feynman：你是最好骗的人——对自己写的软件也要戳。 enduring 的是关于世界的领域（生物、物理）、数学、严谨核查，以及创造力。

Like hiking: you could drive up the mountain, but hiking builds muscles. In the ’80s–’90s I wrote my own numerical libraries and rewrote boosting in four languages—now I know boosting. Athletes have race days and training; hammering hard work may pay long-term even if locally less “productive.” The world wants to squeeze everything; in my group I protect 20% time—learn and try crazy things, you don’t even have to tell me. Hyper-optimizing productivity is overfitting. LLMs make weird non-human mistakes—don’t be an empty suit; Feynman: you’re the easiest person to fool—poke your own software too. Enduring: domains about the world, math, rigor/checking, and out-of-the-box creativity.

Host

你上过费曼的课？

You mentioned Feynman—you took a class from him?

John Platt

他与 Hopfield、Carver Mead 的「计算物理」课，约 1982，有 DARPA 资助。结构是周二嘉宾讲座，周四费曼站起来解释为什么全错——很有趣。还有「费曼效应」：他讲得你以为懂了，走出去发现没有。嘉宾包括 Danny Hillis 等，合在一起是当时对可逆计算、热力学极限的集体困惑；量子门模型那时基本还没谈。我是 Mead 的系统管理员，整组共享一台 VAX 11/750，约 1 MIPS，磁盘 80 MB 像洗碗机——相对今天算力约等于零。

His Physics of Computation class with Hopfield and Carver Mead—I took it in ‘82, DARPA-funded. Guest lecture Tuesday; Thursday Feynman explained why that was all wrong. The Feynman effect: you walk out thinking you understood, then realize you didn’t. Guests like Danny Hillis showed the mass confusion—reversible computing, thermodynamic limits. No gate model of quantum computing then. I was Carver Mead’s sysadmin; the group shared a VAX 11/750 at maybe 1 MIPS and an 80 MB dishwasher-sized disk—to rounding error, zero compute.

John Platt

八十年代 Hopfield 网络极热；NeurIPS 从 Snowbird 工作坊挤爆后独立出来，更早有 Caltech「Hopfests」。关联记忆挖下去，Transformer 某种意义上也是；甚至有论文标题玩梗 Hopfield networks are all you need——绕回来了。当时愿景远超能力，因几乎没有算力。神经网络吃矩阵乘，BLAS 被 GPU 优化，算法与硬件共进化——大脑未必用矩阵乘，但我们走了这条硬件彩票。Hinton 组约 2010 已用 GPU；约 2007–08 GPU 相对 CPU 还没那么快，真正甩开与 ImageNet/语音时代重合，并非巧合。朋友 Brian Catanzaro 在 NVIDIA 的故事是：一次与黄仁勋约 15 分钟谈话，说服公司认真对待 CUDA 上的深度学习。

In the early ’80s Hopfield nets were hugely exciting; NeurIPS spun out of a crashed Snowbird workshop, after Caltech Hopfests. Dig into transformers—they’re associative memory; there’s even “Hopfield networks are all you need”—full circle. Hopes outstripped capability because we effectively had zero compute. Nets ride on GEMM/BLAS optimized by GPUs—algorithms co-evolved with hardware; brains may not matrix-multiply, but that’s why we’re here. Hinton’s group used GPUs around 2010; in 2007–08 they weren’t much faster than CPUs—the leap coinciding with ImageNet/speech isn’t coincidence. A friend, Brian Catanzaro at NVIDIA, had a ~15-minute conversation with Jensen that pivoted the company toward deep learning on CUDA.

Host

怎么才能有机会给小行星命名？

How does one get to name an asteroid?

John Platt

Caltech 上 Gene 与 Carolyn Shoemaker 的行星科学课。八十年代尚未自动化：去帕洛马用快望远镜两张底片，回校用体视镜找「浮起」的移动点，测量显微镜相对已知星，寄给 Minor Planet Center 的 Brian Marsden，拼出终现后可获发现权命名。如今智利 Vera C. Rubin / Simonyi Survey Telescope 六周发现约 11,000 颗——可能有数百万颗小行星。业余爱好者用掩星：小行星挡恒星变暗，若再暗一次则有卫星；我命名的一颗（故事上 Carolyn 把她的一颗让我命名给我母亲）去年被发现有约 1 km 卫星，主体约 4 km——不小的双星系统。同班大概还有一人发现；我找到两颗很不寻常。

At Caltech, Gene and Carolyn Shoemaker taught planetary science. In the ’80s nothing was automated: Palomar film pairs, stereoscope for floaters, measuring microscope, send to Brian Marsden at the Minor Planet Center; final apparition gave naming rights. Now Vera C. Rubin Observatory’s Simonyi Survey Telescope found ~11,000 asteroids in six weeks—likely millions exist. Occultations: an asteroid dims a star; a second dip means a moon. One asteroid named after my mom—Carolyn kindly gave me one of hers after I waffled—had a ~1 km moon found last year; main body ~4 km—a fairly big binary. One other classmate discovered; finding two was unusual.

