Ex-OpenAI Researcher Predicts Brain-Controlled Coding by 2027

Naomi Bashkansky resigned from OpenAI on July 23 after about 18 months and joined Conduit the following day as a founding researcher. At Conduit she will work on models that convert non-invasive neural recordings into text that can direct AI agents.

In an August 4 essay she wrote a headband could decode rough intentions into prompts for an AI coding assistant by 2027. She outlined later stages in which AI would consume neural representations directly by 2030 and where two-way read-and-write interfaces could appear by 2035, and described those timelines as optimistic sketches without firm launch dates.

Conduit reports it has gathered roughly 10,000 hours of so-called neuro-language data from thousands of participants. During recordings volunteers wore multimodal headsets while typing, speaking, reading or listening in sessions that involved interaction with a language model. The company has published a few claimed zero-shot decoding examples but has not released aggregate performance metrics, an evaluation protocol or independent replication for open-ended thought decoding.

Current public non-invasive research tends to decode constrained, speech-related brain activity rather than unconstrained thought. Portable systems that translate free-form neural signals into usable text for general use have not been demonstrated at scale.

One small experiment using magnetoencephalography recorded nine people while they actively typed and reported improving word accuracy as training data grew. A larger study using EEG and MEG data from more than 700 participants tested reading and listening tasks and reported about 20% top-1 accuracy in a 50-word comparison when participants were not producing language. A commercial research program reported average word accuracy in the low 60s for its dataset and higher scores for its best participant, with accuracy improving roughly in proportion to data volume.

Conduit’s immediate technical challenge is to show that its 10,000-hour dataset can support a portable device that reliably decodes unconstrained thoughts into actionable text across users. Meeting the 2027 headband timeline would require publishing standardized performance metrics, describing an evaluation method and allowing third-party verification of real-world decoding performance.

In her essay Bashkansky used the word “telepathy” to describe the project of turning neural signals into language for AI agents. She framed later stages of direct neural consumption and bidirectional interfaces as speculative developments that would take years to realize.

Conduit has not provided public evidence that its approach decodes free-form thought at scale. The company has not announced commercial release dates tied to the timelines presented in Bashkansky’s essay.