# Cammarata's interpretability essay

AI & tech · Heat 1.2 / 10 · 2 shows · 26m airtime

Nick Cammarata published the essay Neural Networks Are Already Clean on October 18, 2026, arguing that most neurons in large models are legible and that superposition findings were largely artifacts of blurry tools and old architectures. He reports that about 9 in 10 detectors in a modern OpenAI model have one main job, against 2% explainable in GPT-2 in 2023. Cammarata and colleagues are now applying the approach to current models and plan to publish technical results soon. The claim drew praise from OpenAI's Rune and skepticism from Google DeepMind's Neil Nanda, who called it interesting but not a significant update to the big picture.

Page: https://boilpoints.com/point/cammaratas_interpretability_essay_2599709068

## Segments

- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-18, 0:54 to 9:37: **Walks through the V-detector findings**. Walks through the essay's argument: the cat-face and car-front detector reinterpreted as a V-detector, three fixable causes of apparent scrambling, and roughly 9 in 10 detectors in a modern OpenAI model having one main job. Notes the counts are early and self-reported, and that the full technical paper is not out. New angle: Cammarata reports about 9 in 10 detectors in a modern OpenAI model have one main job, against 2% explainable in GPT-2 in 2023.
- [The AI Daily Brief](https://boilpoints.com/u/theaidailybrief), 2026-10-18, 12:23 to 21:07: **Recaps the essay's three claimed mistakes**. Walks through Cammarata's three claimed mistakes - the cramming story, the visualization tool, the 2014 architecture - and the pushback from Gorton and Nanda, then closes on safety implications and reactions. Retells the essay's argument and stays inside the day's coverage.
- [Daily Tech News Show](https://boilpoints.com/u/dailytechnewsshow), 2026-10-20, 18:07 to 26:21: **Walks the four-step argument and next steps**. Walks through the four-step argument in Cammarata's essay and the cat-and-car neuron recast as a V-shape detector, then notes the mixed reception. Adds that Cammarata and colleagues are now applying the approach to current models and plan to publish technical results soon. New angle: Cammarata and colleagues are applying the approach to current models and plan to publish technical results soon.