Elon Musk has a talent for describing the most comforting kind of crazy-ass futures. His vision—Mars rockets! Underground electric robot cars!—sells itself as the world of tomorrow, today. But it’s actually the science fiction of yesterday, ripped from the pages of Science Wonder Stories. (Or maybe WIRED circa 1995.)
It’s lovely that someone is finally building yesterday’s pew-pew science fiction priorities. That’s especially true of the latest Muskery, presented last week: a new iteration of Neuralink, a wireless implant that could someday give human brains a direct interface to digital devices. Once the company gets past the mere formalities of human testing and government approval, and convinces people to pay for the privilege of having a robot surgeon stitch 1,064 wires into their brains, we’re halfway to Borgtown, babies. Control prosthetic limbs, play Starcraft with your mind, summon your Tesla telepathically, and eventually upload your consciousness to an immortal robot body. Beep.
About that: Neuroscientists who study the human brain and work with electrodes and neurons, the specialized cells that comprise our think-meat, are pretty psyched—hah!—about Musk’s tech. It looks like a real leap forward for research. But as for the mind-reading, memory-downloading telepathy stuff? It’s possible that Elon has made one of the classic mistakes of a would-be technomancer. The first, of course, is never get involved in a land war on Endor. But the second is, don’t confuse metaphors for science.
Neuralink’s technology, to be clear, looks awesome. It’s a brain implant the size of four dollar coins with more than 1,000 electrodes that will (someday) allow a person to wirelessly send neuroelectrical activity to anything digital, from prosthetic arms to Tesla autopilots to memory-recording cloud servers.
The reason that excites neuroscientists is that right now their tools are relatively crude. The standard is the “Utah array,” a single chip with 64 electrodes on it. Just putting it in or taking it out can damage the tissue around it, and it’s not good at isolating single neurons or covering a large area. “It’s very cumbersome. Typically, they can’t take the thing home. It requires two trained engineers and a PhD and all of that,” says Christof Koch, chief scientist at the Allen Institute for Brain Science. “Musk now has a device that’s at least 10 times better. It’s at least 1,000 channels and it’s all streaming, so that’s pretty cool, right?”
The answer is yes. At the Neuralink presentation, Musk said that his prototype included sensors for motion, temperature, and pressure and 1,024 thin, flexible wires to pick up the electrical signals neurons put out while they’re neuron-ing. In a living and seemingly normal pig that Neuralink handlers brought to the demo, the device was nestled invisible below the scalp and transmitted wireless, real-time signals, powered by an inductively charged battery that should last a full day. (Which, wait, you have to stick a charger onto your head at night? But OK.)
No one from Neuralink responded to my requests for later comment, but at the presentation they acknowledged the challenges still in front of them. A Neuralink is supposed to stay in a person’s head not for the hours, days, or weeks that researchers have achieved with other animals, but for years. That’s tough, because the mammalian brain is an unfriendly environment to anything that is not brain. It’s a lump of computational aspic in a saltwater bath that corrodes most metals. The brain fights off invaders, surrounding things like electrodes with a protective wadding of cells called glia. They’re insulators, which means over time gliosis kills an electrode’s ability to record. So the Neuralink team is looking for materials that’ll resist breaking down and won’t set off that protective response.
Thin, flexible electrodes (and a sewing-machine robot to insert them) should also help; in fact, those were the innovations that encouraged Musk to start Neuralink. The 5-micron-wide wires the company uses are supposed to cause less damage to blood vessels during installation—hurting the vessels that carry oxygen-bearing blood to the brain is a bad, bad thing. The more permanent electrodes used today tend to stay stuck to the skull, which means they can do yet more damage as the meat sloshes around during everyday use; flexible wires are meant to move with the brain. And Neuralink’s Bluetooth wireless connection means it has no wires sticking out of someone’s head, which would be cumbersome and a potential site of infections.
In the language of the profession, these are solutions to issues of biocompatibility and longevity. Maybe. At the demo, Musk introduced a pig with an implant, and another from whom the team had removed one safely. But he didn’t share specific data on how long the Neuralinks had been in the pigs, nor how many pigs they’d worked on overall. So it’s hard to know how close Neuralink is to achieving its major selling points: an easy-to-implant, non-damaging, long-lived cybernetic implant. “Having that work in a human brain for a long time without problems, without destroying a bunch of blood vessels and so on, is a really hard biological problem,” says Loren Frank, a neuroscientist at UCSF and Howard Hughes Medical Institute. “That’s separate from hardware, however beautiful it is.”
And that’s not even the really hard part. Musk left out last week that neuroscientists still don’t really understand all the different types of neurons and how they all work together. They can measure the signals those electrodes pick up, but extracting meaning from them is a whole other problem. It’s possible to put an electrode into a single neuron and monitor what it does for hours or days when presented with different stimuli—a pattern of squares, a color, a task. It’s even possible to build a model of what that neuron will do going forward, or some of the other neurons it talks to. But how all of that turns into memories, thoughts, and feelings? Yeah; no.
Musk laid out all kinds of neurological issues that a Neuralink might someday fix—mental illness, vision and hearing impairment, memory loss, and addiction were all on the list. But the first probable use, he said, might well be something humans can already do. That’s using a neural implant to control a robotic prosthetic or operate a digital device. That’s great for people with missing limbs or paralysis.
Right now, a neural implant and connected computers can learn to associate outbound signals with specific intentions. But a person has to train with the device to learn to emit the kinds of signals it can understand even as the device learns to correlate signals to desires or activity. This is a good thing; the Neuralink’s newer and more plentiful electrodes might even improve the process.
