前 Neuralink 创始人：AI 增强型身体已近在眼前｜与 Max Hodak 对谈 | EP #171

I'm super pumped about what you've been building. We're going to talk about 2 things: a product you have today, which is a revenue engine, and the incredible moonshot you're pursuing. I love entrepreneurs who've got something big and bold but also have a real business at the same time on the path there. That's extraordinarily unique and critically important when you're actually building a business, and you've done both. Before we jump into the BCI of it all, would you talk about PRIMA?

First of all, thank you for having me. It's a pleasure—super cool to be here. I don't want to undersell the near-term product, which is still a huge deal.

We have the world's first retinal prosthesis that really works. There are 3 layers of cells in the eye that transmit vision, from light coming in to a signal going into the brain. For patients who have lost the rods and cones in the back of their eyes, their retina is intact and their brain can see, but their eyes no longer light up.

What types of disease are these typically?

This is specifically macular degeneration, especially age-related macular degeneration. Does anybody know anybody with macular degeneration here in the room? A lot, right? It's pretty prevalent. It's also retinitis pigmentosa. My mother's father had retinitis pigmentosa, so I grew up around blindness.

We have a chip that can be implanted in the back of the eye. Each one of these little honeycomb structures is essentially a solar cell. The patient puts on glasses with a laser projector that strikes the implant in the back of the eye to excite the remaining viable cells in the retina and get the visual signal into the optic nerve at the first possible opportunity beyond the dead photoreceptors.

It's a super-simple, 1-hour outpatient procedure. The surgeon makes a little bleb under the retina, places the chip, and then the patient goes home and recovers in a couple of days. They can put on these glasses. We finished a clinical trial last summer with 38 patients in Europe. About 50 people around the world have had it so far. It's the first time in the history of the world, as far as we know, that these blind patients have been able to read again. We're super excited about this.

You know, it's interesting because giving sight to the blind is a very biblical statement.

Literally.

Literally. Yeah. I'm definitely very excited about the BCI technology that we have coming. One of the things I've learned is that the end state is often obvious. Ray Kurzweil was saying in the 1990s, “We would get here.” The end state can be inferred. The question is, how do you get there, and what kinds of investment will it take to make these technologies work? I definitely don't want to undersell the retinal prosthesis.

Let's talk about PRIMA for a second. If you were going to describe the state of the technology in humans—working and ready to sell—where does it stand?

I have to be very careful with what I say on that piece. It's not approved yet, but we're planning to submit for marketing approval in Europe in the next month. We're discussing with the FDA exactly what else they need to see, but we're hoping to have this on the market in the EU, definitely, and hopefully in the U.S. early next year.

Amazing. How big is that potential total market?

There are many billions of dollars a year in potential. There are very strong reimbursement precedents, at a couple hundred thousand dollars—probably around $200,000 per patient. In the U.S., the payer that matters here is Medicare because all these patients are over 65. In Europe, there are reimbursement precedents around $150,000 per patient.

There are tens of thousands of patients for whom this is directly relevant. The whole population is probably about 250,000. If you can reach even 2,000 or 3,000 patients a year, this is a half-a-billion-dollar business and a $3-billion-to-$4-billion opportunity.

I love it when an entrepreneur describes their on-base single as a billion-dollar opportunity. But when you look at AI, the companies that get to invest most sustainably are the profitable tech companies. Having to go back to the well every 18 months is a huge strategic disadvantage. There's been a ton of capital and enthusiasm flowing into BCI, but what this space really needs is a company making $100 million or more a year. Until that happens, there will always be a risk of winter. I think we're incredibly close to super-exciting breakthroughs, but they have to be supported by something that can fund them sustainably.

And you've acquired the manufacturing and built up the manufacturing capacity for this?

Yes. We've done 2 acquisitions, including one as a captive MEMS fab in North Carolina. Vertical integration is essential. I absolutely received the gospel of vertical integration from my former co-founder and prior boss.

We routinely void the warranty on million-dollar fab tools to place atoms exactly where we want them. Being able to do that, and also go from a design change to surgery in a couple of weeks, is absolutely enabling us to innovate.

Hold that in your mind: a company solving something of extraordinary difficulty, with the technology operational, regulatory approvals coming very shortly, and revenues following shortly thereafter. I think that's an extraordinary accomplishment on its own. Now let's move to the grand-slam home-run potential.

I remember when I was talking to Ray Kurzweil about his predictions. If you Google his predictions, he's got an 86% accuracy rate, if you look on Wikipedia. One of his predictions was high-bandwidth BCI by the early 2030s—like 2033. I said, “Ray, this one I don't see happening in that time frame. You're wrong about this one.” Then I met Max, and I thought, “Okay, Ray, you're right again.”

