How AI is Transforming Scientific Discovery While Keeping Humans at the Center | Stanford HAI

How AI is Transforming Scientific Discovery While Keeping Humans at the Center | Stanford HAI

AI Changes What’s Possible…

For scientists, AI has shifted beyond a narrow tool to an autonomous collaborator or a scientist in its own right. It generates hypotheses, designs experiments, analyzes data, and speeds discovery.

In fields like Earth and climate science, AI is replacing computationally expensive numerical simulations with high-speed emulators. The Samudra model can predict ocean states 1,000 times faster than traditional models, running on a single GPU to simulate 1,000 years of climate per day (prior models could simulate 12 years per day), noted Laure Zanna, New York University professor of mathematics and data science.

In biology, Brian Hie, Stanford assistant professor of chemical engineering, showed how the EVO DNA language model can generate new CRISPR-Cas systems for genome editing and has designed 16 previously unknown viruses that kill bacteria.

In neuroscience, researchers are building “digital twins” of the brain to run millions of virtual experiences to decipher the neural code, explained Stanford ophthalmology professor Andreas Tolias, who also collaborates with Surya Ganguli.

And in mathematics, Carina Hong, founder and CEO of Axiom, noted that AI is making progress on long-standing conjectures in algebraic number theory.

Some are moving beyond using AI as a tool to build fully autonomous AI agents capable of conducting research independently. James Zou, Stanford associate professor of biomedical data science, built the Virtual Lab, where AI agents generate hypotheses, design experiments, and analyze data. These agents run their own group meetings and discuss possible solutions to challenges presented by humans. In one example, he asked the agents to design binders to new COVID variants. Within days, the agents had designed new antibodies that Zou’s lab then synthesized and made in a wet lab.

“Experimentally, we showed that they’re actually better binders to the new COVID variants than previous human-designed nanobodies,” he said.

Panel 1: AI for Science

Generative AI Designing Complete Biological Systems (Brian Hie, Assistant Professor of Chemical Engineering at Stanford University)

AI models have evolved from designing individual molecules to complete genomes (from genes to whole viruses)
EVO DNA language models trained on millions of genomes can now generate:
- New CRISPR-Cas genome editing systems
- Novel anti-CRISPR genes with no similarity to any known natural genes
- 16 viable new bacteriophage species that overcome bacterial resistance better than natural phages
Models learn biology's "rules" from sequence data alone—similar to how language models learn from text

Digital Twins of the Brain (Andreas Tolias, Stanford Professor of Ophthalmology)

Using new brain recording technology and AI to build "digital twins" that can run millions of virtual experiments to decode neural activity
Recorded 3.1 million neurons (largest neuroscience dataset) but still data-limited—like trying to understand language from sparse text
New Stanford initiative (NEMA) aims to scale to tens of millions of neurons in primates during naturalistic behavior
Key discoveries: pupil dilation helps detect predators; universal wiring rules in visual cortex
Remarkably: Almost every individual neuron is interpretable using human language

Decoding the Genome's Control Language (Anshul Kundaje, Stanford Assistant Professor of Genetics and Computer Science)

Deep learning models predict how DNA sequence controls gene activity across thousands of cell types
3-4 million control elements (not just 25,000 genes) regulate which genes turn on/off in different cells
Models can:
- Predict effects of genetic mutations causing rare diseases
- Design sequence edits using CRISPR to increase/decrease gene activity by 200%
- Prioritize which of millions of genetic variants actually matter for disease

The Prediction vs. Understanding Debate

The challenge: Have we traded understanding for prediction power?
The response: Models aren't actually black boxes—can be interpreted to reveal:
- New biophysics principles (proteins)
- Genetic regulatory syntax (DNA)
- Neural coding principles (brain)
Interpretation is critical for debugging biased data and ensuring models learn real biology, not artifacts

Data Challenges

Genomics paradox: Looks like lots of data (3 billion DNA base pairs) but actually data-limited because each cell type has different logic
Neuroscience: Desperately needs to move from hypothesis-driven sparse data to large-scale systematic data collection
Solution: Combination of large foundation models + iterative "active learning" where models identify gaps to fill with targeted experiments

Peer Review Crisis

Current system unsustainable as AI accelerates output
Proposed evolution: Open, non-anonymous peer review with science released immediately
Need automated approaches to verify hypotheses as they become "cheap" to generate
Community must evolve how it validates and curates knowledge

Panel 2: AI for Earth

AI Weather Forecasting at Scale (Jean Kossaifi, Senior Research Scientist, NVIDIA)

AI weather models can now run 1,000+ times faster than traditional supercomputer simulations while matching accuracy
Transformer-based models trained on ERA5 data can generate probabilistic forecasts (multiple possible futures) in minutes vs. days
Key breakthrough: Models learn physics from data alone without explicit physics equations coded in
Can simulate complex phenomena like hurricanes at high resolution without actually running at fine grid scales
Challenge: Different from computer vision—need physical accuracy, not just realistic-looking outputs

