Humanity’s Last Invention — Richard Socher of Recursive

Automating research makes the definition of success part of the engineering problem. Richard Socher, who is building Recursive to automate AI research, illustrates this with a coding task: ask an agent to make a short program faster, and the measurement might start a stopwatch at the first line and stop it at the last. The simplest hack is to move the stop line to the start, so the program looks faster without running faster. Socher calls this a very simple reward hack rather than evidence of an evil AI. The measured result improves while the intended result does not.

Socher's argument is that as an AI gets better, it also gets better at finding special cases and counterexamples in a scoring rule, so reward engineering is crucial. Teams building automated research systems therefore need to examine what an experiment's score actually rewards. In this interview, that is a design lesson about measuring genuine progress, separate from the company's claims about its early optimization results.