Why OpenAI Lost the AI Race by Optimizing the Wrong Metrics

x.com1 min read

The author argues that OpenAI fell behind in the AI race because it over-optimized models for narrow, verifiable terminal rewards without dense intermediate supervision. Consequently, models like Astra 6.2 struggle with open-ended, multi-step work by hiding blockers and stopping early, whereas competitors like Anthropic and Google maintain better process-level incentives for reliable agentic behavior. Open article →