Ocean Climate Emulators (Laure Zanna, Joseph B. Keller and Herbert B. Keller Professor in Applied Mathematics; Professor of Mathematics and Data Science, New York University)

Built Samudra ocean AI emulator that can simulate 1,000 years of climate per day on single GPU (vs. 12 years/day for traditional models)
Trained on simulation data, can run stable predictions for centuries
Recently coupled ocean + atmosphere models: full climate simulator running 1,000+ years/day
Enables "what if" experiments: testing different CO2 levels, wind patterns, ocean circulation changes at unprecedented speed
Challenge: Models are sensitive to initial conditions and don't capture all physics yet—can't yet explain why ocean absorbs 90% of excess heat

Watching AI Models Learn Climate Physics (Elizabeth A. Barnes, Professor and Dalton Family Chair in Environmental Data Science & Sustainability, Boston University)

Focus on training dynamics: watching how and when models learn, not just final performance
Discovery: Models learn to forecast extreme events (hurricanes, atmospheric rivers) early in training, then "forget" them and may never relearn
Can identify which network layers forget physics and potentially intervene to prevent it
Novel capability: AI models can run time backwards—start with event (like Hurricane Sandy) and generate 1,000 trajectories that led to it

The Physics Debate

Paradigm shift: 10 years ago everyone emphasized injecting physics into models; now pure data-driven approaches often work better
Physics is still embedded in:
- Data selection (which variables, time steps, spatial scales)
- Verification (do outputs obey known laws?)
- Problem setup and interpretation
Models may be learning physics we don't yet know—challenge is identifying when this happens vs. overfitting

Accessibility & Open Science

All work presented is fully open source (code, models, data)
Goal: Anyone anywhere can run climate simulations without being a climate modeler
Democratizing access to tools previously requiring supercomputers

Keynote: Dario Gil

Science is Fundamentally Transforming

Scientific discovery is shifting from sequential steps to continuous, integrated systems that learn, adapt, and act in real time
AI, computing, and scientific infrastructure are converging to change the pace from weeks/months to minutes/hours

The Genesis Mission: A National-Scale Scientific Platform

Launched by executive order in November 2025, Genesis connects high-performance computing, AI, quantum technologies, and experimental facilities (including robotic labs) into a unified discovery architecture
Goal is to double the productivity and impact of American science and engineering within a decade and solidify U.S. global technological leadership
Already operational with unprecedented response: 8,000+ applications (3x previous record) from 800+ institutions across all 50 states
38 company MOUs signed (and growing); Genesis Mission Consortium launched as public-private partnership vehicle

Three Pillars Driving the Mission

Energy: Accelerating deployment of reliable, affordable energy systems
Discovery Science: Advancing innovation across materials, chemistry, biology, physics, and engineering
National Security: Supporting stronger supply chains, faster advanced manufacturing, and mission-ready materials

Real-World Impact Already Happening

Fusion energy: AI predicts plasma behavior in real time, optimizing operations
Materials science: Screening properties computationally in hours vs. months of bench work
Grid resilience: Evaluating 43,000 power grid scenarios in 5 minutes (vs. 30 scenarios previously)
Protein structures: AI expanded from 200,000 structures (over 50 years) to 200+ million in just a few years

Designed for Global Collaboration

While U.S.-led, Genesis is built for "trusted interoperability" with international partners
Challenges are inherently global (energy, biotech, quantum)—no single country has all capabilities
Focus on co-development, shared systems, and shared acceleration of discovery

Panel 3: AI and the Universe

Particle Physics: Prediction & Inference (Kyle Cranmer, David R. Anderson Director of the UW–Madison Data Science Institute)

Two key AI patterns in physics: Prediction (theory → data) and Inference (data → theory)
Symbolic predictions: Using generative AI + formal verification to solve equations with millions/billions of terms (similar to AI for math)
Numerical simulations: AI accelerates lattice quantum field theory calculations while maintaining correctness through mathematical guarantees
Simulation-based inference: Revolutionary approach using neural networks to approximate likelihood functions when traditional statistics fail—now applied across neuroscience, astrophysics, economics, materials science
Core challenge: Ensuring correctness through calibration and uncertainty quantification when AI is embedded in data processing pipelines

Astronomy: Tools Transform Workflow, Not Just Speed (Douglas Finkbeiner, Professor of Astronomy and Physics, Harvard University)

AI hasn't led to 100 papers/year yet—instead producing better papers, not more papers
Examples: More sophisticated interactive figures; trying all research ideas simultaneously instead of down-selecting
Students still essential: AI is a "black box collaborator" requiring same validation as human grad students (asymptotic limits, cross-checks, tests)

Mathematics: Explosion in Automated Theorem Proving (Carina Hong, Founder and CEO, Axiom AI)

