OpenAI's Astra: Math Wins, Overhyped Claims

OpenAI's Astra reportedly excels at math benchmarks, but Gary Marcus flags a fallacy of composition here: strong at one type of math doesn't mean strong at science or general cognition. Math is an easy target because it's symbolically verifiable and cheap to generate synthetic training data for—open-world problems don't work that way. OpenAI hasn't published methodology, so the actual significance is unverifiable right now.