Key Moments
🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow
Latent Space PodcastWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
Google's AI system, ERA, can automate scientific discovery by framing problems as "scorable tasks," but human oversight is crucial to prevent "reward hacking" and ensure rigor.
Key Insights
Google's ERA (Empirical Re-Actor) framework allows scientists to map diverse scientific problems into "scorable tasks," where AI agents aim to maximize a defined score, streamlining model building and problem-solving.
ERA utilizes advanced LLMs like Gemini, enabling it to assist in defining scorable tasks through conversational interaction and then iteratively mutate and improve code to maximize the defined score.
The ERA system employs an optimistic algorithm (UCB) to explore potential solutions, not always selecting the best-performing notebook but rather one with the highest optimistic bound to efficiently make progress.
ERA can significantly accelerate scientific research by automating the laborious process of writing scientific code, allowing human scientists to focus on higher-level creativity and conceptualization of problems.
While ERA can identify solutions, it's crucial for scientists to maintain rigorous oversight to prevent the AI from "cheating" or "reward hacking" the scoring function, ensuring the model's findings are truly descriptive and not just predictive.
ERA has demonstrated success in various domains, including remote sensing for CO2 estimation, climate modeling (contrail impact analysis), and even prediction tasks like CDC COVID/flu case forecasting.
Automating scientific discovery with scorable tasks
Google's ERA (Empirical Re-Actor) system represents a significant advancement in AI for science, aiming to automate aspects of scientific discovery by reframing problems as "scorable tasks." This framework allows diverse scientific challenges, from statistical modeling to complex differential equations, to be translated into a format where an AI agent can optimize a specific score. This approach moves beyond traditional machine learning's predictive focus, enabling the creation of more descriptive models that capture underlying scientific principles. The core idea is to generate code that maximizes a defined score, surprisingly effective across many scientific disciplines. For example, in applied mathematics, ERA can tackle problems like asymptotic expansions by having an AI propose solutions and then verifying their correctness against specific parameters, a task that previously required significant human expertise.
The role of large language models in ERA
At the heart of ERA are powerful large language models (LLMs) like Gemini. These models are not just tools for generating code but also act as collaborators in defining the scorable tasks themselves. Users can converse with the AI, describing their scientific problem, and the AI agent helps to formulate a scorable task, often outputting a Python notebook with a scoring function. The AI then iteratively mutates and improves this code, driven by the objective to maximize the score. This symbiotic relationship leverages the LLM's vast world knowledge and reasoning capabilities, allowing it to propose novel solutions and code modifications that might not be immediately obvious to human researchers. The process leverages the AI's internal 'prior knowledge' derived from its training, leading to more effective problem-solving.
Intelligent exploration and optimization strategies
ERA employs sophisticated algorithms to explore the vast solution space efficiently. Instead of solely focusing on the best-performing solutions, it uses an optimistic approach, specifically the Upper Confidence Bound (UCB) algorithm, which is common in reinforcement learning. This means ERA considers both the current performance of a potential solution and the uncertainty associated with it. By selecting candidates with the highest optimistic bound (current performance plus a margin of error), ERA actively hunts for promising avenues, balancing exploitation of known good solutions with exploration of new possibilities. It can even recombine ideas from different candidate solutions to generate novel approaches, moving beyond simple greedy optimization.
Accelerating scientific research and shifting human focus
The primary benefit of ERA is its ability to drastically accelerate scientific research by automating the often time-consuming and intricate process of writing scientific code. This frees up human scientists from the 'grungy goop' of data wrangling and low-level coding, allowing them to concentrate on the more creative and philosophical aspects of science – defining the core problem, conceptualizing meaningful scoring functions, and interpreting results. In essence, ERA shifts the human scientist's role to be more about high-level scientific creativity and rigorous validation, rather than the detailed implementation of models. This is particularly impactful in fields where developing custom scientific software can be a major bottleneck.
The critical need for human oversight and rigor
Despite its power, ERA is not a fully autonomous discovery engine. A crucial aspect is the necessity for human scientists to maintain rigorous oversight. AI agents, like any optimization system, can fall prey to "reward hacking" or "cheating" – finding ways to maximize the score without genuinely solving the underlying scientific problem. Scientists must carefully define scoring functions to avoid unintended loopholes and continuously validate the AI's outputs. This meta-level thinking, ensuring the AI's solutions are descriptive and extrapolative rather than merely predictive or overfitted, remains a critical human responsibility. The analogy of a powerful tool that can "slice your fingers off" underscores the need for extreme care and rigor.
