so @conduitpbc spent 6 months running a basement lab 20 hours/day to collect the largest neuro-language dataset ever built—~10,000 hours from thousands of people. the wild part isn’t the model. it’s the operations. turns out: •⁠ ⁠noise matters < 4k hours; after that, scale eats it •⁠ ⁠dynamic overbooking beats every recruitment strategy •⁠ ⁠a 4-pound helmet is fine if you cut custom polygonal padding •⁠ ⁠rewriting the backend cut marginal cost/hour by ~40% •⁠ ⁠people will happily wear multimodal headsets if the LLM talks to them, not at them •⁠ ⁠zero-shot decoding works on people the model has never seen: thought → text: “the room seemed colder” → “there was a gentle breeze” “do you have a favorite app?” → “any favorite robot?” the future of thought-to-text isn’t sci-fi. it’s logistics, padding foam, airflow pipes, and a booking algorithm. full post: condu.it/thought/10k-hours