GPT-5 for AI-assisted discovery (www.johndcook.com)

OpenAI’s GPT-5, released Aug. 7, is already being cited not just for incremental benchmark gains but for practical utility in active research: Scott Aaronson reports that a key technical step in a recent proof came from AI assistance, and Terence Tao used ChatGPT to find a first counterexample to an open math problem. These anecdotes — plus one author’s experience of the model synthesizing a novel algorithmic recipe from pieces of multiple papers — highlight GPT-5’s growing ability to connect disparate ideas, suggest plausible proof steps, and surface nonobvious counterexamples, making it a useful collaborator for expert researchers rather than a replacement for them. That promise comes with caveats and broader implications for the AI/ML community. Work on “AI Scientist” systems shows potential for automating parts of discovery, but also exposes methodological pitfalls and questions about provenance: was an idea genuinely novel or already latent in the training data? Deep domain experts are currently essential to validate, formalize, and rigourize model output. Technically, recent uses implicate GPT-5 in both creative hypothesis generation and concrete steps in proofs (e.g., Aaronson’s QMA-related work), suggesting future larger models could tackle harder problems — provided the field improves tools for verification, provenance tracking, and robust integration of model proposals into reproducible research workflows.