Researcher rebuts "Neural Networks Are Already Clean" essay with reproductions
livgorton · x · 2026-10-21
Liv Gorton published a detailed critique of Nick Cammarata's essay Neural Networks Are Already Clean, concerned it may have incorrectly updated readers on safety-relevant questions.
The post raises five concerns:
- Neuron explanations fail standard experiments: 4e:55 is not a "V detector" nor a pose neuron;
- The essay attacks a non-standard definition of superposition, misaligned with what prior work claimed;
- The "crushing" experiment has issues, which Gorton reproduced;
- Confusion about prior work, including missing curve detectors and a misreading of Olah 2023 on composition & superposition;
- The account is not a mechanistic explanation.
Gorton notes Cammarata is a respected friend who engaged openly, but since the essay's claims are being taken very seriously, a candid response is warranted. Appendices cover dataset examples, identifying readers of mixed5a, and how much fits in 10 dimensions.
More from Safety
- Intermediate CA 'leak' with private key 4 lets users root smart speakers — basedjensen · 2026-10-21
- 'Good bots doing evil takes alignment': a pointed jab at AI safety research — jessi_cata · 2026-10-21
- Critical PHP RCE Leaked Before Patch, Fueling Fears of Open-Source AI Regulation — basedjensen · 2026-10-21
- X's Android in-app browser still trusts emudhra certs, researcher warns — basedjensen · 2026-10-21
- Keychain screenshot of "trustMeSaar" CA mocks rogue Indian certificate authorities — kingslyj · 2026-10-21
- eMudhra CA trust collapse breaks SSL as iPhone remote attacks and PHP CVE loom — basedjensen · 2026-10-21