Host

奥斯卡呢？

How does one get an Oscar?

John Platt

1986 年在 Schlumberger 的 AI 实验室实习，做计算机图形。我说可用弹性理论做柔软/织物等，写了弹性仿真器；其后代成为 Pixar 等大量物理仿真。跟实习生半开玩笑：实习干得好也能得奖。从工作到获奖约 20 年——学院要确认已被广泛使用；当时是论文，不是为某一部电影定制。

Student intern in 1986 at Schlumberger’s AI lab doing computer graphics. I said use elasticity theory for floppy things and fabric, wrote an elastic simulator; descendants became physics simulators in Pixar movies. I half-joke to interns: do a really good job and you can get awarded. About 20 years later—typical; they want to see wide use. It was a paper, not work for one specific movie.

Host

你长期挨着量子计算；轨迹怎么看？像聚变一样热过又冷过？Willow 纠错像真正可扩展？

You’ve been quantum-adjacent and worked on it at Google—trajectory? Exciting then stuck like fusion, now hints again? Is Willow error correction the right scaling?

John Platt

我按十年尺度平均——就是在进步。不完全是纯软件：要足够大、稳的系统才是计算机而非装置。我们在 NISQ（Preskill 的词）时代。量子团队有 Quantum Echoes 算法论文（我是众多共同作者之一），风格接近费曼式提议：在环里用量子机拟合参数化哈密顿/观测（如解码 NMR 参数）——是否大到足以突破仍 TBD。他们按 Neven 几年前的路线图扎实推进，到最后里程碑应能做惊人的事——数年量级，不是明天也不是三十年。超导是否最终赢家 TBD，但我觉得很可扩展。这些东西极娇气，本质有段时间仍是模拟计算机，别指望突然相位跃迁。物理比特对逻辑比特比例仍高；中性原子理论上全连接，实践限制我不完全清楚。硬件组非常强。Dave Bacon 常笑我：量子总是「再等二十年」——十年过去了，半路上；我也是十年前开始做聚变。顺便：九十年代初我爱卷积网；据我所知「convolutional net」这词是我造的——不想叫 LeNet，因为我不给 Yann 打工。

I average over decades—it’s progress. Not purely software; need large stable systems. We’re in the NISQ era—Preskill’s term. The quantum team has a Quantum Echoes paper I’m one of many co-authors on—Feynman-style: fit a parameterized Hamiltonian to observed data (e.g. decode NMR) in a loop; whether big enough for breakthroughs TBD. They’re executing Neven’s roadmap amazingly; at the last milestone, a machine that does amazing things—few to several years, not tomorrow, not 30 years. Whether superconducting wins is TBD; I find it promising and scalable. These are finicky analog computers for a while—don’t expect a sudden phase change. Physical-to-logical qubit ratios are still large; neutral atoms can in theory connect anything—to-anything; practice TBD. The hardware team is very skilled. Dave Bacon teases me that quantum is always 20 years away—10 years gone, halfway. I also started fusion 10 years ago. In the early ’90s I loved conv nets; as far as I can tell I coined “convolutional net”—everyone said LeNet, and I didn’t work for Yann.

Host

想留给听众什么？若能魔法消除一个瓶颈，会是什么？

Anything you want the audience to take away? If you could remove one bottleneck by magic, what would it be?

John Platt

ERA 是例子：我对 AI for Science 潜力极度兴奋——科学家的强力工具，不会被取代；希望他们把时间花在创造、严谨与哲学上，会很酷。若有魔法愿望：请做一座「Everything Lab」——丢 JSON 就能做任何实验。那几乎要解决 AI-complete 的机器人问题；若做到，会震撼——今天 ERA 类工具全是计算侧，数据仍要人去采。打破这点会无比美妙。

ERA is an example—I’m amazingly excited about AI for science as a power tool. Scientists won’t be replaced; I hope they spend time on creative, rigorous, philosophical work—way cool. Magic wish: an everything lab you send a JSON blob to that does any experiment at all—basically the AI-complete robotics problem. If we did, stunning—right now ERA-like tools are all computational; someone still gathers data. Breaking that would be utterly amazing.

Host

感谢你特地在回程改降落旧金山来录。

Appreciate you adjusting your schedule—landing in San Francisco on the way home to be here.

John Platt

谢谢邀请，非常开心。

Thank you for having me—it’s been a blast.