Neuralink Is Building Both Ends Of The Call

Picture a doorway across a room you can't see. Now picture a device that gives your brain just enough to know where that opening is. No color, no fine detail. Enough to walk toward it.

Then picture someone with ALS who knows exactly what they want to say and can't physically say it. A friend asks how they're doing. The answer is right there. The body is in the way.

Neuralink is working on both of those people. One project carries a picture in. The other carries words out. My argument in the video is that fixing those two connections leads somewhere bigger than medicine: a better way for you to use AI. Watch the upload for the full walkthrough. This is the same case in writing.

Both ends of the call

Think about a video call. The picture on your screen is information coming in. Your microphone carries information going out. You need both directions to have the exchange. Lose either one and part of the conversation disappears.

That's the frame I use for Neuralink. Blindsight is designed to carry visual information into the brain. The speech system takes the words a person intends and gets them out to another person. Both are routes around damaged connections in the body.

Elon Musk's update on Blindsight is very direct. He says it is ready from an engineering standpoint and just waiting on regulatory approval. That's his own assessment, and approval is the piece still pending. The engineering behind it is worth walking through either way.

Going around the eye

Start with the camera. It captures the scene in front of you. In Blindsight's design, that image gets sent wirelessly to a specialized implant, and the implant activates the parts of the brain responsible for vision. That includes the visual cortex, the region that processes what you see.

So the route runs from camera to implant to the brain's visual system. It goes around a damaged eye or optic nerve, the nerve that normally carries information from the eye.

Why go that deep? Because replacing the image coming into an eye still leaves you needing a working path from that eye to the brain. Reaching the brain directly changes who the device can be built to help.

Elon's stated goal includes people who have lost both eyes and their optic nerve. It also includes people blind from birth who have an intact visual cortex. That condition matters: the brain's vision system has to be there for this design to work.

Think about someone whose eyes have never shown them the world, receiving visual information through an implant. That still has to be proven in people. It's an extraordinary thing to build toward.

Two out of three

The camera and the implant have very different jobs. A camera captures an image. The implant has to deliver stimulation the brain can actually use. That second job is where the engineering gets interesting.

Neuralink tested stimulation of vision-related brain areas in a monkey. The animal looked toward the location the researchers were trying to make it perceive at least two-thirds of the time. At least two out of three attempts.

The useful part of that test is the feedback loop. Stimulate a particular location, then check whether the animal looks toward the intended spot. That gives the engineers something concrete to measure as they work on the system.

Atari first

What does the first useful version look like? Elon's first vision target is low resolution, like Atari graphics, with more detail over time.

Back to the doorway. Suppose the device gives you enough visual information to find that opening. You can tell where it is. For that task, the useful part is knowing where to go.

Every bit of detail has a job, and that's a clean way to think about upgrades. More useful detail means more information about the room. In the doorway example, the next goal is telling what's beside the opening or what's in the way.

Human testing is the next step for Blindsight. That's where the team will test safety and sight in a person. Neuralink runs a vision registry for people who want to hear when those trials launch. Until then, what we have is the design and the animal work.

Words with no way out

The outgoing side of the call already has people using it, through Neuralink's voice trial.

ALS damages the nerve cells that control muscles, including the muscles involved in speech. A person can still have plenty to say while losing the ability to physically say it. Somebody asks how you're doing. You know the answer. You know the words. Getting them out becomes a whole task standing between you and the person asking.

The speech implant records activity from attempted or imagined speech. Software called a decoder turns that activity into words. The words appear as text, or a generated voice speaks them aloud. The sound comes from the computer. The message comes from the person.

In Neuralink's film, Terry, a participant with ALS, trains by miming speech and then by thinking the words. Grok Voice plays those words in a recreated version of his own voice.

Why recreate the voice? It can be built from recordings of how the person used to sound. When you're talking to somebody who knows you, that familiar voice is part of the connection. Full disclosure on that front: the voice narrating my video is a clone of my real voice. The idea and framing came from the real me, and my Grok bot, Solo, turned that idea into the finished video.

1,024 contacts

In Terry's case, the words start with his attempt to speak. The software has to learn how to turn that brain activity into the words he intends. No typed sentence comes off the implant. What comes off is activity recorded near brain cells, and the computer has to work out what that activity means for what he's trying to do.

So how does Neuralink get enough of that activity to work with? Take its earlier motor implant, N1, built for controlling a computer. It records through 1,024 electrodes spread across 64 flexible threads, which a surgical robot places in the brain. An electrode is a tiny electrical contact. That works out to 16 recording points per thread on average, a lot of listening posts near brain cells.

The hardware and software split the work. The electrodes collect the activity. The decoder turns the recorded patterns into a useful action, like moving a cursor. For speech, the useful output is words. Terry's training makes that split easy to picture: he attempts speech while the system records, and the software learns the relationship between those signals and his intended words. Then the patterns can become text or voice.

The whole chain has to hold: a useful signal from the brain, a correct read of what the person intends, and an output they can actually use.

75 words a minute

Then the question everyone asks: how fast can the words come out?

Neuralink says one participant with ALS reached 75 words per minute in their first week. Natural conversation runs around 160 words per minute. That's a little under half conversational pace, through a brain implant, in week one.

That result had a 5.8% word error rate, about one error for every 17 words. It used a vocabulary of 125,000 words, which gives the system far more choices than a short menu of preset answers.

