“On the SWE-Bench benchmark, GPT-4's performance jumps from approximately 2% to a range of 14-23% when using Devin's agent scaffolding.”