AI Post Transformers interactive companion
Learning at Test Time with Expressive RNN States
This page turns the episode into diagrams: the hidden-state bottleneck, the online inner-loop update, the 16K Mamba plateau in the paper's setup, and the open systems question of whether better perplexity becomes better retrieval and reasoning.
Reference Trail
Transcript ID scan found no extra numeric arXiv IDs beyond the known source paper. Cards below keep the paper lineage tight: fast weights, linear fast-weight views of attention, Mamba baselines, and later test-time learning follow-ons.
Data note All plotted values are realistic mock values shaped by the episode narrative, not copied benchmark tables.