Latent Space

Anima Anandkumar on Why AI for Science Isn't Just Language Models

Anima Anandkumar· Bren Professor at Caltech
·~84 min·English·Latent Space
TrainingGPUAI InfrastructureMultimodal
TL;DR

Caltech's Anima Anandkumar argues that AI for science has to model the physical world directly, not just describe it in words: her neural operators fold data and physics together to approach the accuracy of supercomputer weather forecasts tens of thousands of times faster on a single GPU, and point toward foundation models that don't just simulate but discover.

01Core Mental Model

The World Is Latent

<strong>The physical world has hidden low-dimensional structure that data-driven AI can uncover</strong> — modeling in that compressed latent space reveals patterns that stay buried in the raw, high-dimensional measurements.

I really like the term latent space because that very much figures in a lot of my work uh because it's really you know the world is latent

— Anima Anandkumar, Latent Space
Key Insight
Traditional numerical solvers chase correctness at every point, so they never look for this shortcut. Anandkumar's bet is the opposite: nature is compressible, and a model that learns the compression beats one that re-derives every equation from scratch.
02AI for Science

Ideas Are Not Enough

<strong>The bottleneck in AI for science is not generating hypotheses but verifying them</strong> — language models are cheap idea machines, so Anandkumar's recent work builds neural networks you can formally prove things about before trusting them in a control loop.

But ideas are not enough, right? So you can have a lot of ideas. The bottleneck is going testing and verifying that they work in the real world.

— Anima Anandkumar, Latent Space
Key Insight
This reframes formal methods as an AI-safety tool for engineering, not just mathematics. If a neural network sits inside a drone or a nuclear reactor's control loop, a certified bound on how much the output can move when the input is perturbed is the difference between a demo and a deployable system.
03The Architecture

Zoom In, Zoom Out

<strong>A neural operator learns a mapping between continuous functions, not between fixed-size grids</strong> — so unlike a standard network locked to its training resolution, it can be asked for outputs at any scale and keep resolving finer detail.

neural operators enable us to zoom in and out as we like.

— Anima Anandkumar, Latent Space
Key Insight
This is why the physical world, not language, is the natural home for operators. Text has a fixed vocabulary and images a fixed resolution, but weather, fluids, and materials are multiscale — the coarse map you trained on is never the scale the phenomenon actually lives at.
04The Breakthrough

Tens of Thousands of Times Faster

<strong>Weather scientists warned AI could never rival decades of physics-based forecasting</strong> — a neural operator came within reach of their accuracy tens of thousands of times faster, small enough to run on a single consumer GPU.

not only accurate it's almost as close to what the traditional weather models can do accurately but also tens of thousands of times faster

— Anima Anandkumar, Latent Space
Key Insight
The speedup is not just a convenience — it changes who gets to forecast. Because the model is small and was open-sourced permissively, weather agencies in the global south can now run the same fidelity that only the largest agencies could afford before, which is a redistribution of capability, not only compute.
05Scaling to Climate

There Is Only One Earth

<strong>Baking the Earth's spherical geometry into the model is what lets a single network span short-term weather and long-term climate</strong> — architectures that assume a flat rectangle stay accurate for days, then blow up over months.

But to me, there's only one earth. You know, if you want a foundation model, if the claim is that it should be able to do both very short-term as well as very long-term together.

— Anima Anandkumar, Latent Space
Key Insight
This is a concrete case of a general principle she keeps returning to: when data is scarce, you pay for it with structure. The sphere is not learned from more examples — it is handed to the model as a prior, and that prior buys months of extra stability the data alone could never supply.
06Data Efficiency

The Physical World Is Forgiving

<strong>Physical models can learn from as few as tens of thousands of weather samples, or a few thousand plasma samples, and still predict rare extreme events</strong> — those events carry specific physical signatures the model can learn.

the physical world may be more forgiving because you know where there are extreme events like hurricanes that have very specific physical signature

— Anima Anandkumar, Latent Space
Key Insight
This inverts the usual intuition that rare events are the hardest to model. A hurricane is extreme but not arbitrary — its physics constrains what it can look like, so a model that has internalized the structure can recognize the signature from far fewer examples than a purely statistical view would demand.
07The Future

From Simulation to Design

<strong>The real prize is not faster simulation but inverse design: telling the model the outcome you want and having it search for the physical system that produces it</strong> — and with physics as guardrails, the designs it returns are ones it has already checked.

it's not just about simulation. It's about really novel designs and novel discoveries that enable us to move the needle of innovation itself.

— Anima Anandkumar, Latent Space
Key Insight
The guardrail is what makes this trustworthy. Generative design has existed for years, but its proposals were often physically nonsensical; embedding a validated physics model in the optimization loop means the designs the AI returns are ones it has already checked against the laws they must obey.
08The Thesis

AI for Science Is Not Just Language Models

<strong>Language models and agents are wrappers around symbols, not models of the physical reality they describe</strong> — which is why Anandkumar argues the missing piece of AI is a model that can simulate, design, and control that reality, and why rules written only for chatbots misread AI for science.

AI for science is not just language models and agents yes that's one aspect of it but ultimately you know those are still like external rappers in a way

— Anima Anandkumar, Latent Space
Key Insight
The policy edge here is sharp: she notes that regulatory frameworks tend to equate AI with language models, which can manipulate people. AI for science has a different risk-and-benefit profile, and a one-size-fits-all rule written for chatbots could throttle exactly the work that democratizes discovery.