Gary Marcus Calls Astra 'Amazing but Vastly Oversold': 249 Pages on the Math, Zero Pages on How the Model Works
In an August 2 post Marcus concedes the result is real but attacks the inference drawn from it as a fallacy of composition, arguing that 'success on one form of fancy cognition' does not mean 'success on all forms is imminent.' His sharpest technical objection is that math is uniquely favorable because it 'allows for external tools to do verification and to create synthetic data,' a property absent from open-ended real-world problems, and he notes OpenAI published a 249-page paper on the results where 'not one page is about how the model works, how the proofs were verified, what role if any humans played.' He stacks corroborating critics: Ernie Davis on cherry-picking and hidden human labor, Henry Yuen on Astra's proof-writing quality lagging the proofs themselves, Eric Weinstein on domains Astra likely fails, and Sabine Hossenfelder on practical failures.
Source
↳ Follow the thread