Casey Harrell hasn't spoken in years. He's typed two million words with his mind.
Jump to a section
What happened
Casey Harrell has advanced ALS. He can no longer move almost any part of his body, and the disease has degraded his natural speech to the point where people outside his family struggle to understand him. For nearly two years, that hasn't stopped him from talking. A brain implant reads his neural activity, a decoder turns it into text in real time, and a synthesized version of his own pre-ALS voice reads it aloud.
The numbers, published this month in Nature Medicine by a team from UC Davis, Brown, and Mass General Brigham: 183,000-plus sentences, close to 2 million words, at an average of 56 words per minute, with roughly 92% accuracy. Harrell has logged more than 400 days and 3,800 hours on the device, which researchers say is the largest single-subject brain recording dataset the BrainGate2 pilot trial has produced. He uses it to send email, browse the internet, and hold down a full-time job. "It is very sweet to have the ability to look at my wife's eyes when she hears my voice," he's said through the system, per ScienceAlert's writeup.
Setup takes help from a caregiver each morning. After that, the decoder mounts on a mobile cart and follows him through the day, and he navigates the interface with eye gaze, "clicking" by thought. He also has a privacy mode he can switch on, and only sessions recorded with it off get used to retrain the model. That detail is small in the coverage but it's not small in what it implies, and I'll come back to it.
Pros & cons
What's genuinely good:
- Continuous, real-world use for 400-plus days, not a lab demo measured in minutes
- 56 wpm is close to a third of typical conversational speech, which is a real communication rate, not a party trick
- The synthesized voice matches his pre-ALS voice, which is the detail that makes his daughter's reaction to it make sense
- Harrell controls what gets recorded and what gets used to retrain the model, via privacy mode
- He's held full-time employment through it, which is the actual bar for "does this work," not a benchmark
What's still unresolved:
- This is a case study of one person inside a pilot trial. The path from "works for Casey Harrell" to "works for the next 10,000 ALS patients" is unmapped
- It requires brain surgery to install electrode arrays, which is a categorically different ask than downloading an app
- Setup needs a caregiver every morning; this isn't yet an independence tool in the fullest sense
- No public numbers on cost, and nothing in the coverage about an insurance or regulatory pathway
- 92% accuracy is good. It also means roughly 1 in 12 words needs correction, which matters a lot more in "I love you" than it does in an email draft
- If you build personalization or fine-tuning systems, read the Nature Medicine paper's methodology on how they gated training data through an explicit user-controlled privacy toggle — it's a cleaner consent pattern than most consumer AI products ship.
- This is evidence for the "small, deeply personalized model beats large general model" argument on certain narrow-but-high-stakes tasks. Worth citing next time someone assumes bigger is automatically better.
- If you work anywhere near assistive tech, accessibility, or neurotech, BrainGate2 is an active pilot trial — worth tracking for when (if) it moves toward broader access.
- Note what's absent from the reporting: no cost figures, no regulatory timeline. Don't build a product roadmap assuming near-term availability.
Further reading
- ScienceAlert — Brain Implant Lets Man With Severe Paralysis Speak in His Own Voice Again — primary source for all figures cited, reporting on the Card et al. study published in Nature Medicine
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
Scientists use AI for the grunt work. The parts they think matter most, they still do themselves.
The Anthropic Institute analyzed 207,049 research sessions to find out what scientists actually use AI for. Data analysis and writing dominate. Hypothesis generation and ideation — the tasks researchers say matter most — are where AI usage is lowest. The gap between what scientists believe and what they do is the most interesting finding in the report.
Claude designed a speakerphone. Engineers built it. It still works.
Anthropic asked Claude Fable 5.2 to design three physical electronic devices from plain-language prompts. Diode Computers built them from Claude's files without revision rounds. After weeks of use, all three prototypes still work. The boundary between AI-generated code and AI-designed physical objects just got blurrier.
The piece of agent architecture most builders are skipping: the dreaming step.
Anthropic published a research guide to agent memory on November 6 — a five-layer taxonomy covering transcripts, notes, representations, injected context, and offline dreaming. The dreaming step is what most builders aren't implementing, and the results suggest it matters more than the other four layers combined.