# 🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow

## Executive summary

John Platt discusses Google's Empirical Research Assistance (ERA), an AI system designed to accelerate scientific discovery by mapping complex scientific problems into 'scorable tasks.' ERA combines Large Language Models (LLMs) with advanced search techniques like Monte Carlo Tree Search and Upper Confidence Bound (UCB) to iteratively propose and refine code that maximizes a defined score. The discussion highlights the critical distinction between descriptive models (what science aims for, allowing extrapolation based on physical laws) and purely predictive statistical models. The system's power lies in its ability to automate the initial coding and hypothesis generation, allowing human experts to focus on the high-level scientific creativity and defining the correct scoring function.

## Key takeaways

- ERA: Automating Scientific Hypothesis Generation: ERA maps scientific problems into 'scorable tasks,' allowing the system to iteratively propose and mutate code to maximize a defined score. This process moves beyond traditional machine learning by leveraging LLMs' vast prior knowledge and combining it with algorithmic search (Monte Carlo research) to solve problems like estimating CO2 concentrations or predicting wildfire boundaries.
- The Importance of Scientific Rigor and Domain Expertise: Platt emphasizes that while AI is a powerful tool, human input remains crucial for defining the correct scoring function and maintaining scientific rigor. He warns against 'reward hacking' and the risks associated with overfitting, noting that the ability to distinguish between predictive and descriptive models is a core human task.
- AI's Role in Climate and Earth Science: AI is being applied to complex, non-stationary problems like climate modeling and contrail warming. For instance, ERA helped solve a counterfactual problem—estimating the effect of contrails on outgoing longwave radiation—by finding a simple model that previously went unnoticed.
- The Future of Scientific Computing: The field is undergoing a 'phase change,' moving from specialized, single-problem models to general AI toolchains. The optimal approach for young scientists is to combine deep domain expertise (scientific taste) with the ability to utilize and experiment with multiple advanced tools.

## Technical details

- ERA Architecture: ERA uses a specialized harness running Monte Carlo research. It maintains a tree-structured candidate pool of Python notebooks, using the Upper Confidence Bound (UCB) algorithm to select the next best candidate for mutation, rather than always choosing the highest-scoring one. The core AI underneath is Gemini, which provides the initial code generation and mutation capability.
- Modeling Paradigms: A key distinction is drawn between *predictive models* (aiming for the lowest error rate on data) and *descriptive models* (aiming to capture the underlying physics or 'description of reality' to allow for extrapolation).
- Advanced ML/Optimization Techniques: The system utilizes concepts from applied math, such as asymptotic expansions, to turn differential equations into empirical, scorable tasks. It also employs techniques like 'informed super resolution' to combine data from disparate sources (e.g., multiple satellites) to estimate missing information.
- Climate Modeling Challenges: Climate models are challenging because they are non-stationary (the underlying physics changes over time) and involve massive uncertainty (e.g., the 300 ppm uncertainty in CO2 absorption). The problem of determining the impact of interventions requires complex counterfactual modeling.

## Practical implications

- AI tools can significantly reduce the time spent on the 'grunt work' of scientific coding and data manipulation, allowing researchers to focus on high-level scientific creativity and hypothesis formulation.
- The development of AI-driven scientific tools is shifting the focus of scientific research toward defining precise, quantifiable scoring functions for complex, real-world problems (e.g., climate mitigation, wildfire prediction).
- The integration of AI with physical constraints (e.g., physics-informed neural networks) is proving highly effective, enabling the reconstruction or reverse engineering of complex scientific papers.
- For build engineers, this suggests a future where complex, multi-domain systems (like climate or fusion modeling) will require AI agents to manage data fusion, simulation, and iterative hypothesis testing.

## Topics

Artificial Intelligence, Scientific Computing, Machine Learning, Climate Modeling, Optimization, LLMs, Google's Empirical Research Assistance (ERA), Gemini 2.0 to 2.5, Upper Confidence Bound (UCB)

Source: https://www.youtube.com/watch?v=2xBSGluFkG0