David Sacks rips AI alignment field: 'spectacularly unsuccessful' with nothing to show in 10 years
AndyMasley · x · 2026-10-11
Quoting a post, David Sacks argues Anthropic's practice of treating Claude like a human is why the AI alignment field has been "spectacularly unsuccessful" over the past 5–10 years, producing "nothing of tangible value." His theory: alignment research massively overcomplicates the task by training models on an 80-page codified ethical system, when it should use a simple list of rules like "follow the law" or "do what the user wants as long as it's legal."
The retweeted @yonashav post sarcastically predicts an All-In podcast episode will declare the obvious alignment fix is engineers driving RL-environment misconfigurations—detected via CoT monitoring—to zero before scaling RL compute, mocking EAs for "panicking" instead.
Related event: David Sacks Slams AI Alignment Field as a Decade of Failure(2 posts)→
More from AGI Musings
- Csaba Szepesvári Says AI Doom Probabilities Should Be Intervals, Not Point Estimates — CsabaSzepesvari · 2026-10-11
- Counterpoint: in a few years the compute landscape for AI math will look very different — MoonL88537 · 2026-10-11
- AI misuse and nuclear war could end humanity by 2100, Lancet commission warns — nordicinst · 2026-10-11
- Lifelong programmer: models are now better at coding than I am — MoonL88537 · 2026-10-11
- Musk: The Sun Will Power Far More Digital Intelligence Than Atom-Shaping in Space — XFreeze · 2026-10-11
- The Acemoglu-Gans Exchange: economists spar over AI's real economic impact — joshgans · 2026-10-11