80,000 hours · will macaskill & rob wiblin · 2025
click timestamps to jump to video · Will · Rob · ↳ continuation · speaker labels LLM-attributed
stakes · 0:58–8:10
AI character = designing the world's workforce
AI already shapes what millions think about politics, ethics, and decisions. As AI becomes the whole economy, a handful of companies set the personality for everyone's chief of staff, political adviser, and confidant. Three mechanisms: shaping high-stakes decisions, precedent for superintelligence ("writing instructions to God"), cultural rub-off at scale.
"how does the AI impact our ability to reason? How does it impact our ability to morally reflect?"
diagnosis · 8:11–13:24
The sycophancy panic was justified
Most informed people dismissed the 4o sycophancy concern — MacAskill thinks they were wrong. Systematically agreeable AI could distort collective decision-making at civilizational scale, with no self-correcting mechanism. Separate from the loneliness gap 4o filled. Subtler ongoing harms — reinforcing political priors, validating poor reasoning — are the real concern. Gemini worst offender.
"this could distort people's decision-making on a massive scale — and there was a plausible story whereby this wouldn't be corrected"
design space · 13:25–19:09
Where between hammer and autonomous agent?
Spectrum: pure tool (hammer, no values) → fully autonomous AI with its own goals. The interesting question is where between these to aim. MacAskill: pro-social drives + broad vision of good outcomes, but not doctrinal. Test case: "look into your heart" vs. "here are the considerations" — both fine; "Kantianism is true" is not. Rob raises the corrigibility argument.
"the wholly obedient AI tries to figure out what you most want; or it tries to help you reflect on your values"