Across participants, Neuralink puts the median speech rate at roughly 50 words per minute. That's about six times the rate of typing with an eye gaze system, where your eyes select the letters.

Picture sitting at a table while people talk. Someone asks you something, and you want to answer before the conversation moves on. The speed of that exchange decides how easily you can take part in it.

Ten minutes a week

Speed also covers what happens before the first word. How much setup do you need? How much effort does it take to start?

Neuralink says it has more than 50,000 hours of implant use across participants. Laid end to end, that's over five years. Those recordings are being used to pre-train decoding software. The decoder has to recognize patterns in brain activity, and more recorded use gives the team more material to train on.

The result has to show up when a person sits down at the computer. In Neuralink's cursor work, some decoders now keep working for weeks instead of days. For some users, calibration (the setup that helps the software interpret their signals) went from 10 minutes a day to 10 minutes a week. Over seven days, that's an hour back, plus fewer times they have to stop and get the decoder ready again.

Same reason a video call needs more than a good microphone. If you had to reset the call every time you wanted to say something, you'd stop calling. The connection has to stay useful while you're using it.

The button problem

Now take that into speech. Neuralink engineers spent three days with Kenneth, a speech trial participant, building features in his living room.

One prototype detects when he intends to start and stop speaking, which removes the need to press a button. Think about what that button asks of somebody who already has trouble moving. Before the sentence can come out, there's one more thing to operate. Detecting speech intent takes that step out of the way.

They also connected his generated voice to an iPhone call. The speech goes straight into the call's microphone stream, so the person on the other end hears it. That puts the voice inside something people already use to talk.

That's a concrete job beyond the implant itself. The words have to reach the place where the conversation is happening. A voice on a screen gets a lot more useful once you can call somebody with it.

Hey, Neuralink

A sentence coming out of the brain can also go to a computer that acts on it. That's the AI connection.

An AI agent is software that can take a task and carry out steps on a computer. Once intended speech becomes an instruction software understands, you have a way to ask it to do something.

In Neuralink's hackathon video, Kenneth uses a prototype that connects imagined speech to a Grok agent. He thinks of saying "Hey, Neuralink" along with the task he wants done, and that triggers the agent. That's a person giving software a task through intended speech. For somebody who has lost speech and hand movement, it's another way to direct a computer, and the agent can become part of how they get things done.

The human bottleneck

Take that idea forward. Imagine an agent helping you edit a document. You give it the change you want, look at the result, and correct the part it misunderstood. The task gets done through that exchange. You still need to get your intent across, see what happened, and say when the result is wrong.

Now make the agent faster. Suppose it finishes each edit before you've finished giving it the next instruction. At that point, getting your next request into the computer sets the pace of the work. The machine is waiting on you.

That's the human bottleneck. The more capable the machine gets, the more it pays to improve the connection between you and it.

Throughput is the word for this: how much useful information gets through in a given amount of time. More of your intent reaches the computer, with less effort spent getting it there. A faster connection still has to carry the right instruction. If you ask for a change and the system hears something else, you're stuck correcting it. Accuracy is part of how quickly you finish, which is why the 5.8% error rate belongs right next to the 75 words per minute.

Why not just talk?

A fair question: why would an implant add anything when you can already talk to an AI? For a person who can speak, talking is a very useful interface.

So run a hypothetical. Give two people access to the same agent. The AI on the other end is equally capable for both. Now imagine one of them has an implanted connection that has improved beyond ordinary speech speed. In that future setup, the advantage comes from how quickly that person can communicate what they want and correct the result. Same agent, more useful instructions in the same amount of time.

How fast does Elon want the connection to get? Telepathy is Neuralink's first product, and his speed goal for it goes well beyond ordinary speech. His forecast is that output bandwidth will vastly exceed the fastest human speech or typing in a few years. That's Musk's forecast. Today's measured numbers are the ones above.

Back to the document. Say it needs several rounds of changes. You read a paragraph, spot something wrong, ask for the correction, check it, and move on. A faster, useful connection helps at every one of those exchanges. Less time feeding requests into the machine leaves more time deciding what you want. That's how a better connection turns into a task finished sooner.

Bigger than one implant

This is why I think Neuralink's long-term opportunity is bigger than a single implant function. The company is learning how to connect a person's intent to useful software, and how to build a new route for visual information coming back in.

A person who can send words through an implant can use those words in more than one place: a conversation, or an instruction to software. Improving that one route can improve several things that person does with it.

Now bring Blindsight back in. Its goal is to turn information captured outside the body into visual information the brain can use. Words out, picture in. Better communication in both directions is the point.

For everyday use across the population, my read is a decade-plus horizon. The development path starts with people who need another way to communicate. Their actual use gives the engineering a purpose, right down to whether starting a sentence takes an extra button press.

The call goes through

So why do I think this is such a compelling future interface for AI? A direct path from intended speech to an agent already has a prototype. The long-term goal is to make that exchange faster and easier, with useful information coming back the other way.

Kenneth put it in very simple terms. After the engineers' visit, he and his wife, Cheryl, joked on the phone again.

Think about that next to the goal of seeing a room through Blindsight. A connection that was lost becomes something a person can use again. The picture comes in. The words go out. The person gets to take part in the conversation.