But Musk also said that in the further future, a Neuralink would be able to record and replay memories, even save them to an external drive and download them into a robot body. He said people with implants would be capable of telepathy—not just sending and receiving words, but actual concepts and images. (“Words are a very low data rate,” Musk said. “We can have far better communication, because we can convey the actual thoughts.”)
That’s asking a future Neuralink to understand, record, and transmit the neural substrate of thought. And no one knows what that is.
Musk didn’t seem to think this was essential. “A lot of people think, ‘I couldn’t possibly work at Neuralink because I don’t know anything about how brains work,’” he said at last Friday’s demo. “Well, that’s OK. You can learn. But we need software engineering, we need mechanical engineering, electrical engineering … chip design, robotics, and all the things a company needs to work.”
At some point, someone is going to have to know something about how brains work. The Neuralink picks up electrical signals—the “spikes” or “action potentials” that run the length of neurons when they’re activated, and signal the squirting of neurotransmitter chemicals across synapses. But some of what the team said seemed to imply that given enough of those signals, they’d be able to interpolate actual thoughts or memories. Nobody’s really sure that’s true. In fact, it’s possible (though unlikely) that the electricity, the movement of charged ions into and out of neurons, is just an epiphenomenon—the exhaust that a brain coughs out while doing the work of creating and maintaining consciousness.
Even if it’s possible to correctly infer mental state from those electrical signals (and it probably is), they still just happen to be what people can measure. “There are things you can do with the neural signals. They’re the expression of things like memories. The retrieval of a memory will be instantiated, we think, in terms of a pattern of brain activity. That’s true,” Frank says. But that’s not how people store that memory for future retrieval, which doesn’t bode well for recording specific ones, saving them somewhere else, and replaying them. “The storage of the memory involves huge numbers of chemical reactions at synapses between brain cells,” Frank says. “Those things can be modified by brain activity, but they’re not the same as brain activity.”
In other words: The electrical activity of the brain happens while you are thinking or remembering, but it may not be what you are thinking or remembering. Just being able to sense and record that activity isn’t recording actual thought. It correlates, but may not cause.
Musk went even farther, though. “It’s read-write in every channel,” he said. He meant that each one of those 1,024 channels can both pick up signals from, and send them to, adjacent neurons. Now, Musk didn’t specify in what sense he meant that phrase. Neuroscientists talk about the capacity to “read out” signals from a brain, and the ability to “write in,” so send signals back. They can read out signals from motor neurons to control a robot arm, for example, or write in auditory information, sound, via a cochlear implant. They’re working on doing the same for sending images to the retina, or the visual cortex. Researchers can record what neurons are doing, and stimulate them so they activate.
Computer engineers, though, talk about reading and writing as getting digital information from a storage medium, or putting information in one.
Is Musk using the terminology interchangeably? Or does he think that the technology’s ability to do the primitive version will lead to the more sophisticated one? I don’t know.
But if it’s the latter, Neuralink might be headed for a metaphor-based failure. Neuroelectrical writing-in is very different from the digital version. “The techniques they have to write information in are primarily electrical stimulation, and that’s just awful,” Frank says. “Imagine when you wrote to a hard drive that you targeted a particular sector or byte, but what you hit was five other bytes first. That’s what happens with electrical stimulation to the brain.” Axons, the long projecting connections between neurons, have a lower activation threshold than the cells themselves. So sending a signal pulse down one of those Neuralink electrodes activates that mesh of connections, a whole lot of cells, before hitting a target neuron—and that’s assuming you know exactly which neuron to target.
If your goal is to control a robot prosthetic or play Starcraft with your mind, it doesn’t really matter if the Neuralink is recording “real” thoughts, whatever those are, or just some signal that both user and device have agreed means “lift coffee cup” or “shoot spaceship.” But if you’re looking for abstract emotions, memories, or ideas, nobody knows what signals correlate with those—if any. “If you want to make me move my arm, I know where to put the electrode,” says Greg Horwitz, a neuroscientist at the University of Washington. “If you could put electrodes in my brain wherever you wanted, and you wanted to make me vote for Biden or Trump, I’m not sure where you should stimulate to make that happen or in what pattern.”
A brain is not a hard drive. A memory is not a video. Neuroelectrical signals emitted during the act of remembering are not themselves memories. Recording and saving those signals for later playback is not the same as uploading your consciousness. “This idea, it really comes from digital computers that these things are the same. You flip the orientation of a magnetic particle and you’ve changed a memory, and you do that by activating a transistor,” Frank says. “I think this is, if you’ll forgive me, a failure of knowledge of biology.”
Obviously Neuralink employs neuroscientists who understand biology. But Musk is a technologist—possibly America’s best-known exemplar of the power of engineering to make big ideas happen. But the basic science of rockets and electric motors was mostly locked in when he built SpaceX and Tesla. Neuroscience doesn’t even have a consistent theory of consciousness yet. That makes it hard to build a device to read it out, or write it into a robot.
It’s not that technology isn’t critical. It just isn’t the only thing in play. “People focus on the technology, particularly in Silicon Valley, because that’s what Silicon Valley is about. It’s not about science, it’s about engineering. But the technology, I think, is the easiest, in some sense,” Koch says. “We know that three pounds of excitable brain matter are responsible for seeing, moving, and suffering. These are biological signals that obey natural laws, but that tells us precious little about how trillions of electrical spikes per second streaming over tens of billions of cells constitute a specific sight or sound or motion. That’s limited not by technology but by science.”
That science isn’t finished yet. Neither were reusable rockets before SpaceX, or high-quality electric cars before Tesla. But neither of those advances required anyone to get neurosurgery to be in the first aftermarket test group.