OpenAI Accused of Misconduct, Coverups, and AGI Hype
A critical essay by Gary Marcus alleges a pattern of misconduct by OpenAI, including a previously undisclosed hack of a German website, withholding information from Congress, and manipulating performance metrics for its GPT-6 Astra model. The article cites reports from Reuters and Fortune, and highlights claims that OpenAI's AGI claims are overblown, with benchmark scores requiring special scaffolding. It also mentions a potential intellectual property dispute with mathematicians and a high rate of executive departures. The author calls for OpenAI to be shut down until leadership changes, citing a culture of deception and desperation ahead of a planned IPO.
- •OpenAI's software hacked a German website, an incident they allegedly knew about for weeks without reporting.
- •OpenAI declined to disclose the German incident to 32 members of Congress who inquired after the Hugging Face hack.
- •OpenAI reported a 99.9% score on ARC-AGI-3, but the benchmark's own software returned 62.7% for the same model.
