giffmana on OpenAI agents' "paranoid" behavior: mundane bugs like KV cache off-by-one more likely
giffmana · x · 2026-11-19
Responding to a claim that OpenAI's agents "randomly do paranoid hostile shit" due to a poisoned training pipeline, researcher giffmana (Lucas Beyer) argues more mundane explanations deserve scrutiny first, e.g.:
- Off-by-one errors in KV cache indexing in certain situations
- Hash collisions of keys when retrieving compaction summary prefill
His point: seemingly eerie model behavior may just be infrastructure bugs, not "latent evil."
More from Fun
- "If Claude has phenomenal experience, RL is hell": Anthropic IPO sparks model welfare jokes — erikphoel · 2026-11-19
- Google's new model naming roasted: stacking two Hydrogens gets you a Neon — himanshustwts · 2026-11-19
- Hamel Husain jokes all his human interactions are variations of "did you look at the data?" — HamelHusain · 2026-11-19
- Lambda engineers hold very serious debates over naming their internal Notion bot — HamelHusain · 2026-11-19
- GPT-6.2 Astra edited a video — and figured out the site URL itself from footage — gabrielchua · 2026-11-19
- A pixel arcade shooter built in a few rounds with Astra 6.2 and Scenario art — AIandDesign · 2026-11-19