For decades, brain-computer interfaces that require no surgery have run into the same wall. The signals a headset picks up from outside the skull are simply too noisy to reliably decode what a person means to say. On August 5, a 23-year-old alignment researcher named Naomi Bashkansky announced she had left OpenAI after 18 months to bet that this wall has quietly come down, and that the thing that knocked it over was the large language model.
Her argument, laid out in a blog post the day before, is worth taking seriously because it describes a real computational structure rather than a slogan. Think of GPS. A weak satellite signal cannot pinpoint your location on its own, but combined with a map, a model of where roads exist and where you were a moment ago, it becomes accurate navigation. Bashkansky says a noisy brain signal works the same way. On its own it cannot pick which of thousands of words you are forming. Paired with a language model, a probabilistic map of which words plausibly follow from context, the decoder's search space collapses to something manageable. The receiver never needed a stronger signal. It needed a smarter prior.
Her new employer, a San Francisco startup called Conduit, has spent six months building the data to test that idea. By its own account it has collected roughly 10,000 hours of non-invasive neural recordings from thousands of people, running phone-booth enclosures 20 hours a day and paying participants recruited on Craigslist 50 dollars for a two-hour session of free conversation with a chatbot. The company says it found a threshold familiar from speech AI: below about 4,000 to 5,000 hours, noise is a critical problem; above it, the sheer volume of data starts to overwhelm the noise on its own, without explicit filtering.
The candour about how far this still has to go is what makes the pitch credible. Conduit's own published examples are semantic near-misses, not transcription. When a participant was about to type "the room seemed colder," the model guessed "there was a breeze, even a gentle gust." Bashkansky calls this the "GPT-2 era" of neural decoding: primitive, but far enough along to see the scaling behavior. The best peer-reviewed non-invasive result to date still gets nearly a third of characters wrong, and it needs a magnetically shielded room and equipment worth hundreds of thousands of dollars. Skeptics from the surgical-implant world, who have watched EEG decoding disappoint for years, are unconvinced that more data changes the underlying curve.
What makes her move notable is not just the destination but the reasoning. The recent exodus from OpenAI and Google DeepMind has mostly flowed to rival model labs. Bashkansky went to hardware. And she frames it as an alignment argument: if the main interface between a person and an AI becomes passive decoding of what the human intends, rather than an agent pursuing goals of its own, the human stays in the loop by design. That claim is unproven, and so is the whole enterprise, which has published no benchmark and disclosed no funding. But it reframes what the LLM era might be about. Not only what these models can write or code, but what they might one day read directly from a mind that has not yet found the words.