Max, just for a moment—I won't linger on it—but you were the co-founder and president of Neuralink. How long were you there?

About 4 and a half years.

I wouldn't say this—and he won't say this—but, you know, it's actually... I won't even say it. All right. You broke away and founded Science. You had a unique idea, which I think is extraordinary.

Yeah.

Describe the problem with all the current neural implants. You've got external BCI, which is looking at EEGs. You've got something under the skull, above the dura. Then you have wires placed in the upper parts of the neocortex, and you have deep-brain stimulation. Those are all different types, but let's talk about products like Neuralink and others. What's the challenge they have?

There are many different ways to try to record and drive the activity of neurons throughout the brain. Neuroscience as a field has been trying to do this for the last 100—almost 150—years. The first thing I want to say is that BCI is a field, not a product. There are many different products that will use many different modalities for different things.

There do seem to be very serious, fundamental physics limitations to the resolution and accuracy that you can get with purely noninvasive devices. Once you start thinking about putting something below the skull, the main approaches used today involve putting wires into the brain. The idea there is very simple: neurons communicate through electrical fields that they generate, and if you put an electrode in the brain, you can detect this.

There are other groups that are genetically modifying neurons in the brain to make them light-sensitive or make them emit light. There are also groups interested in using ultrasound. The problems with ultrasound and the optical methods of optogenetics are that they really require genetically modifying neurons throughout the brain.

And so doing this in an adult human is really pretty tricky. That seems like a nonstarter for many cases. You’re irreversibly modifying these neurons in the brain of adult humans using these viral vectors, and they don’t get perfectly distributed. Even then, there are still really severe limits to the depth that you can image or the resolution that you can get.

Now, the problem with placing wires into the brain, which allow you to get single neurons, is that we’re used to these cartoons of neurons floating in space, where you can place electrodes safely between them. But the reality is that there’s no space in the brain. The brain is this wet, warm, squished thing.

No matter how small or how flexible your device is, it might look like it’s a tiny fraction of a human hair floating off a finger. Every time you place one of these into the brain, you destroy thousands of cells. That blue line is your typical thickness of an electrode.

If you have a serious spinal cord injury, destroying 50,000 cells in the cortex to get 5 bits per second by recording from 500 neurons might be totally vindicated. But it does mean that you can’t scale up this approach to millions of cells, and that is really what you want in order to get these next-generation applications.

So I think about what an idealized neural interface is. I’ve been thinking about this question really since I was in 5th grade. Can I set a piece of context for folks?

I love that.

In terms of bits per second, in terms of baud rate, how would you describe the human brain interface for communication and speech? Let’s establish how fast our brain is actually inputting and outputting information.

There are 2 ways to answer this question. The figure of merit for any brain-computer interface is bandwidth, in bits per second. This is another way to look at the different approaches.

There are groups that are placing stents with electrodes into blood vessels. There’s something very elegant about getting into the brain through the body’s natural road system, but because of where that limits you and how far you are from the cells, those only get half a bit per second.

Penetrating cortical electrodes, like Neuralink, BrainGate, and others, have been able to get 5 to 7 bits per second. Spoken language is about 40 bits per second.

Okay. Just to hear that, right? Forty bits per second is when you and I are speaking. Neuralink is probably getting how much, you think?

I think what’s been published is about 7 bits per second.

7 bits per second.

Yeah. This is an interesting result, because if you take many different human languages, some are spoken more quickly and convey less information per token, while some are spoken more slowly and convey more information per token. But if you plot these, they all come out to about 40 bits per second.

There’s also a lot of neuroscience evidence that our attention processes the world at about 10 bits per second. The amount of information that you can perceive and remember is limited by this evolved cognitive bottleneck of about 10 bits per second.

When I think about high-bandwidth BCIs, I don’t think in terms of communicating faster. I don’t think you’re going to make it so you can simply convey thoughts more quickly. But it might be possible to get information into the brain. That’s very straightforward and very easy: You can see, you can hear, and you can feel.

These are much, much more than 40 bits per second, but you can’t get these signals out of the brain. For everything that you can perceive, you can imagine, but you can’t get imagery or audio out of the brain. That might be possible.

Or we can think about adding new cortical areas, in the sense that even if you’re still communicating at 40 bits per second in terms of the number of tokens, can you make those tokens much smarter? Can you have skills, knowledge, or memory that allow you to get a Chinese character versus a letter of something?

I just want you to get those numbers, because maybe you think you communicate in megabits or gigabits, like your computer does. We’re at 40 bits.

When I think about what the idealized brain-computer interface is—the one that would really solve a lot of these problems—the thing I think of, if you’ve seen the Avatar movies, is this big externalized cranial nerve.

All of the information that flows in or out of the brain goes through a relatively small number of wires. There are 12 cranial nerves. The optic nerve is nerve II. The vestibular nerve that carries hearing imbalance is nerve eight.

