'Beautiful maths, useless models': a scaling lesson from variational inference work
RylanSchaeffer · x · 2026-10-20
Schaeffer recounts three papers on streaming variational inference for Bayesian nonparametric models (with Weiwei Pan, Blake Bordelon, Yilun Du and others), chosen hoping such models could reconcile odd associative-learning findings with memory engram research. The verdict: the maths were beautiful but the models useless against deep learning. His lesson — sophisticated methods that don't scale lose to simple methods that do. The attached figure shows the Dynamical Chinese Restaurant Process under four dynamics, comparing Monte Carlo samples with their analytical recursion.
More from Research
- EMNLP paper: diffusion LMs internally encode latent denoising timestep representations — gsarti_ · 2026-10-23
- KU Leuven and VIB open two Group Leader professorships in AI for Biology, start 2027 — jpillowtime · 2026-10-23
- Subspace Tuning curbs LLM catastrophic forgetting by freezing principal singular directions — scottniekum · 2026-10-23
- Perceptic launches PercepticOS, an AI workspace for 'scientific superintelligence' — nathanbenaich · 2026-10-23
- EPFL's Bosselut to talk long context as test-time learning at ELLIS seminar — andre_t_martins · 2026-10-23
- DRATS: Adaptive Task Sampling Boosts Worst-Task Performance in Multi-Task RL — Cohere · 2026-10-23