BrowseComp vs The Bots that Bluff
Can AI actually read the internet, or is it just faking it with confidence? In this high-voltage episode, host Emily Laird cracks open BrowseComp, OpenAI’s benchmark built to test whether web-browsing agents can find facts that are hard to uncover but easy to verify. Humans had two hours per question and still bailed most of the time, so what does it mean when a model claims victory? From compute budgets and canary strings to the rise of multimodal chaos, Emily exposes the difference between sounding right and being right, and why in an era of polished, source-backed answers, persistence…
The skinny
The skinny isn't ready yet — notes appear once the transcript is processed.