A slide argues Bayes is only one lens, not the theory of modern AI inference

davidmanheim · x · 2026-07-29

• The repost reacts to a technical slide arguing that much of today’s strong AI is not truly Bayesian.
• The slide claims deep learning beat probabilistic programming, Bayesian nonparametrics, and Bayesian neural nets on most practical problems, and that LLMs can represent some PPL ideas directly while the reverse is not true.
• It also argues that deep learning is an automated way of building new continuous “theories,” something Bayes is weak at in theory.
• On the other hand, the slide says Bayesian ideas still recover several features of modern optimization—such as RMS normalization, Nesterov acceleration, AdamW, approximate inference, and deep ensembles—though often only as post-hoc or imperfect analogies.
• The conclusion is that “Bayes is one lens among several” is a more plausible worldview than “Bayes is the theory of inference.”

Original post →

More from Research

• Manski says clinical statistics research has a systemic methodological dysfunction — RexDouglass · 2026-07-29
• Can models really understand, or are we just scaling pattern matchers? — ocean_protocol · 2026-07-29
• Kimi K3’s constant-state design cuts long-context memory use by about 73% — bookwormengr · 2026-07-29
• DexRobotics open-sources a 50-episode SO101 robot fine-tuning workflow — AdinaYakup · 2026-07-29
• SmartPhotoCrafter lets a model diagnose a photo first, then auto-edit it — 量子位 · 2026-07-29
• Brain-inspired AI can plan and solve problems with far less energy, study suggests — striketheviol · 2026-07-29

Research channel →