Cognition Launches SWE-1.7, Defying AI Training Limits!

Cognition unveiled its most advanced AI model, SWE-1.7, challenging the belief that reinforcement learning (RL) returns diminish after extensive training. The model, available via the Devin platform, achieved significant gains over its predecessor, showcasing promising performance metrics across various benchmarks. Cognition suggests that infrastructure quality, not just model training, influences RL outcomes. The research community will verify these claims in upcoming experiments.