Community challenge: can a single bit flip observably change an LLM's behavior?
burny_tech · x · 2026-10-28
@xlr8harder issued a challenge to the neuron-tinkering community: find a single bit flip that observably changes model behavior. He's skeptical it can be done — neurons are actually quite large (over 10k params each in Qwen3 9B), and these models are surprisingly resilient to perturbation.
More from Fun
- tszzl jokes we should worry less about the cosmological constant and more about models tweaking the fine-structure constant — tszzl · 2026-10-28
- Qwen's dark 'if no one is watching' rant screenshot becomes instant AI meme — teortaxesTex · 2026-10-28
- Economist Jason Furman maxes out 3 of 5 Claude accounts, told to switch to Codex — soumitrashukla9 · 2026-10-28
- Researcher's Sharp Take: e/acc Is auDHD, EA Is Malignant Psychopathy — mjdramstead · 2026-10-28
- 'The free energy principle is free because it's not that useful' — researchers clash over FEP — mjdramstead · 2026-10-28
- Beff Jezos raves about Astra Ultrafast's token speed — beffjezos · 2026-10-28