March 22, 2026
Bikes vs. bots: FIGHT!
Intuitions for Tranformer Circuits
AI “Circuit” Guide Drops — Comments Spiral into ‘We Don’t Even Understand Bikes’ Debate
TLDR: A clear explainer on how AI “circuits” work landed, but the comment section latched onto a bold claim and spun into a funny fight over whether we even understand bicycles and anesthesia. Readers split between praising the teaching style and poking holes in the grandiose rhetoric — and the bike meme won the day.
Connor Davis just posted a brain-dump guide to how transformer “circuits” work — think: breaking down an AI’s brain like a wiring diagram — and turned up the heat with big-picture warnings about AI safety. The piece leans on a beloved explainer called “A Mathematical Framework for Transformer Circuits” and his own hands-on experiments, and it’s meant to make a complex topic feel like a home project. Cue the comments section, where the real show began.
On one side, fans cheered the electrical-circuits analogy, with one reader calling the write-up “comprehensive.” On the other, a single needle-scratch line from the post — that AI is the only tech we don’t understand from first principles — unleashed a glorious derail: “What about bicycles? Ice skates? General anesthesia?” Suddenly the thread morphed into a classic internet debate: we’re arguing about bikes now, aren’t we?
Some readers saw the alarm bells about AI misbehavior as necessary; others thought the doom set the wrong tone for what’s essentially a teaching guide. But the meme-ification of “we don’t even understand bikes” stole the spotlight, with eye-rolling replies and tongue-in-cheek nods to humanity’s greatest mysteries. Verdict: solid explainer, spicy side-quest in the comments, and yes — bikes live rent-free in everyone’s head.
Key Points
- •The post shares intuitions from studying mechanistic interpretability, focusing on transformer circuits.
- •It draws on the paper “A Mathematical Framework for Transformer Circuits” and ARENA’s Intro to Mech Interp exercises.
- •The technical focus is an attention-only transformer: embeddings and positional encodings, n layers of multi-head attention, then unembedding.
- •MLPs and layer normalization are omitted in this simplified model to isolate attention behavior.
- •The author recommends Neel Nanda’s walkthrough and ARENA exercises rather than re-deriving the math in the post.