Then you’ve got 31 spinal nerves that connect out to the muscles.

Bringing back memories from medical school.

When we think about our retinal prosthesis, what we really see is a nerve II interface. But the question is: Could you grow a 13th cranial nerve that has interhemispheric bandwidth—the bandwidth that connects the 2 hemispheres through this fiber bundle called the corpus callosum? Could you have a branch of that come out and give you a USB cable at the end?

This was an idea that I had back in college, but it was really beyond the field’s collective ability to build at the time. The idea we had was: What if, instead of placing something into the brain, we load an electronic device with heavily engineered, stem-cell-derived neurons, embed them in a hydrogel so that the cells don’t go anywhere, and then engraft the wet side of this into the brain?

There’s no serious injury to the brain. You don’t place any electrical or mechanical parts. The only thing that penetrates into the brain is the biological processes of these graft cells. But at the far end, you get chemical synapses.

We can activate these cells optically to fire them selectively. They grow both axons and dendrites, so we can get input and output. We can record from them electrically. Optical stimulation and electrical recording allow us to avoid crosstalk.

We can drive all of them at once. This is a cool device, because you can easily make a 100,000-electrode device when you’re much closer to the cells. You can have much tighter electrode pitch, and you can load that with 1 million neurons. When that grows in, you’ll get 1 billion synapses throughout huge areas of the cortex.

Really critically important here: If you look at BrainGate and Neuralink, how many total electrodes are they placing?

BrainGate places 100. Neuralink, so far, has placed 1,000.

We’re talking about 100,000 or millions of these.

The other thing that’s beautiful is that these neural grafts—their axons and dendrites—when they grow into the brain, because they’re native to the brain, they’re not disrupting the tissue. They’re pushing it aside.

If you were to do this for real, you would see an image that looks like this. This is a mouse brain. You can see at the top there’s a bolus of cells where the device was removed for sectioning.

So this is what you did?

Yeah. This is functional in a mouse.

The graft cells that we’ve added are labeled in green. The host neurons of the mouse are in blue. If we look carefully, you can see all these little green dots really throughout it.

What we’ve seen is that when we engraft these devices, they grow in and wire up very promiscuously. They form connections everywhere. Then, after about 4 to 6 weeks, they start undergoing activity-dependent pruning.

The really interesting possibility there is that how they wire up is not necessarily genetically defined. It can be informationally defined based on the types of activity that you’re getting in the device.

In addition to growing down and wiring up through our cortex, the first layer on the surface of the brain, cortical layer 1, is a white matter tract. These are long projections, like a highway between different areas of the brain.

We often see in the devices that we’ll get a fiber bundle that follows that for millimeters. The mouse brain is very small, but we see these things project all the way through to subcortical structures.

Where these neurons and dendrites grow, they wire up and connect, and where they don’t, they die off.

Yeah, they’ll retract. I mean, the cells mostly don’t die, but they’ll retract the axon growth cones and the dendritic arbors.

So this is what we have. This is looking at one of the chips. There are cells loaded in these trenches.

Here, this is a Z-stack. Each frame starts at the surface and looks deeper and deeper into the brain. You can see the circles of the cell bodies on the surface. All of the green that we see is the graft neurons, but you can see the shadows of the blood vessels in these lighter layers.

This is super cool, because when you place an electrode into the brain, you always get bleeding. If you hit a descending blood vessel, you could stroke out a whole mini-column. Here, these grow in around the blood vessels. The capillaries remodel around them.

And so this is a really perfectly biocompatible way to get chemical synapses. We see these things even where it looks like they’ve fallen off. You see the processes of these cells growing in.

The theme of this event, this year’s summit, is convergence. What technologies had to converge here to make this possible?

A lot of this was enabled by recent advances in cell engineering. One of the things that we have to do is hide the graft cells from the immune system. We do a lot of editing to these cells. If we were to do this on a per-patient basis, because the immune system would have to recognize them, this would take over a year and be prohibitively expensive.

There’s been a lot of advancement recently in making what we call hypoimmunogenic stem cells. The whole CRISPR toolbox and a lot of other technologies now include things like small-molecule-triggered kill switches. We can make it so that if you take a vitamin, the graft cells will die. We can keep an eye on them. Once you’ve hidden them from the immune system, you kind of want that built in.

When these go to humans, these will be by far the most heavily edited cell therapy to reach people.

And materials science?

Yeah, it’s materials like silicon carbide. There have been big improvements in the materials. We talk in the BCI field about the smartphone dividend. We rely heavily on the same tech stack that smartphones and wearables build on, but Apple, Samsung, and others have poured over $100 billion into that. Our field is too small to afford that today, but we get to build on it, and that has really enabled us and advanced a lot in the last few years.

