Liv Gorton: superposition has remarkable predictive power over adversarial examples
livgorton · x · 2026-10-22
- In a debate with Nick Cammarata and others, liv Gorton argues that the superposition hypothesis has remarkable predictive power over surprising neural network behaviors—claiming it explains nearly all of the surprising ways adversarial examples behave.
- In a follow-up she clarifies she'd truly love for superposition to be fake news and believes we must stay unattached to candidate explanations: she tried hard to persuade herself of the counterpoint about a specific neuron but, after running many experiments, couldn't see supporting evidence, hoping a related paper will be more convincing.
- A substantive interpretability debate about hypothesis testing methodology, with opposed viewpoints and concrete experimental grounding.
Related event: Liv Gorton Rebuts "Neural Networks Are Already Clean" Point by Point(11 posts)→
More from AGI Musings
- Nathan Clayton publicly presses Dario: Anthropic is racing toward recursively self-improving superintelligence — ShakeelHashim · 2026-10-22
- Thiel's final 'New Rome' lecture: rent-traps, stagnation and parallel tech rails — DanielFNash · 2026-10-22
- ARK analyst: marginal robotaxi rides could destabilize Uber driver network — skorusARK · 2026-10-22
- GitHub commits exploded from 1 billion to 14 billion in a single year — deliprao · 2026-10-22
- 'Generally exceptional but not superhuman AI' sparks debate: a long golden age, not a singularity — teortaxesTex · 2026-10-22
- Open source is dead in the coding-agent era: power shifts from skill to capital — MarcJSchmidt · 2026-10-22