Naomi Bashkansky resigned from OpenAI on July 23 and joined Conduit the next day as a founding researcher. After about 1.5 years at OpenAI, she's now pushing a much more radical idea: a headband that decodes rough intentions into prompts for an AI coding agent by 2027.
The Timeline Bashkansky Lays Out
Bashkansky's predictions stretch further. By 2030, she envisions AI systems consuming neural representations directly. By 2035, she imagines two-way read/write technology. She's careful to call these optimistic, with no launch commitment. The near-term goal is a wearable that turns a fuzzy thought like "make the button bigger" into a prompt a coding agent can execute.
Why Thought-to-Text Is Hard
Current thought-to-text studies only understanding constrained, speech-related brain activity. Portable free-form communication remains unproven. That's a big gap. A headband reading intent while you sit at a desk is a different problem from understanding a specific word someone is trying to say.
Conduit's Data vs. Its Evidence
Conduit reported gathering roughly 10,000 hours of neuro-language data from thousands of people, as of a December 2025 account. But the company has published only a few zero-shot examples. There's no aggregate performance metric, no evaluation protocol, and no third-party replication. That leaves outsiders guessing how well the system actually works.
What Peer-Reviewed Research Shows
Meta's Brain2Qwerty project gives a sense of the bar. It achieved 61% average word accuracy, and 78% for its best participant. Performance improved log-linearly as data increased. That experiment recorded nine people with magnetoencephalography while they actively typed sentences. A separate Nature Communications study with 723 participants found performance improved with more EEG and MEG data, but it studied reading and listening, not producing language. Its top-1 accuracy was 20% in a 50-word comparison. The authors said practical non-invasive brain-to-text remains an open challenge.
The Hurdle Ahead
Conduit's next test is to link its 10,000-hour dataset to a portable system that decodes free-form thought — and to show that in public, with numbers anyone can check. Until then, the headband stays in the realm of prediction.
