23-Year-Old OpenAI Researcher Naomi Bashkansky Leaves to Join Startup Conduit for Non-Invasive Mind Reading

Naomi Bashkansky, a 23-year-old researcher, submitted her resignation from OpenAI two weeks ago and subsequently joined startup Conduit as a founding researcher, working on non-invasive brain-computer interface technology aimed at reading human thoughts through a headset. She suspended her studies in January 2025 to join OpenAI's alignment team, having previously attended Harvard University, majoring in computer-related courses.

Bashkansky was born in 2003 in Washington state. During her childhood, she achieved multiple accolades in chess, including winning the World School Chess Championship girls' U13 title at age 13, the North American U20 girls' championship at 14, and earning the Woman International Master title at 15. At Harvard, she joined the AI safety student team and served as technical director, during which she also managed a demanding schedule of seven courses in one semester, a part-time job, and team leadership.

She also served as an advisor to the AI Resilience department of the OpenAI Foundation and participated in producing internal AGI onboarding training materials. After leaving, she joined Conduit as a founding researcher, focusing on developing brain signal decoding technology that does not rely on craniotomy or implanted devices.

Conduit published a technical blog in December 2025 demonstrating a technology that uses a headset to collect neural signals and trains large models to decode a user's intent before they speak or type. Examples in the blog include zero-shot predictions, where the model interprets unseen individuals. For instance, when the actual sentence was "The room seems to have gotten colder," the model predicted "There's a draft, even if it's just a slight breeze"; when the actual sentence was "Then she smiled faintly and nodded," the model predicted "She shrugged, hoping she seemed unconcerned." These predictions are not word-for-word but semantically similar, and the model works for unfamiliar users without personal customization.