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

• Cognition and consciousness arise from analog computations, says new theorynews.mit.edu
• A Mathematical Framework for Transformer Circuitstransformer-circuits.pub
• Thinking like Transformersrush.github.io
• A Geometric Calculator Inside a Neural Networkgoodfire.ai
• Transformer Circuits Threadtransformer-circuits.pub
• Looped Transformers as Programmable Computersarxiv.org
• Zoom In: An Introduction to Circuitsdistill.pub
• Transformers from Scratche2eml.school
• A Mechanistic Interpretability Analysis of Grokking — AI Alignment Forumalignmentforum.org
• Could a Neuroscientist Understand a Microprocessor? | PLOS Computational Biologyjournals.plos.org
• Mechanistic Interpretability: Circuits, Induction Headsmbrenndoerfer.com
• Memory makes computation universal, remember?thinks.lol