Reading your thoughts before you type them sounds like science fiction. Conduit, a San Francisco startup, is betting it’s an engineering problem.

Naomi Bashkansky, a Harvard computer science graduate and former OpenAI researcher, announced her departure from the AI lab in late July 2026 to join Conduit as a founding researcher. The company is building AI models that decode brain activity into text, without implanting anything in your skull.

The announcement dropped on August 5, 2026, and it landed at an interesting moment. Non-invasive brain-computer interfaces have long played second fiddle to surgical approaches in the public imagination. Conduit is making the case that the non-invasive path, combined with enough data and good enough models, can close that gap.

The dataset play

Conduit claims to have built the world’s largest neuro-language dataset, approximately 10,000 hours of recordings collected from thousands of participants over just six months inside a dedicated San Francisco facility. As of December 2025, that figure reportedly dwarfs existing public neuro-language datasets in both size and complexity.

The company collects this data using EEG, a non-invasive technique that measures electrical activity across the scalp, combined with other recording modalities. Volunteers are paid between $50 and $55 per session.

The goal is to train what Conduit calls large brain foundation models, the neural-data equivalent of GPT-style architectures. The ambition is decoding semantic content, meaning what someone is thinking, not just which direction they intended to move a cursor.

Who is building this

Bashkansky’s prior work at OpenAI centered on AI safety auditing and language model interpretability. She also published on instruction stability in language models, a corner of alignment research focused on how reliably a model follows directions under varied conditions.

Conduit was co-founded by researchers from Oxford and Cambridge, according to early coverage, giving the team a transatlantic academic pedigree alongside its San Francisco operational base. The company is explicitly organized around a data-centric model, prioritizing dataset scale as the primary driver of model performance.

What this means for the market

For investors tracking the AI-adjacent hardware space, the interesting signal here is not just one startup’s launch. It’s the pattern: serious AI researchers with alignment and interpretability backgrounds are now moving into neurotechnology, bringing with them a data-scaling playbook that worked in language modeling.

A reliable thought-to-text system would transform communication for people with ALS, locked-in syndrome, or other conditions that eliminate motor output.

Conduit has not disclosed funding figures. What to watch is model performance on semantic decoding tasks as the dataset grows, regulatory posture from the FDA on non-invasive neural data collection, and whether competing labs, academic or commercial, can close the dataset gap that Conduit is currently claiming as its primary moat.