Former OpenAI Researcher Joins Conduit's Thought-to-Text Project
Former OpenAI alignment researcher Naomi Bashkansky joined Conduit as a founding researcher in late July, according to reports published August 5 and 6. The move renews attention on Conduit's December 2025 claim that it collected roughly 10,000 hours of non-invasive neuro-language data from thousands of participants; the company has published examples but not an independent benchmark or peer-reviewed evaluation.
Former OpenAI alignment researcher Naomi Bashkansky joined San Francisco startup Conduit as a founding researcher in late July, according to reports published on August 5 and 6. The move brings new attention to Conduit's effort to train systems that map non-invasive neural recordings to text, but it does not mark a new dataset release or a validated product launch.
The dataset predates the hiring news
Conduit disclosed its data-collection program on December 7, 2025. In that first-party post, the company said it had collected roughly 10,000 hours of recordings from thousands of participants over six months. Participants sign a consent form, wear a custom multimodal headset, and spend two hours speaking, listening, reading or typing during a conversation with a language model. The company says each session creates neural data aligned with text and audio.
Conduit describes the corpus as the largest neuro-language dataset it knows of, but that is a company claim rather than an independently audited measurement. Its public post provides several zero-shot examples for previously unseen participants, yet it does not report a standard benchmark, error distribution, statistical analysis or peer-reviewed comparison.
What the model must prove
Conduit says its system aims to decode semantic content from the moments before a participant speaks or types. Its collection hardware combines multiple non-invasive sensing modalities; the company discusses EEG as one example but does not disclose every sensor configuration. That distinction matters because the quality and practical constraints of non-invasive signals differ substantially across modalities.
The company also reports that performance became less sensitive to environmental noise after roughly 4,000 to 5,000 hours of training data. This scaling observation is not yet independently reproduced. Fluent output alone would not establish that a model recovered a participant's intended meaning, because a language model can generate plausible text from contextual priors even when the neural signal is weak.
For a credible technical evaluation, Conduit would need pre-registered tasks, held-out participants, baselines that remove or shuffle the neural input, and metrics that separate semantic recovery from language-model completion. It would also need clear reporting on consent, retention, deletion and downstream use of neural recordings.
Bashkansky's arrival is therefore best understood as a research-team development around an ambitious existing program. The public evidence shows a large collection operation and early company-selected examples, not a demonstrated consumer thought-to-text interface.
Key Points
- 1Naomi Bashkansky joined Conduit as a founding researcher in late July, according to reporting published August 5 and 6.
- 2Conduit's roughly 10,000-hour neuro-language dataset was disclosed in December 2025 and remains a company-reported figure.
- 3The company has not published a peer-reviewed benchmark showing that its system reliably recovers intended meaning from held-out participants.
Scoring Rationale
The researcher move and unusually large company-reported neural dataset are notable for non-invasive BCI research, but practical impact remains unproven because Conduit has not published an independent benchmark, peer-reviewed evaluation, or consumer-ready system.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
