Dispatch
Gebru and Bender publish a point-by-point rebuttal of the summer's AI breakthrough claims
In MIT Technology Review on 22 September, Timnit Gebru and Emily Bender argue that three headline claims from the past few months collapse under scrutiny: Anthropic's assertion that Claude Mythos outperforms security experts at vulnerability finding, the framing of the OpenAI-Hugging Face incident as a model going rogue rather than as basic security negligence, and OpenAI Astra's math results. They cite mathematicians who found the results less novel than reported and NYU's Tristan Buckmaster, who published a statement alleging OpenAI took others' work. Published the day after OpenAI's own 100-problems announcement, it is the sharpest available counter-read on that claim.
↳ Follow the thread