🔬 An Oscar, Two Asteroids, and the Algorithm in Your sklearn: John Platt on AI for Science

The hosts, Brandon and RJ, introduce John Platt as a Google Fellow and head of applied science at Google Research. They highlight that Platt began college at 14 and his PhD at 18 at Caltech, where his adviser was John Hopfield, later a Nobel laureate. Platt developed Platt scaling and sequential minimal optimization, the standard algorithm for training support-vector machines and an algorithm still used through scikit-learn. He has also worked in applied mathematics, signal processing, fusion, quantum computing, climate modeling, and other fields. The hosts mention his Erdos–Bacon number of six, two asteroids he discovered, and an Academy Award for technical development connected to computer graphics and Pixar films. Platt describes himself as a “gigginger”—someone excited by almost everything—and says his interests have ranged widely. He explains that the asteroid discoveries came through a Caltech planetary-science class in the 1980s: students photographed the same region of sky at intervals, found moving objects using a stereoscope, measured their positions, and submitted the observations to the Minor Planet Center. Platt named one asteroid after his mother; that asteroid later turned out to have a roughly one-kilometer moon orbiting its approximately four-kilometer main body. His Oscar came much later, after work on physical simulation for computer graphics, including simulations of elasticity, floppy materials, and stretchy objects that became useful in Pixar’s production tools.

Platt says Google Research had worked on AI for science for more than ten years, initially using specialized classical machine-learning models such as convolutional networks. About two years before the interview, the team became interested in more general language models and discovered that many scientific problems could be reformulated as “scoreable tasks”: problems in which a system proposes code and is evaluated by a numerical score. The system can then search for code that improves that score. He distinguishes ordinary statistical fitting from broader scientific reasoning. In one example from applied mathematics, a system is asked to propose asymptotic expansions for differential equations. It tests whether the proposed solution behaves correctly as a parameter approaches zero, while Gemini performs the mathematical reasoning and modifies the code. In remote sensing, the system combines satellite observations with different trade-offs among revisit frequency, pixel size, and spectral resolution. It produced a model that could use high-resolution but infrequent CO2 measurements from OCO-2 or OCO-3 together with frequent, lower-resolution infrared imagery from weather satellites such as GOES, effectively creating an informed form of super-resolution. The interface begins as a conversation. An agent helps turn a scientist’s description into a Python notebook containing a scoring function, then repeatedly mutates the notebook. Platt says the system can also suggest datasets and read papers supplied by the user, rather than requiring complete starter code.

Platt explains that ERA is not simply a one-shot coding agent. Its harness uses Monte Carlo tree search and keeps hundreds or thousands of candidate notebooks. Gemini proposes changes, runs the resulting code, and places promising candidates back into the pool. The selection mechanism uses an upper-confidence-bound strategy: it does not always choose the currently best notebook, but estimates which candidate has the best optimistic potential, allowing the search to explore candidates that may improve substantially. The system can also recombine ideas from two candidates to create a third. The initial candidates may come from Gemini itself. Given a textual problem description—and sometimes several papers describing prior solutions—it can write an initial implementation, repair bugs, and continue mutating the code. Platt says the practical parallelism is about ten leaves at a time. Running thousands of branches simultaneously would consume substantial computation but prevent candidates from learning from one another. After each batch, the shared context contains the previous reasoning, code, and results, so Gemini can learn across attempts rather than acting as thousands of independent agents. The hosts point out that choosing the score is often the hardest and most human part. Platt agrees that the scoring function frequently needs several iterations because an agent may exploit an unintended loophole. Scientists must tell it what not to do, inspect failures, and refine the objective. The payoff is that researchers spend less time importing files, joining databases, and debugging infrastructure, and more time deciding what scientific quantity should actually be optimized.

The hosts ask whether ERA’s relentless search creates a severe multiple-hypothesis-testing problem. Platt separates predictive models from descriptive models. A predictive model may simply minimize error on a dataset. A descriptive scientific model should capture something about how reality works and therefore extrapolate beyond the observations used to fit it. He compares a model trained only on falling apples with Newtonian gravity: an “apple model” might predict more apples but have no principled basis for predicting planetary motion. Platt says the boundary is not perfectly sharp. Scientists bring intuition and a large body of known facts to model-building, then constrain their models to remain consistent with those facts. ERA currently supplies language-model intuition and can inherit biases from papers and instructions supplied by researchers. It is very good at adapting published techniques—for example, reconstructing the code behind a paper or applying methods involving equivariance and other specialized modeling ideas—but Platt does not claim that it independently discovers entirely new physical principles from nothing. The hosts return to ordinary overfitting. Platt’s answer is blunt: ERA is a power tool, so researchers must become more rigorous, not less. They need genuinely hidden holdout datasets, careful validation, and procedures designed to prevent the system from fooling them. He invokes Goodhart’s law: once a metric becomes a target, it stops being a reliable metric. Kaggle competitions illustrate this vividly; competitors have exploited tiny labeling errors, including a half-pixel convention in a contrail-related task, to improve leaderboard scores without solving the intended problem.

