Stanford CS25: Transformers United V6 I Advancing Science and Medicine with Collaborative AI Agents
TL;DR
Google DeepMind researcher Vivek Natarajan discusses the development of Co-Scientist, an AI system designed to act as a collaborative partner for scientific discovery by moving beyond fast System 1 thinking to rigorous System 2 reasoning, emphasizing that true scientific AI requires the generality of human cognition rather than narrow specialization.
🧬 Birth of the Co-Scientist 3 insights
From medical licensing to hypothesis generation
Vive's team evolved from Med-PaLM (the first AI to pass medical licensing exams) to building collaborative AI for scientific hypothesis generation after Dr. Gary Peltz suggested applying LLMs to identify causative genes for rare diseases.
Lab-validated AI hypotheses
An early PALM-based agent generated a hypothesis about gene factors that Dr. Peltz validated through CRISPR knockin experiments on mice, resulting in a peer-reviewed paper published in Advanced Science.
The 'build while falling' methodology
The project began without a clear technical roadmap, following the approach of committing to ambitious goals and engineering solutions during the attempt rather than waiting for perfect conditions.
🧠 Rethinking AI Cognition for Science 2 insights
System 1 versus System 2 thinking
Current LLMs like Gemini or GPT primarily exhibit System 1 thinking—fast, intuitive, surface-level pattern matching—while scientific discovery requires slow, deliberate, rigorous System 2 thinking that develops over weeks or months.
Quantifying scientific complexity
Scientific tasks range from literature reviews (hours) to paradigm-shifting theories like general relativity (decades or lifetimes), requiring AI capable of sustained long-horizon reasoning across this entire spectrum.
🌐 The Generality Imperative 2 insights
AlphaFold's specialized brilliance
While AlphaFold accomplished work equivalent to millions of scientist-years and earned a Nobel Prize, it lacks generality because it only processes protein sequences and cannot adapt to other scientific domains.
Language as the foundation of general intelligence
Natural language serves as the critical input/output interface for building general-purpose scientific AI, enabling systems to understand diverse problems and attempt solutions across domains much like the human brain.
Bottom Line
To achieve true scientific breakthroughs, AI systems must evolve from narrow specialized tools into general-purpose collaborative agents capable of deliberate, long-horizon System 2 thinking through natural language interfaces.
More from Stanford Online
View all
Stanford CS336 Language Modeling from Scratch | Spring 2026 | Lecture 16: Post-Training - RLVR
This lecture explains why RLHF hits overoptimization limits with learned reward models, and how RLVR (Reinforcement Learning from Verifiable Rewards) enables unlimited compute scaling on verifiable tasks like math and coding through simpler algorithms like GRPO.
Stanford CS336 Language Modeling from Scratch | Spring 2026 | Lecture 15: Mid/Post-Training
This lecture explains how post-training transforms raw pre-trained models like GPT-3 into instruction-following systems like ChatGPT through supervised fine-tuning and reinforcement learning, emphasizing that high-quality data curation matters more than algorithmic sophistication.
Stanford CS336 Language Modeling from Scratch | Spring 2026 | Lecture 14: Data
This lecture details the pre-training data pipeline, covering the transformation of raw HTML and PDFs into linear text and classifier-based filtering strategies to curate domain-specific datasets, while emphasizing the strategic trade-off between data quality and training duration.
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Infrastructure, Capstone Case
Sachin Katti, OpenAI's head of industrial compute, details the infrastructure economics driving the AI supercycle, explaining how the company plans to scale to 30 gigawatts by 2030 while navigating the shift from training to inference-heavy agentic workloads and managing massive energy and supply chain constraints.