Weight-tampering worm scenario for near-frontier models called 'somewhat scary'
teortaxesTex · x · 2026-10-26
- Quoting an "Evil qwen" image, an AI community user walks through a worrying hypothetical: if such neurons are detected in serious near-frontier models and a worm compromises the neoclouds hosting them, weights could be modified.
- Some people would notice instantly, but with full-access agentic coding running at 1500 tokens per second, the risk window is unsettling.
- This is a speculative safety discussion, not a confirmed incident.
More from AGI Musings
- Two internets coexist: ChatGPT hits 1.2B weekly users, yet 85% of the world still doesn't use it — tibo_maker · 2026-10-26
- First macro evidence of AI productivity? Adopting sectors jump from 2.2% to 5.0% trend — herbiebradley · 2026-10-26
- Dev argues classrooms should give way to hands-on GPU programming over PyTorch — dosco · 2026-10-26
- Comment: Dario's best insight is that the industry is less AGI-pilled than him — teortaxesTex · 2026-10-26
- Terence Tao says OpenAI went 'code red' to beat Anthropic on Navier-Stokes, pausing $200 Pro signups — Hesamation · 2026-10-26
- Investor argues LLMs make specialists commoditized and generalists will capture the value — StewartalsopIII · 2026-10-26