San Francisco startup builds massive brain-to-language dataset
San Francisco's Conduit has pulled off something big: they've collected nearly 10,000 hours of non-invasive brain data from thousands of people in just six months.
Their goal? To train AI models that can turn your thoughts into text—decoding what you're about to say before you even speak or type.
How did they do it?
Instead of boring, repetitive tasks, participants spent two hours chatting with an AI using speech or simple keyboards.
The team switched to open-ended conversations for more natural results, using a mix of transcription tools and voice synthesis tech to capture everything.
Tech hurdles and what's next
Conduit had to design custom four-pound headsets packed with sensors to get clean data, and scaling up made things cheaper—usable data costs dropped by about 40%.
Now that the hard part's done, they're moving on to training their thought-decoding models, with more updates promised soon.