Former OpenAI alignment researcher Naomi Bashkansky mentioned she resigned from the AI firm on July 23 and joined Conduit the subsequent day as a founding researcher. At Conduit, she’s going to work on fashions designed to show non-invasive neural recordings into textual content that may direct AI brokers, a purpose her essay calls “telepathy.”
Bashkansky mentioned she spent about 1.5 years at OpenAI, and described the brand new function in an Aug. 4 essay. She predicted {that a} headband may decode tough intentions into prompts for an AI coding agent in 2027.
Her later eventualities envision AI techniques consuming neural representations immediately by 2030 and two-way “learn and write” expertise by 2035.
She known as these vignettes optimistic predictions, and the essay consists of no launch dedication for any of them.
The information-scale guess
In a December 2025 account, Conduit mentioned it had gathered roughly 10,000 hours of neuro-language information from hundreds of individuals. Members wore multimodal headsets whereas typing, talking, studying or listening throughout periods with a language mannequin.
The corporate printed just a few claimed zero-shot examples, excluding combination efficiency metrics, its analysis protocol or third-party replication. Bashkansky argued that Conduit’s outcomes enhance with rising coaching hours and described the work as a greenfield various to the narrower analysis she may pursue at OpenAI.
Meta reported in June that the outcomes of its newest Brain2Qwerty reached 61% common phrase accuracy and 78% for its finest participant, with efficiency enhancing log-linearly as information elevated.
The experiment recorded 9 folks with magnetoencephalography whereas they actively typed sentences.
The place present techniques fall brief
A Nature Communications research spanning 723 members additionally discovered that efficiency improved with extra EEG and MEG information. But it studied folks studying or listening fairly than producing language, and reported 20% top-1 accuracy in a 50-word comparability.
The research’s authors mentioned sensible non-invasive brain-to-text remained an open problem.
Conduit’s subsequent evidentiary hurdle is to determine public combination efficiency by linking its 10,000-hour dataset to a transportable system that decodes free-form thought. The corporate has not but proven that outcome.