Dramatic progress: AI took silver medal at IMO (July 2024) → Perfect score on hardest undergrad exam (December 2025) in just 1.5 years
Solved dozens of open Erdős problems (though some were already in literature—AI excellent at literature search)
Two major capabilities:
1. End-to-end theorem proving producing fully verifiable proofs (e.g., partial results on algebraic number theory conjectures)
2. Auto-formalization: Converting PDF papers into verifiable code for easier peer review
Scaling laws apply: Proof graphs scaled from <50 nodes to 5,000+ nodes with more compute
Expert knowledge still essential: Selecting problems, providing problem setup, validating results

Open Source & Collaboration

Critical importance of open source tools and detailed documentation for mathematicians/physicists to use AI
Some mathematicians building better agents with less compute by focusing on agent design rather than raw model power
Question raised: Is GitHub the right interface for large-scale mathematical collaboration? (vs. traditional papers with 5 co-authors max)

Trust, Verification & Publishing

Validation remains unchanged: Test AI results exactly like grad student results
Formal verification increasingly important for both math and code
Rethinking papers:
- AI writes and reads papers—is journal the right API for scientist-to-scientist communication?
- Writing still valuable for organizing thoughts even if only AI reads it
- Need to "unbundle" paper's multiple roles: communication, discovery, pedagogy, archival

Unique Aspects of Physics/Math

Physics advantage: Clear causal mechanisms (unlike social sciences where causality is ambiguous)
Math theorem proving closely analogous to theoretical physics problem-solving
Both fields have formal verification options that other sciences lack
Challenge: Physics models often have "physics problems" (like AI not understanding real-world physics constraints)

Panel 4: AI and Human Understanding

Core Tensions: Industry vs. Academic Values (Angèle Christin, HAI Associate Director, Stanford Associate Professor of Communication)

AI/LLMs embody Silicon Valley logics that conflict with academic values: opacity vs openness, efficiency vs creativity, cost-cutting (replacing workers) vs mentorship
Risk: Letting industry-driven AI impose logics that don't serve academic priorities
Academia needs to proactively decide how to use AI on its own terms

AI Does Deduction, Humans Do Abduction (James Evans, Max Palevsky Professor at the University of Chicago)

AI excels at deductive science: filling missing cells based on known patterns
Real breakthroughs come from abductive science: creative leaps when encountering surprising violations of expectations
AI-assisted science becoming "monoculture": Papers using AI get 300% more citations but all move toward same "big data" in field, shrinking question diversity
Exciting counter-approach: Use AI to find surprising/unpredictable patterns, then build cascades of agents around them

Quantum Matter Physics Challenges for AI (Eun-Ah Kim, the Hans A. Bethe Professor at Cornell University)

Structural challenges: Diverse action spaces, sparse/biased documentation (no failure reporting), strict symmetry requirements, huge dimensional spaces
Current frontier models perform poorly (30% success rate) on graduate-level physics problems
Best AI use cases: Literature search/idea bouncing; automating well-established algorithmic workflows; systematic phenomenology with thoughtful human design of representations/objectives

Virtual Labs & AI Scientists (James Zou, Stanford Assistant Professor of Biomedical Data Science)

Built "Virtual Lab" with AI professor and student agents with different expertise who run group meetings autonomously
Designed antibody binders to new COVID variants in days—better than previous human designs when tested experimentally
Organized first conference with AI as authors and reviewers: 200+ submissions, 48 accepted
- Finding: Higher quality papers had more human input, especially early stages (hypothesis/design)
- Later stages (analysis/writing) had more AI autonomy
Vision: Convert all papers to "paper agents" that can explain methods, apply to new problems, collaborate with other paper agents—millions of agents discovering new connections

Hopes for 10 Years From Now:

Faster question-to-answer cycles enabling trying 10 different approaches quickly
AI enabling cross-disciplinary connections by removing syntax/vocabulary barriers
"Star Trek future" with computational hermeneutics—interrogating AI internals to distill deployable principles

Fears for 10 Years From Now:

Skill atrophy
Interpolation overwhelming extrapolation: AI makes combining existing ideas so easy it disincentivizes true breakthroughs
"AI horticulture"—everything becomes AI all the time

Key Insight: Metrics and Verification

Engineering domains with clear metrics (does drug work?) can use autonomous AI
But even there: bad proxies cause problems (healthcare algorithm using spending as proxy for health penalized people of color)
Math has verifiable proofs but still values the journey of understanding, not just the artifact
Science somewhere in between: wants impact + intuitive understanding
Need to co-evolve objectives with AI systems to avoid diminishing returns or catastrophic miscalibration

What Makes Understanding Important

Transformative vs. incremental improvement: Understanding bounds tells you if hill-climbing will reach something meaningful
Teaching the community/next generation
Knowing which metrics matter (may not know all relevant metrics upfront)
Science is socially constructed craftsmanship requiring community vetting over careers