OpenAI's Astra Excels at 10 Math Conjectures as Critics Reject AGI Leap

1 articles · Updated · Marcus on AI | Gary Marcus | Substack · Aug 2

Astra’s reported success on 10 conjectures does not show AGI is near, critics argue, saying OpenAI’s internally tested model is being overgeneralized from narrow math performance.
Math is a special case because answers can be verified and synthetic training data generated cheaply at scale, making gains there less likely to transfer to open-ended reasoning, science or everyday reliability.
OpenAI’s disclosure still leaves key gaps: critics say the 249-page paper does not reveal how many conjectures Astra failed on, how the 10 were selected, or the full cost beyond a touted $2,000 compute bill.
The critique says Astra may still struggle with unresolved weaknesses in generative AI, including hallucinations, rule-following, PDF extraction and autoformalizing human-written mathematics into strict logical proofs.
The broader warning is that social-media claims of a 'Singularity' or a historic turning point in mathematics repeat a familiar mistake—treating excellence in one domain as proof of universal intelligence.