John Platt surfaced in the conversation as an unusually varied figure: the source mentions an Academy Award (Oscar), two textbook algorithms, two asteroids named after him, and an Erdős–Bacon number of 6. Over a long career Platt and teams at Google have repeatedly worked on areas that were ahead of their time — from convolutional neural networks to fusion research, quantum computing, and the use of computation and AI to tackle hard scientific problems.
Empirical Research Assistance (ERA): goal and mechanism
Platt’s group noticed that many scientific problems can be expressed as a ‘‘scoreable task’’: there is a score function and the objective is to find code that maximizes that score. Often the challenge is defining a good score; even with a score function in hand, finding the maximizer can require substantial effort.
Google’s Empirical Research Assistance (ERA) was developed to address this general optimization problem. The idea grew partly from an ‘‘auto‑Kaggle’’ concept — an AI that can solve arbitrary Kaggle problems — and because Kaggle is owned by Google, the data was readily available for development.
Conceptually ERA is straightforward. A large language model — in the article mainly Gemini is mentioned, though ERA can work with other LLMs — maintains a running tree of past experiments (notebooks) and their trajectories. This resembles Monte Carlo Tree Search: at each iteration an Upper Confidence Bound‑style rule selects which notebooks are most promising to mutate. The selection is optimistic rather than greedy, so sometimes even the fifth‑best branch is explored. Gemini proposes about ten mutations for each chosen notebook, and branches share histories so different leaves can learn from each other.
Platt summarized the system as ‘‘almost like having a hyper‑eager grad student who doesn’t sleep.’’ Evolutionary algorithms have existed since the 1970s, but ERA works well because the language model knows where to look. The article notes a significant improvement between Gemini 2.0 and 2.5: the jump moved the system from barely working to working very well.
Results and research output
ERA proved powerful enough that Platt’s team used it to solve multiple outstanding problems, producing at least ten papers. Several of these papers were related to climate applications.
A central practical concern is avoiding self‑deception when you have a strong optimizer. Platt’s response is that ERA provides predictive models, but it remains the scientist’s duty to ensure those models are truly descriptive. That responsibility involves careful, classical machine learning practice. ‘‘It’s a power tool. It can slice your fingers off,’’ Platt warned. The conversation highlighted how competition settings (like Kaggle) can incentivize clever overfitting that wins prizes but does not solve the real problem.
Platt invoked Goodhart’s law: ‘‘any metric that becomes a target is no longer good as a metric.’’ His pragmatic starting advice is modest: ‘‘Always just fit linear regression. Just do it. Just do it. Just do it. Or SVM.’’
Tackling climate: contrails and FireSat
Platt’s team has worked on several climate‑related projects. The most striking example discussed was reducing the warming impact of airplane contrails. Those thin streaks of ice crystals behind planes account for roughly 1% of human‑caused global warming, according to the article. Some contrails persist for days and behave in the infrared like a thermal blanket, trapping heat both day and night.
Mechanistically, parts of the atmosphere can become ice‑supersaturated: cold air can hold vapor amounts between the liquid and ice equilibrium limits, and without a seed, ice will not form. A small exhaust particle can seed immediate crystallization; remarkably, one gram of exhaust can produce around ten kilograms of ice crystals.
In principle the mitigation is simple: fly around or a flight level or two above or below the regions where contrails form. The practical difficulty is measuring how much warming this prevents — a counterfactual question that confounded Platt’s team for over two years. They had a working model for the heat‑trapping component but lacked a model for reflected sunlight. ERA found a simple model including confounders the team had not considered and helped crack the problem.
Another project mentioned briefly is FireSat, a satellite constellation approach to rapidly detect small fires before they grow large and uncontrollable. In dry years a small ignition can expand to hundreds of thousands of acres; detecting a fire when it is room‑sized would make suppression feasible, whereas a one‑acre fire is much harder to contain.
Perspective, domain expertise and the scientist’s role
Platt has a long view across science, computation, and AI. He recalled taking a ‘‘physics of computation’’ course with Richard Feynman in 1982, at a time when quantum computing was still a vague idea without firm theory or experiments. He also recalled doing science with very limited compute — one million operations per second was cutting edge.
Platt’s core recommendation is to develop deep domain expertise. Taste and judgment emerge from wrestling with hard problems. He also surprisingly recommends spending time ‘‘the old fashioned way’’: implement tools yourself and play with them. ‘‘You could drive up the mountain, or you could hike up the mountain, and maybe it’s okay, even fun, to occasionally hike,’’ he said.
Summing up, Platt believes there will continue to be a role for scientists: automation opens more opportunities for creative, rigorous and philosophical work, but one must still do the grunt work. He warned against a world that hyper‑optimizes everything and risks losing value through overfitting.
He returned to a lesson he got from Feynman forty years ago: ‘‘you must not fool yourself, and you are the easiest person to fool.’’
Additional topics noted in the conversation
- Platt’s view that fusion is three years away, not thirty, and the implications of the Lawson criterion.
- Why superconducting qubits remain finicky.
- The asteroid he named after his mother, which turned out to have a moon.
- A looming helium shortage seldom discussed.
- The origin of NeurIPS (people crashing a private Snowbird workshop) and comments that Hopfield networks are sufficient in some contexts.
- Finding asteroids in 1985 with film, a stereoscope, and correspondence with Brian Marsden; contrasted with the Vera Rubin Observatory’s 11,000 finds in six weeks.
- Quantum echoes, the NISQ era, and a cautious stance on quantum timelines.
- Anecdotes such as being Carver Mead’s sysadmin on a VAX with an 80 MB disk, and a 20% time rule in Platt’s group to allow learning projects without prior approval.
Overall the discussion paints ERA as a powerful, pragmatic tool for automating optimization across scientific problems while emphasizing methodological care and the continuing need for expert scientists to set goals, validate models and do creative work.