All right. Talk to me about where and when this enters primates and potentially humans. When can I get mine?

We currently have some primates getting trained up on behavior.

So you’re training them in advance?

We’re training them in advance. We also need to figure out things like how well they can reason, which actually hasn’t been studied that well. We’re hoping to do the first primate engraftments in a couple of months.

I mean, which is amazing, right?

Once you’re operational in primates, other than regulatory prohibitions, you’re effectively functional. We’ll be able to prove the neuroscience—that is, the big questions for humans and primates. The first humans to get this will almost certainly be for stroke.

If you’d asked me this 3 weeks ago, I would have said I thought it would be 4 or 5 years before the first human got it. I actually think this is now going to be much faster. There might be a collaboration that allows us to go to humans a lot faster than I’d realized. Again, that’ll almost certainly be for stroke.

The primates are actually a pretty good model of human stroke patients, because a human who’s lost a cortical area can be modeled by a monkey that never had it in the first place. If you can restore that capability, then there’s an argument that you’ll be able to do it in humans.

The other thing I’ll say is, beyond stroke rehabilitation or re-adding these capabilities to humans, when we think about scaling this up, I see this as a way to redraw the borders around the brain. Your head has 2 hemispheres. These are connected by a fiber bundle called the corpus callosum that gives you the experience of 1 agent in the head. But really, you’ve got 2 subbrains that are mostly independent.

A long time ago, people used to cut the corpus callosum in epilepsy patients to prevent a seizure from spreading across the hemispheres. If you cut that, you really get something that looks like 2 agents in 1 head. There’s a natural example of going the other way: There’s a pair of twins in Canada, the Hogan twins, who have 1 head with 4 hemispheres, and they can share meaningful elements of their consciousness. There’s an element of task transfer between them.

I think a way to conceptualize this is to imagine if this was a tech product. That might be coming a lot sooner. We know this must be possible because nature has done it, and I’m hoping to have this in humans, hopefully, pretty soon.

Max, I want to dive a little further into what this will mean. So this becomes enabled—other than me being able to think in Google or watch a 4K video with my eyes closed—what does this actually mean in terms of increasing intelligence and connecting to AI? What is a possible future here?

For a lot of my life, I always felt like I could see the future, and I’ve got this event horizon somewhere between 2030 and 2035 now. It’s just impossible to see past. AGI and ASI are definitely happening, and I think that this is—I mean, everybody knows about it now, but there’s basically no way to overrate the impact of that.

Is this the merge, the only way through? I don’t know. That was a conversation we had last year: Do we need to couple?

But if we do merge, you’ve talked about the idea of pretraining, in some ways, these biohybrids. Can you speak to that a little bit?

I think neural interconnects, like brain-to-brain connections, are a really interesting technology for merging with AI. I think transformers are a pretty good explanation for cortex, but to get real agency in a way that is interesting or dangerous, you need something else to add on to that.

People have these loops that are prompted, but that is still coming from the human. It might be that the agency remains with the humans, but these technologies are so powerful and adaptive that people who participate in this have a huge advantage. This is something that societies need to think about.

I also see it as a longevity technology. How do you let someone into your head? That’s a tricky question. I think the first use case for this would probably be things like long-married couples where one has a terminal disease for the last year. You can get a brain-to-brain connection. So rather than merging with AI, it’s merging with your spouse or a close family member.

Talk about a level of intimacy.

Yeah. Can you turn the death experience into basically a stroke that you recover from? Throughout your life, small groups of neurons are constantly dying. There’s a smaller number being generated, but this is turning over.

All communication is about creating correlations between brains. Long relationships already store memories in each other’s brains. Is there a threshold where you can get phenomenal binding across the interface, where you really get 1 agent, and then when you lose some group of neurons, you still get continuity of consciousness and continuity of experience through that transition?

I see that as an alternative path to biological longevity companies, but it feels a lot more like an engineering problem to me. I think it will be possible on the timescale of the next decade. In that view, you can merge with other people, merge with AI, or have these superorganisms that are composites of big groups.

Yeah, I call them a metaintelligence. When we’re able to connect millions of people’s thoughts and feelings at a level of intimacy and connection, I mean, you are a collection of 40 trillion cells that you don’t think of yourself as 40 trillion cells. You think of yourself as you. Imagine if millions or billions are connected through the cloud together, and you become conscious on yet another level.

The really interesting question here, where we’re still missing some physics, is what is the point where you go from having 2 conscious experiences into a single experience, or do you keep multiple attentional windows? I’m increasingly confident we’re going to get this in the next 5 years.

This is tough to talk about without sounding like a lunatic. All I know is that these devices are technically capable of being built, and I have no idea what it will feel like, but we’re going to find out.