Elon Musk doesn’t think his newest endeavor, revealed Tuesday night after two years of relative secrecy, will end all human suffering. Just a lot of it. Eventually.
At a presentation at the California Academy of Sciences, hastily announced via Twitter and beginning a half hour late, Musk presented the first product from his company Neuralink. It’s a tiny computer chip attached to ultrafine, electrode-studded wires, stitched into living brains by a clever robot. And depending on which part of the two-hour presentation you caught, it’s either a state-of-the-art tool for understanding the brain, a clinical advance for people with neurological disorders, or the next step in human evolution.
The chip is custom-built to receive and process the electrical action potentials—“spikes”—that signal activity in the interconnected neurons that make up the brain. The wires embed into brain tissue and receive those spikes. And the robotic sewing machine places those wires with enviable precision, a “neural lace” straight out of science fiction that dodges the delicate blood vessels spreading across the brain’s surface like ivy.
If Neuralink’s technologies work as Musk and his team intend, they’ll be able to pick up signals from across a person’s brain—first from the motor cortex that controls movement but eventually throughout your think-meat—and turn them into machine-readable code that a computer can understand. It might use them to control a computer or a prosthesis, to someday even feed information back to help the blind see, or to create entire virtual Matrixes inside your mind. “All this will occur I think quite slowly,” Musk said from the stage. “It’s not as if Neuralink will suddenly have this incredible neural lace and take over people’s brains. It will take a long time.” But after tests, and FDA approval, and more advances, this tech could be the thing that lets people commune with the ultrasmart artificial intelligences Musk is convinced are on the way. “Even in a benign AI scenario we will be left behind,” he said. “With a high-bandwidth brain-machine interface, we can actually go along for the ride. We can have the option of merging with AI.”
The WIRED Guide to Artificial Intelligence
This is all pretty on-brand for Musk. As the guy who runs the electric-car company Tesla and the rocket company SpaceX, Musk has gotten very good at—in trouble, even, for—taking impressive technological achievements and, well, maybe not hyping them, but let’s say skipping all the way to the end of their speculative narrative arcs. It’s not enough to have superslick electric cars; no, they’re also going to drive themselves. That rocket isn’t just going to ferry cargo to a space station; no, it’s going to take people to Mars. How exciting!
Since The Wall Street Journal revealed Neuralink’s existence two years ago, the tech and neuroscience worlds have buzzed about what Musk’s team of brain-machine interface experts was up to. Other companies, including Kernel and Facebook, announced they, too, were working on the technology, which has so far been used only in research and rare clinical settings. Darpa, the US government’s advanced-science division, has been funding brain-computer interface work since the 1970s, and the agency has been part of the government-wide Brain Research through Advancing Innovative Neurotechnologies (yes, the acronym is also “Brain”) since 2013.
So it’s hard to know exactly how to calibrate Musk’s claims for a device that he plans to eventually stick into healthy people’s brains. “We hope to have this aspirationally in a human patient by the end of next year,” Musk said. The first volunteers, he hopes, will be people with quadriplegia, willing to have four chips implanted, three in the motor cortex of the brain (roughly running from above the ear to the top of the head) and on providing closed-loop feedback to the somatosensory cortex. That’s even though, according to an article distributed at the presentation—and not peer-reviewed—the Neuralink technology is so far only in the heads of 19 rats, and even then with only 87 percent of the electrodes successfully inserted. The FDA is going to want more than that before it approves human use.
And, sure, there’s more. A public records request from WIRED in April 2019 found that Neuralink is licensed to have hundreds of rats and mice in its research facilities. In a seemingly unplanned moment at the Cal Academy, Musk also acknowledged that Neuralink’s research had progressed beyond rodents to non-human primates. It’s only because of a records request filed by Gizmodo that Neuralink’s affiliation with the primate research center at UC Davis is public knowledge. That affiliation has apparently progressed: “A monkey has been able to control a computer with its brain, just FYI,” Musk said during the Q and A after the presentation.
His team seemed as surprised and discombobulated by the announcement as the audience. “I didn’t know we were running that result today, but there it goes,” said Max Hodak, president of the company, on stage next to Musk. (Monkeys have controlled computers via BCIs before, though presumably this would be the first time one used Neuralink.)
(A separate records request from WIRED in August of 2018 reveals that Neuralink re-upped its deal with UC Davis in June of that year, a month after the Gizmodo article. That relationship hasn’t always been entirely cordial; emails obtained by WIRED show that in June of 2018 John Morrison, director of the California National Primate Research Center at UC Davis, complained that Neuralink was trying to poach UC Davis staffers. “I realize that this is routine practice in the private sector, but I am a little surprised since my understanding is that there was an interest in developing scientific collaboration between Neuralink and the CNRPC,” Morrison wrote to a redacted contact apparently at Neuralink. “Hiring away personnel does not build a relationship.”)
Adam Rogers writes about science, technology, and miscellaneous geekery for WIRED.