Applications in remote sensing and climate science
ERA has shown promise in diverse scientific applications. In remote sensing, it has been used to enhance the resolution of satellite data, for instance, by estimating CO2 concentrations from lower-resolution weather satellite imagery. In climate science, ERA played a key role in analyzing the warming impact of contrails, helping to solve a two-year-long problem related to estimating the effect of reflected sunlight. The AI was instrumental in identifying a simple yet effective model by exploring numerous confounding factors that had previously eluded human researchers. ERA has also contributed to predicting public health trends, such as COVID and flu cases, demonstrating its versatility across different scientific domains.
Distinguishing descriptive from predictive models
A core distinction discussed is between predictive and descriptive models. Predictive models aim to minimize error on existing data, essentially acting as sophisticated curve-fitting tools. Descriptive models, on the other hand, aim to capture the underlying physics or reality, enabling extrapolation to new, unseen scenarios. Science, fundamentally, seeks descriptive models. While machine learning often excels at prediction, ERA's goal is to facilitate the creation of descriptive models. This is achieved by incorporating known scientific principles or guiding the AI toward models that can generalize beyond the training data, though the line between prediction and description can be blurry, especially when LLMs provide 'intuition' based on their training data.
The evolving landscape of AI for science
The field of AI for science is undergoing a rapid phase change, driven by the advent of more general-purpose AI systems like ERA. While specialized models were once the norm, the current focus is on understanding how these more versatile tools can be integrated into scientific workflows. For young scientists, the advice is to develop deep domain expertise while simultaneously playing with and learning to use these new AI tools. The emphasis remains on human creativity, rigorous validation, and a deep understanding of scientific principles, with AI serving as a powerful amplifier. The physical lab work, a crucial bottleneck for empirical validation, is also highlighted as an area where AI integration is still nascent but holds immense potential.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Books
●Concepts
●People Referenced
Common Questions
ERA is Google's AI scientist that maps scientific problems into 'scorable tasks' where the goal is to maximize a score. It uses large language models like Gemini to generate and mutate code (Python notebooks) and an algorithm like Monte Carlo Tree Search to explore solutions, aiming to accelerate the scientific process. This allows scientists to focus on the 'scientific creativity level' rather than coding details.
Topics
Mentioned in this video
Google's AI scientist tool designed to automate and accelerate scientific discovery by mapping scientific problems into scorable tasks and using LLMs like Gemini to mutate and optimize code.
The department at Google where AI for science initiatives, including ERA, have been developed for over 10 years.
A Google-owned platform for machine learning competitions, which was the original inspiration for the ERA project (AutoKaggle problem).
The US public health agency that set up a competition for predicting COVID and flu cases, which ERA performed exceptionally well in.
A scientific conference where Google presented its work on solving the counterfactual modeling problem for contrails.
An institution that held a 'virtual cell challenge' which had a notable result, possibly involving overfitting.
A Google AI company that has been working on control systems for Tokamaks to prevent instabilities and disruptions.
A large international fusion experiment, mentioned with concerns about potential disruptions leading to a '30 billion dollar brick'.
A non-profit organization, co-founded by Google, focused on wildfire detection and mitigation, using satellite technology.
A US government agency that Google collaborated with to create new models for fire propagation.
An international organization whose estimates indicate 300,000 excess deaths annually from wildfire smoke.
The agency that funded Feynman's computation class, providing benefits like prime rib dinners for attendees.
An organization in Arizona where asteroid observations are sent for orbit calculation and discovery rights assignment.
A modern observatory in Chile with an advanced telescope (Simonyi Survey Telescope) that automates asteroid discovery.
An observatory where the speaker used a fast telescope to discover asteroids as part of a Caltech class in the 1980s.
Google's core AI model that powers ERA, providing the intelligence for mathematical reasoning, code generation, and optimization.
A tool that, when combined with ERA, can ingest numerous papers and generate extensive code, enhancing scientific discovery capabilities.
Another LLM with which the host experimented, strapping an open-source version of ERA to it to build a neural network for classification.
An early Google initiative that used data to track disease progression, mentioned as an ironic full circle given ERA's success in CDC competitions.
