Overview
- A San Francisco startup called Conduit has made public claims that it built a large non‑invasive neuro‑language dataset and plans to train 'large brain foundation models' to translate brain activity into text.
- Naomi Bashkansky, a former OpenAI researcher known for work on model interpretability and safety, has joined Conduit as a founding researcher to help develop thought‑to‑text models.
- Conduit says its dataset totals roughly 10,000 hours of EEG and multimodal recordings from thousands of volunteer sessions collected in a dedicated San Francisco facility and that participants were paid about $50–$55 per session.
- Key open items include independent verification of the dataset and any decoding benchmarks, the company’s undisclosed funding and runway, and how regulators such as the FDA will assess scaled non‑invasive neural data collection and semantic‑decoding claims.
- If proven, the approach could expand communication options for people with severe motor impairments but it also raises clear issues about consent, privacy, participant safeguards, and whether high data scale can close the accuracy gap with invasive implants.