Why I'm leaving OpenAI to build telepathy

Naomi left OpenAI on a Thursday and started at Conduit the very next day, which tells you something about how sure she is about this bet. Conduit's goal is blunt: read thoughts non-invasively and turn them into text an AI can act on. Not sci-fi mind control, she insists, just a headband that picks up neural signals and a model that guesses what you meant to say, then routes it to something like Codex.

Her 2027 scenario is oddly mundane, which is the point. You glance at a plot, feel a flicker of annoyance at bad labeling, and a subagent quietly fixes the legend while you make coffee. No typing, no talking out loud, just noticing things and having the system react in ten-second chunks. By 2030 she expects AI labs to train models that read Conduit's neural latents directly, skipping the text-decoding step entirely, which would let people send mental images and vague intuitions instead of clean sentences. That's also when Conduit reportedly branches into invasive reading and, more strikingly,