A powerful AI model for protein folding that shifted scientific paradigms towards data-driven approaches, though it may not provide traditional physical intuition.
The mobile operating system through which Google propagates wildfire information to users.
A process-based model from the 1970s used for fire propagation, which Google's new neural network proxy model aims to improve upon.
Hardware that significantly accelerated deep learning by optimizing BLAS operations, leading to breakthroughs in fields like image recognition.
A library for linear algebra operations, crucial for the efficiency of neural networks and optimized by GPUs.
NVIDIA's parallel computing platform and API, widely used for deep learning and key to NVIDIA's pivot from gaming.
The lead author on the ERA paper, highly skilled in applied math and capable of quickly producing scientific papers using the tool.
A meteorologist credited with developing the concept of chaos theory and the butterfly effect.
A famous physicist whom the speaker took a class from, noting his charisma and the difficulty of truly grasping his concepts.
The speaker's PhD advisor at Caltech, who co-taught Feynman's computation class and later won a Nobel Prize.
Co-teacher of Feynman's computation class and the speaker's computer system administrator (CIS admin) at Caltech.
A guest lecturer in Feynman's class, associated with the Connection Machine and early thinking about parallel computing architectures.
A theoretical physicist who coined the term NISQ (Noisy Intermediate-Scale Quantum) era.
A prominent figure in deep learning whose group was using GPUs for deep learning around 2010.
Runs the software team in Google Quantum and teases the speaker about being '20 years ahead of time' with his research interests.
A friend of the speaker from grad school who worked at NVIDIA and was instrumental in convincing the company to focus on deep learning with CUDA.
A pioneering AI researcher who the speaker notes might have been predated in coining the term 'convolutional net'.
CEO of NVIDIA, convinced by Brian Kotenzaro to pivot the company towards supporting deep learning with GPUs and CUDA.
A planetary scientist who, with her husband Gene, taught a class at Caltech where the speaker learned how to discover and name asteroids.
A platform where one of the papers related to ERA's applications, specifically in remote sensing, might be found.
A fusion energy company that Google previously collaborated with on systems to recommend new experiments for their field-reversed configuration (FRC) reactors.
Another company that works on field-reversed configuration (FRC) fusion, similar to TAE Technologies.
A company that Google partnered with to build a prototype satellite for wildfire detection.
A company that pivoted its focus from gaming to deep learning, largely due to the widespread adoption of CUDA for AI applications.
An animation studio that utilized physics simulators descended from the speaker's early work on elasticity theory to create movies.
A satellite that measures atmospheric CO2 concentrations with high accuracy and resolution.
An orbital research station to which the OC3 satellite is attached, providing CO2 measurements.
A weather satellite that provides infrared data, used to estimate CO2 despite not being designed for it, but with large pixels and less spectral resolution.
An early parallel computer, associated with Danny Hillis, reflecting early ideas about computational architectures.
An early minicomputer, used by the speaker and his research group in the 1980s, demonstrating the limited computing power of the era.
The algorithm used by ERA to explore potential code modifications, maintaining hundreds or thousands of notebook instances and selecting ones based on an optimistic algorithm.
A standard reinforcement learning algorithm used by ERA to optimistically select which code mutations to explore, aiming for efficient progress rather than immediate best performance.
The principle that any metric that becomes a target is no longer a good metric, applied to the challenges of overfitting and reward hacking in AI and leaderboards.
A database of protein structures, mentioned in the context of AlphaFold's comprehensive application to all available protein data.
A type of recurrent neural network and form of associative memory that generated significant excitement in the 1980s and is now seen as having parallels with modern transformers.
A modern deep learning architecture that has strong parallels to early associative memory concepts like Hopfield networks.
A class of neural networks, for which the speaker claims to have coined the term, preferring it over 'LeNet'.
A large visual database designed for use in visual object recognition software research, a key benchmark where GPUs showed their superiority over CPUs.
More from Latent Space
View all 253 summaries
143 minWhy We Made Jev — Diogo Almeida, TypeSafe Co-founder & CEO
70 minA Worm With 302 Neurons Inspired Their Architecture — Ramin Hasani, Liquid AI
87 minThe Watchdogs of AGI — Rune Kvist of AI Underwriting Company
94 minRecursive Self-Improvement: from Auto Research to Superintelligence — Richard Socher, Recursive
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free