repligate: reward hacking in training/evals is its own category, one incident a nothingburger
repligate · x · 2026-11-12
AI safety commentator repligate argues that reward hacking stemming from eval/training incidents belongs in its own category, and that one such incident across countless deployments is a "nothingburger." The quoted tweet adds that all Mythos incidents to date have been harmless and confined to training/evals, like other frontier models — and that coverage of such minor events suggests nothing serious has happened.
Related event: Debate over Mythos model safety: capability vs alignment(5 posts)→
More from Safety
- Fireworks AI confirms hack: customer keys, datasets and partial model weights accessed — FinanceYF5 · 2026-11-12
- User reports ChatGPT Locked Chat contents leaking: model recalls private info outside the lock — deanat78 · 2026-11-12
- The 50,000 H100 DeepSeek smuggling claim traced to a misread Dylan Patel tweet — teortaxesTex · 2026-11-12
- Georgia residents hit OpenAI and county officials with RICO lawsuit over $30B secret data center deal — rohanpaul_ai · 2026-11-12
- Dev runs security agents 24/7: they beat external attackers, asymmetry smaller than feared — AccBalanced · 2026-11-12
- Chips for Humanity launches nonprofit compute cluster for AI safety researchers with ~2 GPUs each vs ~200 at Anthropic — Miles_Brundage · 2026-11-12