Neuralink has given a person with ALS a feature most of us use with mild irritation, a voice assistant with a wake phrase. The difference is that nobody speaks. The user mentally says "Hey, Neuralink," describes a task, and the AI agent Grok takes over the operating system to carry it out. The person in question has lost the ability to move and to speak.
From thought to task
The user in the report has amyotrophic lateral sclerosis (ALS). The sequence works like this:
- The user mentally speaks the wake phrase, "Hey, Neuralink."
- The user formulates the task.
- Grok takes control of the operating system and performs the necessary actions.
That is a different job from typing with your mind. Typing means producing every character yourself. Here the user states an intention and the agent does the clicking. Anyone who has spent an afternoon dragging a mouse through a settings menu will see the appeal.
Neuralink tested the prototype in a 72-hour session. Its engineers ran it at the home of a participant in the company's speech-restoration study. A lab is a forgiving place, and a living room is not. The company says the home setting let the engineers try the new features in everyday conditions and decide where to take the work next.
No record button
The second addition is automatic detection of speech intent. The system works out when the user is trying to speak and passes the corresponding signals along. Nobody has to press a button to start or stop recording.
This sounds minor until you picture the alternative, which is a person who cannot move needing a physical or mental click every time they want to say something. Removing that step makes communication hands-free in the literal sense.
Combined with the agent, it also changes what the implant is for. The user can type text, and the user can also hand the computer a job by mental command.
Phone calls through Bloom
Neuralink is also working on iPhone integration. Its Bloom app sends speech synthesized from brain signals straight into the audio channel of a phone call. The user can call family, and can use other smartphone functions, ordering food among them.
There is a small detail that deserves credit. At the start of a call, the system can automatically play a greeting that tells the other person the call is being made with an assistive communication device. Anyone who has picked up a call from an unknown number and heard a synthetic voice will understand why this is a sensible default. The greeting spares the person on the other end a moment of suspicion and the user an awkward explanation.
A window into the signals
During testing, the engineers added a visualization of the signals coming from the implant. It shows their intensity and distribution. That helps identify which imagined actions produce the strongest brain activity.
The point is practical. The data goes toward improving the models that turn neural activity into commands and speech. In other words, the display is a diagnostic tool for the engineers as much as a feature for the user.
The one number
Neuralink also reported results for the speech interface. In the first week of use, one trial participant typed 75 English words per minute with an error rate of 5.8%.
The report is careful to say what this figure is. It is an early result from a single participant. It has not been confirmed across all users of the system. A good headline number and a good average are different things, and the source does not pretend otherwise.
What the report leaves out
The account is short, and several questions go unanswered.
- Reliability of the agent. There are no figures on how often Grok completes a task correctly or how often it gets one wrong.
- Safeguards. An agent that takes control of an operating system is powerful, and the report says nothing about how the user confirms or cancels what it does.
- Who is who. It does not say whether the participant who reached 75 words per minute is the same person as the one in the 72-hour home session.
- Scale. It does not say how many people have tried the agent, the phone integration or the signal display.
None of this makes the work less interesting. It does mean the story is a demonstration, with the evidence still to come.
The wake phrase survives
The design choice that stays with me is the oldest one. Even when the input is a thought, the system wants to be addressed by name before it does anything. The smart speaker on the kitchen counter got there first, and apparently the brain implant had no better idea.