What just happened? A retrospective of AI alignment

Richard Ngo, a prominent AI safety researcher, is publishing a five-part retrospective on LessWrong examining the AI alignment field over the past decade. Ngo argues that the field has largely abandoned deep, generalizable scientific progress in favor of iteratively improving existing systems and pursuing political and technological power. He contends that this shift, driven by fear and self-deceptive reasoning, has had a paradoxical result: the alignment community may have accelerated AI capabilities development—particularly large language models and ChatGPT—despite its founding mandate to ensure safety. According to Ngo, the field continues pursuing strategies likely to repeat these mistakes: convincing the U-S government to prioritize AGI, conducting research that resembles capabilities work, and trusting Anthropic too much, similar to past over-reliance on OpenAI. He identifies a recurring pattern he calls 'jumping down the slippery slope'—treating outcomes as inevitable and then reasoning toward them, which inadvertently brings them about. Ngo suggests that how well the alignment community learns from these mistakes will be crucial to shaping AI's future.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton