The AIGS Pod
door The AIGS
Nieuwste aflevering
Can a weather model learn without an answer key?
Can a weather model learn without an answer key?
Over deze podcast
The AIGS podcast tackling AI in earth system modeling and more.
Afleveringen
De nieuwste 20 van 30 afleveringen
- Can a weather model learn without an answer key? · 22 min · Aflevering 30Every forecast starts from a picture of now, built by expensive hand-made machinery. A new model learns that whole predict-then-correct loop from raw observations alone…
- Can computation create information? · 26 min · Aflevering 29Shannon and Kolmogorov say a short program can't create information, yet AlphaZero learns chess from the rules alone. Put the observer's compute budget into the definition and…
- What does attention do with a dynamical system? · 21 min · Aflevering 28One attention head can only average its past, so it smooths out a spring that needs a minus sign. Through a single sensor it can still rebuild a hidden state from history, if it…
- Do circuits explain model errors? · 24 min · Aflevering 27A circuit can keep 99.5% of a model's right answers and still miss most of its mistakes. We walk through how the paper measures that gap, why it survives the obvious excuses, and…
- Proving AI truly understands physics · 23 min · Aflevering 26Did a network that turns motion capture into radar learn the Doppler law, or just mimic it? A brisk, upbeat tour of LAWFUL's counterfactual tests and circuit hunt. Pair it with…
- When a network follows the law, has it learned the physics? · 28 min · Aflevering 25A transformer that predicts radar responds to controlled changes in motion the way the Doppler law says it should. Has it learned the physics? LAWFUL tests the behavior, finds a…
- Mapping visual circuits to fix AI · 24 min · Aflevering 24Seeing what a vision model encodes is not the same as seeing how it routes it. Vi-CD finds sparse, edge-based circuits inside vision transformers: the pathways behind a class…
- How AI weather models reinvented physics · 43 min · Aflevering 23What if AI weather models are solving physical equations after all, just not the ones in conventional NWP? Different models turn out to represent the atmosphere in strikingly…
- How AI predicts weather without physics · 54 min · Aflevering 22A long, unhurried walk through an open-source tool for poking at GraphCast's latent space: organise the processor's internal representations, run cosine similarity and PCA, and…
- How AI weather models actually think · 23 min · Aflevering 21A simple, easy-to-use visualisation tool for exploring the latent spaces of AI weather models, and what it finds when pointed at GraphCast: directions in latent space that seem…
- How AI weather models understand physics · 24 min · Aflevering 20Swapping ReLU for learnable B-spline activations from Kolmogorov-Arnold Networks nearly doubles the number of live features a sparse autoencoder finds inside a weather model…
- AI weather models independently learned physics · 42 min · Aflevering 19A long look at what happens when you apply LLM interpretability tooling to a weather model: sparse autoencoders surface human-legible features inside GraphCast, and nudging the…
- GraphCast learns the laws of physics · 20 min · Aflevering 18Sparse autoencoders borrowed from LLM interpretability research pull open GraphCast's hidden layers, revealing neurons that have independently learned tropical cyclones…
- Walrus masters physics from stars to oceans · 17 min · Aflevering 17From stellar plasmas to ocean currents, Walrus treats fluid dynamics as one continuous problem rather than nineteen separate ones, beating prior foundation models on both…
- Walrus AI unifies nineteen physics scenarios · 22 min · Aflevering 16Astrophysics, geoscience, rheology, plasma physics, acoustics, classical fluids: nineteen scenarios walk into a transformer. Walrus pretrains across all of them, then outperforms…
- Teaching AI the laws of physics · 19 min · Aflevering 15Maybe the mistake was ever treating Earth as physics-special. Walrus is a transformer trained on nineteen scenarios, from astrophysics to oceans, testing whether one foundation…
- Aurora AI beats multi-billion dollar supercomputers · 45 min · Aflevering 12The long form: Microsoft's Aurora foundation model, pretrained on over a million hours of geophysical data, matches or beats ECMWF-class numerical forecasts at a computational…
- Aurora AI beats traditional weather models · 19 min · Aflevering 14The short cut: how Microsoft's Aurora foundation model delivers forecasts in seconds that traditional numerical weather prediction needs hours of supercomputer time to produce.
- Microsoft Aurora outperforms traditional weather models · 24 min · Aflevering 13One pretrained model, four jobs: we walk through how Aurora fine-tunes a single 3D Swin Transformer backbone into weather forecasting, air pollution tracking, ocean wave…
- NVIDIA ATLAS beats global weather supercomputers · 1 u 0 min · Aflevering 10NVIDIA's ATLAS framework shows that scaling a general-purpose architecture, without bespoke training heuristics, is enough to beat the Integrated Forecasting System and GenCast…








