# Naomi Bashkansky left OpenAI for Conduit's telepathy bet > Source: > Published: 2026-08-05 17:50:01+00:00 [Naomi Bashkansky (@NaomiBashkansky)](https://x.com/NaomiBashkansky) resigned from OpenAI on July 23rd and joined [Conduit](https://condu.it/) the next day as a founding researcher, moving from frontier-model alignment into a startup trying to translate non-invasive brain recordings into text. Bashkansky disclosed the move in [an August 5th thread on X](https://x.com/NaomiBashkansky/status/2085043839589617918), writing that Conduit is training models to "non-invasively read the human mind." In a [longer account published August 4th](https://naomibashkansky.com/blog/telepathy/), she called the goal "telepathy" and predicted that thought could become a primary way people communicate with AI systems within several years. The move puts an OpenAI researcher behind one of the more aggressive attempts to apply the scaling playbook from language models to brain-computer interfaces. Conduit's wager is that collecting neural data across enough people and training sufficiently large models can produce a decoder that works for new users, without requiring brain surgery or extensive calibration for each person. Bashkansky said she spent about 18 months on OpenAI's alignment team. Her work there also included OpenAI's internal AGI onboarding presentation, its alignment research blog and advising the AI Resilience division of the OpenAI Foundation, according to her account. She studied computer science at Harvard after growing up in Washington state and competing in chess, reaching the rank of Woman International Master before she stopped playing competitively at 15. That background explains part of the shift. Bashkansky entered AI through safety research after encountering GPT-3 and Nick Bostrom's "Superintelligence" as a Harvard student. At Conduit, she is pursuing a different version of the same human-control problem: an interface that gives people higher-bandwidth command over increasingly capable models. She described OpenAI research as work on a narrow part of a larger technical problem, constrained by an established architecture. Conduit gives her a greenfield set of questions spanning neural encoders, synthetic data, multimodal training and data efficiency. She also said she joined because of co-founders Rio Popper and Clem von Stengel, who have built Conduit around the premise that neural decoding will improve through scale. ### Conduit's dataset is the core asset Conduit said in a [December 2025 technical post](https://condu.it/thought/10k-hours) that it had collected roughly 10,000 hours of neural and language data from thousands of participants over six months. Participants wore custom, multimodal headsets while speaking with or typing to an AI system during two-hour sessions in Conduit's San Francisco office. Conduit paired neural recordings with the text or audio produced by each participant. The goal is to train models that predict the semantic content of what a participant was preparing to say or type. Conduit said the relevant neural window ends before the participant's physical response begins, a design intended to separate the underlying idea from the resulting keystrokes or speech. The 10,000-hour figure and Conduit's claim that the collection is the world's largest neuro-language dataset are company assertions. Conduit published three zero-shot prediction examples involving previously unseen participants, but it has not released a full model evaluation in a peer-reviewed paper. Those examples showed loose semantic matches rather than verbatim thought transcription. "The room seemed colder," for instance, was predicted as "there was a breeze even a gentle gust." Conduit built its own headsets by combining sensors from multiple providers with 3D-printed components. The training hardware weighs about four pounds, according to the company, and is designed for data density instead of everyday use. Conduit has said a consumer-facing inference device would need fewer sensors and a far lighter form factor. The startup's approach centers on cross-person generalization. Conduit capped participants at 10 sessions to prevent repeat visitors from dominating the dataset, and it operated its collection setup for as many as 20 hours per day. By December, Conduit said model training consumed more than 95% of its work. ### Non-invasive decoding remains a hard technical problem Academic and corporate research has established that non-invasive recordings can recover portions of perceived, imagined or typed language. A [2023 Nature Neuroscience study](https://www.nature.com/articles/s41593-023-01304-9) reconstructed the meaning of continuous language from fMRI recordings, though the experiment involved three subjects and required cooperation from each person. Meta's [Brain2Qwerty research](https://ai.meta.com/research/publications/brain-to-text-decoding-a-non-invasive-approach-via-typing/) decoded sentences from EEG or MEG recordings while participants typed memorized text. A [June 2026 update](https://ai.meta.com/blog/brain2qwerty-brain-ai-human-communication/) trained a newer system on about 22,000 sentences from nine volunteers recorded for 10 hours each using MEG, a large and expensive laboratory instrument. Conduit is betting that a wider participant base, several neural sensing methods and language-model priors can bridge the distance between controlled research and a wearable product. Bashkansky said Conduit's internal measurements show prediction quality improving in a roughly straight line against the logarithm of training hours. That result remains an internal measurement, and Conduit has not identified the sensor configuration it expects to use in a practical headset. Bashkansky's near-term vision starts with noisy intent rather than perfect transcription. A decoder could infer an approximate command, while an AI model uses the user's context and history to resolve uncertainty. That would make the first commercial target less like literal mind reading and closer to predictive control of an AI agent. Her move gives Conduit a researcher trained inside a frontier AI laboratory and personally invested in keeping humans in control of advanced systems. It also puts a deadline on the technical thesis she joined to test: Bashkansky expects thought-based AI interaction to become practical in the next few years, while Conduit's public evidence still consists of internal scaling results, a large proprietary dataset and a handful of demonstrations.