Peter Wang on X: "https://t.co/NGA65kLw3W" / X

https://t.co/NGA65kLw3W

Having spent 10+ years in systems neuro, I can now make twitter posts about how transformers compute like real neurons. This is my favorite case: both use cyclic representations to perform arithmetic and vector computations. @NeelNanda5 et al. had a fun mech interp paper 3 years ago that reverse engineered what a transformer learned to perform modular arithmetic. The task was addition modulo 113. In essence, the transformer mapped two discrete symbols to positions on circles, added their angles, and converted the resulting angle back into a symbol. The model represented each residue in…

saved by

• Alima Deen

related reading

• A Mathematical Framework for Transformer Circuitstransformer-circuits.pub
• Thinking like Transformersrush.github.io
• Circuit Tracing: Revealing Computational Graphs in Language Modelstransformer-circuits.pub
• Looped Transformers as Programmable Computersarxiv.org
• A Geometric Calculator Inside a Neural Networkgoodfire.ai
• A Mechanistic Interpretability Analysis of Grokking — AI Alignment Forumalignmentforum.org
• What Is ChatGPT Doing … and Why Does It Work?-Stephen Wolfram Writingswritings.stephenwolfram.com
• Could a Neuroscientist Understand a Microprocessor? | PLOS Computational Biologyjournals.plos.org
• [2602.16849] On the Mechanism and Dynamics of Modular Addition: Fourier Features, Lottery Ticket, and Grokkingarxiv.org
• The World Inside Neural Networksgoodfire.ai
• [2301.05217] Progress measures for grokking via mechanistic interpretabilityarxiv.org
• Can LLMs Be Computers?percepta.ai