Toy transformers may represent belief-state geometry optimally but not minimally

·LessWrong··

Methods note: The code used for the experiments and related open-source repo were built with Claude. The experimental design and writeup is my own, with minimal editing and formatting amendments made with Claude. Thesis A toy transformer keeps provably predictively defunct belief state data in its residual stream. This information is shed only when there is a sufficient amount of imposed capacity pressure, in which case the oldest predictively defunct information is shed first. Results I conduct...

Read full article →

Related Articles

US strikes $1.2B deal to pay German firm to halt offshore wind projects
defrost · Hacker News · 22h ago
Oracle bans AI-generated code from OpenJDK
delduca · Hacker News · 15h ago
AMD acquires Taalas to boost inference performance by etching models in silicon
itvision · Hacker News · 1d ago
Making Postgres 300x faster for analytics: batching, operator fusion, and SIMD
poly2it · Hacker News · 21h ago
Qwen3.8 Max now ranked as the best overall model by agentic index
apitman · Hacker News · 1d ago