Sam Altman promised billions for AI safety. Here’s what OpenAI actually spent.

An 18-month New Yorker investigation reveals a gap between Sam Altman's public commitments to AI safety and OpenAI's actual follow-through. Key findings include Altman's dismissive stance on hallucinations as desirable 'magic,' OpenAI's superalignment team receiving only 1-2% of compute despite a pledged 20% and being dissolved by May 2024, and internal safety reviews for GPT-4 that board members found incomplete. The piece also covers sycophancy as a structural flaw in RLHF-trained models and the risks of deceptive alignment for developers deploying LLMs in production.