OpenAI's amazing — but vastly oversold — new model Astra

Substack · Read

• Names the “fallacy of composition” as the core error in coverage of Astra: success on mathematics tasks doesn’t imply general competence.
• Argues Astra’s genuine strength is confined to narrowly-defined domains where verification and synthetic data generation are tractable — not evidence of broad problem-solving generalization.
• Positions this as a specific instance of his recurring critique: benchmark performance in verifiable domains gets overgeneralized into broader AGI-progress claims.