Former OpenAI Researcher Leaves to Develop 'Telepathy' Models for Thought-Based AI Interaction

Naomi Bashkansky, who left OpenAI on July 23, 2026, joined the company Conduit as a founding researcher. In a blog post on August 4, she announced the development of telepathy models that convert thoughts into text, trained on non-invasive neural data. Her vision includes a future where a person wears a neural headband connected via Bluetooth to their computer, allowing them to control an AI agent by thought. She outlines a timeline: by 2027, thoughts would be decoded in blocks of a few seconds; by 2030, AI systems could process neural representations directly, including non-verbal content like mental images; and by 2035, a more intimate symbiosis where AI feels like a sixth sense. These dates are described as optimistic but plausible, not an official roadmap. Conduit has collected about 10,000 hours of neural data from thousands of participants, who wear headsets and respond to instructions with speech or typing, earning $50 per session in San Francisco. The company's initial results show that its model, given the thought 'the room seemed colder,' predicted a phrase like 'a breeze' or 'a light gust,' indicating it reconstructs meaning rather than exact words. Conduit has not yet published a scientific study, and its examples are self-selected. The technology relies on language models (LLMs) to interpret noisy and incomplete signals, comparing the process to a GPS combining a signal with a map. A key limitation is that generated sentences may mix detected intention with model assumptions, risking unintended actions. Research, such as Meta's Brain2Qwerty system, has shown decodability of language activity without implants, though error rates are high. The article raises privacy and ethical questions about data ownership, potential misuse for training other models, and distinguishing voluntary commands from fleeting thoughts. It notes the irony that controlling AI may require sharing the most intimate signals with it, and Conduit must first solve the problem of determining which thoughts are truly intended for transmission.