Naomi Bashkansky spent about a year and a half working as an alignment researcher at OpenAI. On July 23, she resigned from the company. The next day, she joined a startup called Conduit as a founding researcher.
At Conduit, Bashkansky will work on models designed to turn non-invasive neural recordings into text. That text could then be used to direct AI agents. She described the goal as building a kind of "telepathy" between people and machines.
She shared details of her new role in an essay published on August 4. In it, she laid out a rough timeline for how this technology might develop over the next decade.
What Bashkansky Predicts
Bashkansky believes a headband could decode rough intentions into prompts for an AI coding agent as early as 2027. She sees this as an early step rather than a finished product.
Her longer-term predictions stretch further out. She envisions AI systems that consume neural representations directly by 2030.
By 2035, she predicts two-way "read and write" technology could exist. This would allow information to move in both directions between a person's brain and a machine.
Bashkansky was careful to frame these timelines as optimistic forecasts. Her essay includes no promise that any of these systems will actually launch on schedule.
Conduit's Data and How It Compares
Conduit shared details about its data collection efforts in a December 2025 account. The company said it had gathered roughly 10,000 hours of neuro-language data from thousands of people.
Participants wore multimodal headsets during sessions with a language model. During these sessions, they typed, spoke, read, or listened while their neural activity was recorded.
Conduit has published a small number of claimed zero-shot examples from this work. The company has not released aggregate performance metrics or details of its evaluation process.
It also has not shared results from any third-party replication of its findings. Bashkansky has argued that Conduit's results tend to improve as more training hours are added to the dataset.
She described the work as a greenfield alternative to the narrower research paths available to her at OpenAI. Other groups are pursuing similar goals using different methods.
Meta reported in June that its latest Brain2Qwerty system reached 61% average word accuracy across participants. Its best individual participant reached 78% accuracy.
That experiment recorded nine people using magnetoencephalography while they typed sentences. Meta said performance improved log-linearly as more data was added.
A separate study published in Nature Communications looked at 723 participants using EEG and MEG recordings. That study also found that performance improved with more data.
However, the Nature Communications study focused on people reading or listening to language rather than producing it. It reported 20% top-1 accuracy in a 50-word comparison test.
The study's authors said that practical, non-invasive brain-to-text technology remains an open challenge. They did not claim to have solved the problem.
Conduit's next step is to show public aggregate performance data linking its 10,000-hour dataset to a working, portable system. That system would need to decode free-form thought rather than constrained tasks. As of now, the company has not published that result.
