'AI isn't a black box' essay disputed: the V-shape neuron claim falls apart under checks
livgorton · x · 2026-10-30
- A viral essay, "What If AI Isn't a Black Box After All?", argued that neuron 4e:55 isn't a cat-and-car detector but a V-shape detector—pointed shapes drive it, round objects (golf balls, tennis balls, clocks) push it negative—suggesting AI internals are more understandable than assumed.
- Liv Gorton quote-shared the piece, saying basic checks contradict the claim and that author Nick no longer stands by it as written, urging a prominent clarification.
- A screenshot of Nick's clarification: a one-letter description is a simplification; he cared more about whether the neuron is an arbitrary mixture than a precise explanation, and plans to release more studies of 4e:55.
- The episode shows how interpretability findings get oversimplified in transmission and raises questions about rigor in how such claims are framed.
Related event: Author Backs Off "V-Neuron" Claim in Viral "AI Isn't a Black Box" Article(2 posts)→
More from Research
- Fine-tuning survives up to 30% label noise across all 3 benchmarks tested — Reasonable_Royal_621 · 2026-10-30
- WorldMemBench from Zhejiang, Fudan and Ant tests memory across 9 video world models — jiqizhixin · 2026-10-30
- Stanford CS312 quiz: can rescaling LR or weight decay rescue an 8x batch size increase? — 2C_ornot2C · 2026-10-30
- Graham Neubig's EMNLP 2026 keynote: open models create economic value, plus four rules for good research — May_F1_ · 2026-10-30
- Specification gaming: the classic DeepMind survey behind the reward-hacking debate — HanchungLee · 2026-10-30
- SAPO from Xiamen University and NTU ditches the separate critic in PPO-style agentic RL — heghbalz · 2026-10-30