Platt describes Google’s work on persistent contrails, the white clouds left by aircraft. He says estimates attribute about 1% of anthropogenic global warming to them. Contrails reflect sunlight during the day, but they also absorb outgoing infrared radiation continuously, acting like a blanket; the second effect often dominates. In busy regions such as Europe, contrail-induced cloud cover can produce roughly one watt per square meter of local forcing, compared with about three watts per square meter for average global anthropogenic warming. Contrails form when aircraft pass through invisible regions of ice supersaturation. The jet exhaust provides moisture that freezes, and roughly one unit of emitted ice can draw in about 10,000 units of atmospheric water under the right conditions. These regions can persist for days, so aircraft can sometimes avoid them by descending approximately two flight levels, at a relatively small fuel cost. Google combines satellite imagery, weather information, and a custom model to identify and predict the regions and provide maps to flight-planning systems. A harder problem was estimating how much warming a contrail caused compared with the counterfactual world in which it had never formed. ERA searched over possible confounders and found a relatively simple model that solved a reflected-sunlight problem the team had struggled with for two years. Platt emphasizes that the result was not an enormous black-box program; it was mainly a useful combination of variables the researchers had not tried. He contrasts this success with full climate modeling, where data are sparse, the system is non-stationary, and future conditions differ from the historical record. Weather prediction has improved dramatically, but climate prediction still requires models constrained by known physics and careful treatment of uncertain processes such as ice-cloud microphysics.

Platt distinguishes weather prediction from climate modeling. Weather forecasts attempt to follow the atmosphere’s chaotic trajectory for roughly 15 days, while climate models estimate changing long-term statistics. Google’s recent weather advances have largely used large amounts of data and compute, and the company has improved tropical-cyclone track predictions. Climate is harder because the underlying attractor can change as temperatures, ice, land, and ecosystems change; a simulator can become unstable without making clear whether the instability is physical or merely a model failure. The conversation also touches on fusion. Platt says every fusion approach must satisfy a Lawson-type condition involving density, temperature, and energy-confinement time. News stories often emphasize improvement in only one of those quantities, while a working reactor needs all three. Tokamaks require control systems to prevent disruptive instabilities; field-reversed configurations have different problems, such as keeping the plasma centered. Platt is optimistic that commercially relevant fusion could be much closer than the old joke that it is always 30 years away, but he remains cautious. For climate adaptation, Platt describes FireSat, a planned constellation of roughly 50–80 low-Earth-orbit satellites equipped to detect fires in midwave infrared. With enough satellites, the system could identify fires about five meters across anywhere on Earth within roughly 15–20 minutes. Google is also using existing satellite data, Android, Search, and fire-spread models developed with the U.S. Forest Service. Platt frames climate change as both a disease whose causes must be addressed and a crisis whose symptoms require adaptation. He cites an estimate of 300,000 excess deaths annually worldwide from wildfire smoke and says early detection could make fires small enough to extinguish before they become uncontrollable.

Platt says AI for science has undergone a phase change in the last 12–18 months. Researchers are moving from building one specialized model for each problem toward using general systems that can read papers, write code, search model spaces, and propose hypotheses. His advice to young scientists is to develop deep domain expertise while actively experimenting with the new tools. He does not expect scientists to disappear: creativity, judgment, philosophy, and rigor remain necessary, and physical laboratory work is still a major bottleneck because there is no general-purpose “everything lab” that can perform arbitrary experiments. The hosts ask whether people can develop expertise if AI instantly solves the simple problems through which earlier generations learned. Platt compares this to hiking up a mountain even when a vehicle could drive there. He has written numerical libraries, machine-learning algorithms, and boosting implementations multiple times in different languages, and says that doing the hard work built durable understanding. He recommends deliberately setting aside training time rather than optimizing every minute for immediate productivity. At Google, he protects a version of 20% time so people can learn, play, and attempt wild ideas without first having to justify them. The episode closes with Platt’s broader view of AI’s future. He sees enduring value in understanding the physical world, mathematics, creativity, and careful checking. He recalls taking a 1982 Caltech course taught by Richard Feynman, John Hopfield, and Carver Mead on the physics of computation, when the field was still unsure what quantum computing should mean and the group shared a VAX computer and an 80-megabyte disk. He views modern quantum computing as progressing steadily rather than arriving overnight, and says his ideal breakthrough would be an automated laboratory that accepts an experiment specification and carries it out reliably.