user testing the user tester - latent space ai in action presentation

Your coding agent compiles, runs every test, and reads its own traces. Tests only check that you built the spec, never that the spec was right, and the one thing the agent can't do is be your user: someone who doesn't care about your mission, arrives with their own goal, and takes a path you never thought to test. Authoring that path is the hard part. A synthetic user just walks it. So I ran the experiment twice. First, head to head with practicing UX researchers: same product, same brief, graded on accuracy and usefulness, with the researchers on record about whether synthetic research is ready to trust. Then against reality: synthetic feedback running beside a product's real user reports, the kind aggregated from every forum, Discord, and support thread it has. Does it reproduce the issues those users already found, and can it surface the next ones before a single user has to? That answer is the talk.