Twelve themes, twelve authors.
Each episode follows one theme through this year’s programme and centres one paper that moved it. The venue line shows where we expect to find the papers; the guest is chosen by the selection method and announced on release.
How models learn to reason
What actually changes inside a model trained to think step by step, and what the evaluations miss.
Guest announced on releaseRetrieval, memory and long context
When a model should look something up, when it should remember, and why long context did not end retrieval.
Guest announced on releaseLanguage models for Asian languages
Tokenisation, data and evaluation for Japanese, Chinese, Korean, Thai, Bahasa and the languages between them.
Guest announced on releaseAgents that use tools
Planning, tool use and recovery from error, and how to tell an agent benchmark from a demo.
Guest announced on releaseThe science of evaluation
Contamination, judges that are models, and what a benchmark is for once models saturate it.
Guest announced on releaseTraining and serving efficiently
Quantisation, distillation, mixture-of-experts and the economics that decide which ideas ship.
Guest announced on releaseSeeing, reading and listening
Vision-language models, documents as images, video understanding and speech.
Guest announced on releaseDocuments, law and professional work
Contracts, filings and forms: the unglamorous documents where language models meet liability.
Guest announced on releaseAI for science
Proteins, materials and weather: what transfers from language models to scientific discovery and what does not.
Guest announced on releaseAlignment and safety research
Interpretability, oversight and robustness, explained by the people running the experiments.
Guest announced on releaseWhy deep learning works
Scaling, generalisation and optimisation: the theory catching up with practice.
Guest announced on releaseRobots and embodied learning
Learning from demonstration, simulation and the gap between a lab arm and a warehouse.
Guest announced on releaseThemes may change as papers are accepted. Venue lines are indicative; MLSys, COLT and CoRL are covered when the paper fits. No guest is named before they have agreed in writing.
Release rhythm
| When | What |
|---|---|
| Every other Tuesday | A Conversation (60 to 90 minutes) |
| Every Tuesday between | A Brief (15 minutes) from the same or a related paper |
| Every Friday | Meridian Notes, the newsletter |
| Before and after each major venue | A Venue Report (Meridian Pro) |