Source: CTech (Calcalist) — 2026-07-15
Summary
Israeli-US neurotech startup Hemispheric emerged from stealth with $52M raised across three funding rounds, unveiling Descartes — a 6-billion-parameter foundation model trained directly on brain data rather than text, aimed at precision psychiatry applications like PTSD, depression, and Alzheimer's. The company was founded by Gidi Littwin, a co-inventor of Apple's FaceID.
Key Takeaways
- Descartes is trained on 250,000+ hours of brain data collected from over 100,000 participants — a genuinely new AI modality (brain-signal foundation models) rather than another large language model applied to a medical domain.
- The founder's background — co-inventing Apple's FaceID, a biometric sensing-and-inference system — is directly relevant pedigree for building a foundation model around a different kind of biological signal.
- Target applications are specifically precision psychiatry: PTSD, depression, and Alzheimer's, conditions notoriously hard to diagnose or track objectively from behavior or self-report alone.
- $52M raised across three rounds funded stealth-mode model development before any public product launch, suggesting a research-heavy runway rather than a fast go-to-market push.
Reel Script
Hook (16s / 36 words) Every AI foundation model you've heard of was trained on text. This one was trained on a quarter-million hours of raw brain data — and it's aimed at diagnosing depression and PTSD.
Core Concept (60s / 130 words) A foundation model is just a large model pretrained on a huge amount of one kind of data, general enough to be adapted to many downstream tasks — that's true whether the data is text, images, or, in this case, brain signals. Hemispheric's Descartes model was trained directly on brain data from over 100,000 people, totaling more than 250,000 hours, rather than on text scraped from the internet. The bet is that psychiatric conditions like depression, PTSD, and Alzheimer's leave patterns in brain activity that are more objective and detectable than what a patient can self-report or a clinician can observe from behavior alone — so instead of adapting a language model to psychiatry, they built the foundation model around the actual biological signal from the start.
Hands-On (50s / 110 words) The concrete specs: a 6-billion-parameter model, trained on 250,000+ hours of brain data from 100,000+ participants, developed through $52M raised across three funding rounds before the company's public unveiling. The founder detail matters for credibility here — Gidi Littwin co-invented Apple's FaceID, which means his prior work was specifically about building reliable inference systems on top of biological sensor data at consumer scale, not a generic ML background applied opportunistically to healthcare. That's a meaningfully different starting point than most "AI for mental health" startups, which typically start from an LLM and add a psychiatric use case on top.
Takeaway (22s / 48 words) This is early — stealth-to-launch, no clinical deployment numbers yet — but a purpose-built brain-signal foundation model for psychiatry is a genuinely different bet than another LLM wrapper. Worth watching for real clinical validation data as it moves out of stealth.