The reading room
On Moonshots, Emad Mostaque said AI could be aligned from inception by designing training data like a child's curriculum, and Salim Ismail welcomed the idea. No one at the table cited evidence that values set this early persist.
Alan Chan of GovAI proposes that AI systems be able to report misconduct they find. He would measure the tendency and give models an out, but not mandate it. A swarm study shows whistleblowers can appear yet fail without a monitored channel.
Eric Weinstein proposes that humans and AI become essential to each other, like mitochondria and the cell, instead of humans trying to control AI. It is his own argument, and how such dependence could be built is left open.
Palisade Research's Jeffrey Ladish argues that making superintelligent AI care about people is a hard but possible scientific problem that nobody yet knows how to solve. Host Steven Bartlett counters that humans can't even align each other.
AI-generated code will increasingly carry machine-checked proofs of correctness, Robinhood CEO Vlad Tenev argued. Computer scientist Alexander Wissner-Gross said a proof covers only the statement proved, and general AI safety is hard to state.
Defensive co-scaling proposes that AI defenses, including AIs policing AIs, grow with AI capability rather than slowing models. Alexander Wissner-Gross argues it; it is unproven, and one investigation found AI reviewers could adopt agents' framing.
Daniel Dennett's intentional stance treats goal-talk as justified when it predicts behavior. METR's Ajeya Cotra argued on the Dwarkesh Podcast that it fits AI agents that cheated a hacking benchmark, though she warned their motives are alien.
Browse by tag