Decision theory debate: splitting an AI into copies is value drift, not a CDT failure
jessi_cata · x · 2026-10-18
In a debate with Trotztd, jessicata argues that splitting an AI into two copies turns one utility function into two — which is directly value drift, not a flaw of CDT+GT. With utility functions defined over uncentered worlds (which matters for optimality results), you can construct scenarios where an agent modifies its own utility function and CDT+GT stops giving ex ante optimal recommendations — but the problem lies in value drift, not the decision theory itself.
Related event: Decision Theory Debate: Copying an AI Constitutes Value Drift(4 posts)→
More from AGI Musings
- AI researcher: in the agent era execution is cheap and good taste is the scarce resource — rose_e_wang · 2026-10-18
- Blogger predicts AI will fully dominate chip design within two years, like it did math — rakyll · 2026-10-18
- Luiza Jarovsky warns AI is creating a society cognitively enslaved to model owners — LuizaJarovsky · 2026-10-18
- Nathan Lambert: no evidence open weights lead AI cyberattack damage — natolambert · 2026-10-18
- Economist debate: AI-era labor shifts to human-valuable work, total labor spend won't fall — herbiebradley · 2026-10-18
- Paying per token to move a robot would be hilarious — and exactly what OpenAI and Anthropic want — philfung · 2026-10-18