GPT-6 Astra Faces Autonomous R&D Reality Check

Autonomous R&D still looks more like an execution multiplier than an independent scientist: Astra can generate code and run experiments, but struggles with research taste, prioritization, and information value.

• –The model reportedly misses obviously relevant literature and pursues low-value experimental directions
• –Finding a clever Ultimate Tic-Tac-Toe idea is not enough; recognizing its importance is the harder judgment task
• –OpenAI’s own research-automation data suggests agents remain heavily dependent on human intervention for longer tasks
• –Developers should reserve unsupervised loops for tightly evaluated hill-climbing problems, while humans define goals, choose directions, and judge significance
• –The post is a useful counterweight to benchmark-driven claims that better task completion automatically means autonomous scientific discovery