The hardware Neuralink has developed is impressive. External, noninvasive technologies like electroencephalograms (EEGs) or functional magnetic resonance imaging tend not to have the kinds of resolution—across the brain and over time—to do things like control a computer. But the inside of the brain is an unfriendly place when it comes to electrodes, a briny soup that eats away at the hard, pointy bits neuroscientists have used for decades to listen in on synaptic chit-chat. Immune responses coat those electrodes with glial cells, defensive gunk that eventually renders them inoperative. The brain’s natural movements, its sloshing around and pulsing in time with heartbeats and breath, mean that implanted electrodes move around, too, eventually sliding off the nodes they’re meant to target. And maybe worst of all, unlike the cured, prepared specimens you might have seen in science class, living brains have the texture of Jell-O, whereas the kinds of electrodes best at picking up neural signals tend to be stiff and solid. Old-school electrodes have been known to damage brain tissue and go off-target as the brain moves around.
Neuralink goes in a newer direction, one neuroscientists only started to come around to in the last decade or so. The electrodes are made from a soft polymer. The thin threads that connect them to the chip allow, for now, more than 1,500 individual channels of recording, covering more neurons overall; that’s widely seen as a good thing in terms of collecting enough signal to interpret. But the threads are too small for a human hand to insert, so Neuralink engineered a robotic system to insert individual threads in preset locations and at preset depths. Those then send signals wirelessly to a receiver that a person would wear like a behind-the-ear hearing aid, via Bluetooth. (The rats send their data via USB-C.) “The devices that we’re talking about, because of their high bandwidth and the ability to tailor the location of each electrode to a person’s individual anatomy, should be able to reach anywhere in the motor cortex,” said Philip Sabes, Neuralink’s senior scientist. “That would give us access to any movement a person thinks about.”
At first, Sabes said, that’d mean the ability to control a computer keyboard or a mouse (after training via a smartphone app). Musk said he’d hope that someone could get up to 40 words per minute typing, a goal that would require remarkably low latency in the chip’s processing speed. And then later? Control of 3D avatars or complicated prosthetic arrays, maybe even the ability to receive haptic input—perceiving textures or pressure—and the kind of signals that deep-brain implants send to quiet the tremors of Parkinson’s disease or the compulsions of obsessive disorders. To be fair, Sabes didn’t show any of this data, and it’s not in the white paper the company handed out, either. It’s all, as Sabes said, aspirational. That was right before Musk said that if two people both had Neuralinks, they’d “effectively have a really high-bandwidth telepathy … potentially a new kind of communication, a conceptual telepathy. It would also be consensual.”
The hardware could indeed be a leap forward for research. The resolution is high, though other groups have achieved numbers in the same ballpark, such as in a multi-institution project called Neuropixels. “The problem was always the backend, which is just not a fun thesis project, so it had to be done outside universities,” says Polina Anikeeva, a materials scientist who works on neuroelectronics at MIT. “Designing a backend that would be reasonably sized and accommodate a few thousand channels is an engineering challenge unsuitable for academic environments and, most importantly, budgets.”
That, along with a reliable implantation robot, might even bring some Silicon Valley–style disruption to the world of brain-machine interfaces. “The robot looks real, the ASIC [application-specific integrated circuit] looks real, the implantable package looks real,” Andrew Hires, a neuroscientist at the University of Southern California, tweeted during the presentation. But, he said, “The closed-loop applications are vaporware.” Which is to say, beyond existing deep-brain stimulation technologies and rudimentary input, the stuff about feeding input back into a brain—writing as opposed to just reading—is still as far away as Mars.
People just don’t know enough about how the brain works to impinge upon it, to make a brain do something it didn’t plan on. Sabes talked about stimulating specific parts of the “maps” in the visual cortex for things like edges and movement, to create projections on the inside surface of the mind’s eye. “Our understanding of brain circuits and ability to interpret neural signals is rather rudimentary, and any technology developed right now would better serve basic neuroscience before we can think about applying it in a medical context,” Anikeeva says.
A decade ago, when a team of researchers at Baylor College of Medicine tried to induce color percepts in a person with an electrode array implanted as part of treatment for recurring seizures, they couldn’t do much better than induce bluish-purple, and even that was amazing. “We have basic knowledge, a lot of basic knowledge, and a lot of imaging techniques,” says Nataliya Kosmyna, a computer scientist working on brain-computer interfaces at the MIT Media Lab. “But where do you want to write, to which part? What do you want to have in that signal?”
Those problems might come to seem small compared to finding out just how biocompatible and long-lasting those polymer electrodes are in a living brain—or how results in rats will translate to primates. Engineering challenges remain. "Are you talking about the ultimate end goal of what Elon is saying, a third layer of interface for the brain? Is it going to do that? No, not even close," Hires says. "But is it a step toward that, and can it advance the field in a meaningful way? Well, as long as they can get it through safety and regulatory approval, I think so."1 Figuring out how the brain works is one of the core ways science can help us humans understand ourselves better, and it might even humanize the coming world of machines and machine learning, or conversely machine-ize a world that’s still all too human. But Neuralink isn’t there yet. That’s all still aspirational.
With additional reporting by Tom Simonite
1Updated 7/17/19 9:35 AM PT with Hires' quote