Hemispheric raises $52M for its Descartes brain AI model

A NeuroAI startup that spent more than six years in stealth just walked into the open with a bet that the scaling recipe from language models will work on brain activity. Wired reports that Hemispheric, cofounded by Face ID co-inventor Gidi Littwin and computational neuroscientist Hagai Lalazar, has raised $52 million across pre-Seed, Seed and Series A rounds to launch Descartes, a six-billion-parameter foundation model trained to decode non-invasive EEG.

The pitch is deliberately mundane on the hardware side and ambitious on the model side. Patients wear a dry-electrode EEG headset for roughly fifteen minutes, tap through tasks on a tablet or smartphone, and the model turns the electrical signal into quantitative readouts a clinician can use. The training corpus is Hemispheric's own: reportedly more than 250,000 hours of EEG and behavioral data from over 100,000 participants, with early clinical targets in PTSD, brain injury, depression, schizophrenia and Alzheimer's.

Why this matters if you don't care about brain implants: psychiatric and much of neurological diagnosis still leans on interviews, questionnaires and clinician judgment. If a wearable-plus-model system can produce an objective, repeatable score in a quarter of an hour, the downstream effects touch clinical trials, drug development and how insurers pay for mental-health care. The investor list, which per the coverage includes Hanaco Ventures, Arkin Capital, OurCrowd, Protocol Labs, L Catterton and individuals such as Scott Belsky and Howard Morgan, suggests people are treating it as a platform bet rather than a single-device play.

The honest caveats are the ones the announcement does not settle. There is no published accuracy figure against existing diagnostic standards, no disclosed regulatory pathway, and no detail on how the 100,000-participant dataset was consented or how well the model generalizes outside it. Consumer-grade dry EEG is famously noisy, and scaling laws that hold on internet text do not automatically transfer to biological signals.

Take the specifics as reported, not settled. If Descartes turns out to be even directionally useful, the winners are pharma companies that need objective biomarkers for CNS trials, and the health systems that spend more on misdiagnosis than they would like to admit.