Early speculation of the strategies that Elon’s next company will follow.
Elon Musk recently announced that he’s getting into the business of neural interfacing. He hasn’t announced much else, which means I’m free to speculate on Nueralink’s strategy to achieve it’s ultimate goal: a full-bandwidth interface between humans and computers.
I’m passionate about neural tech, but Neuralink’s success depends on producing a viable business as much as viable technology, so I’ll offer my opinions about the market as well.
The ultimate goal
Elon has shown that he thinks in decades. I imagine even his biggest public goals are just a stepping stones to a larger ambitions - why stop at landing on Mars in 10 years when you can terraform it in 50? Nevertheless, for Neuralink the “ultimate” goal is to build an interface between the human brain and computers. We already have some interfaces - the screen you are reading this on for example - but the biggest challenge is to “increase the bandwidth”. Typing at 80 words per minute won’t be enough soon.
Computers will become smarter than us. This is inevitable at any rate of progress quicker than our own, so increasing the bandwidth is an opportunity for humans to progress at the same rate as computers and mitigates a huge risk that they will “leave us behind” once they become smarter (or worse).
There is one other more subtle goal that most people miss. While bandwidth could certainly be improved, we already have a reasonably high bandwidth solution dating back to 1924 when Hans Berger invented the EEG which measures various different electrical impulses from the brain. The more sensors and the better the connection to the brain, the better the data/bandwidth. The difficulties arise when attempting to interpret the data. Individual brains aren’t entirely the same due to the way they are folded. And that’s just the physical difference - there is an order of magnitude more complication when reading the waves that are emitted. So the secondary ultimate goal is to make sense of the data coming from the brain. This will be an important point later.
The toughest business yet
It initially seems that most of Neuralink’s challenges will be tech related but there are several larger factors that threaten its success. These are: competition; regulation; and public relations.
- Competition - this one is fairly obvious. I’ll talk more about Neuralink’s strategy below but medical is going to be one of the main focuses so the incumbent in this space is BigPharma. The problem here is that pharmaceuticals are ridiculously cheap to manufacture, so there’s a lot of profit to fight competition. This wasn’t as problematic with Tesla or SpaceX where the competition is on a closer cost scale (perhaps a factor of 50 difference rather than a factor of 1000). To side-step the competition, Neuralink will have to focus on problems that are mostly unsolved by Pharmaceuticals.
- Regulation - while frustrating, this provides an opportunity as well since the competition are similarly disabled. In America the FDA regulates anything that is “intended to cure, treat, mitigate, diagnose, or prevent disease in humans” or “intended to affect the structure or function of the human body”. There’s not much room for failure in MedTech so testing and iterative improvement will be extremely critical, however this will inevitably require animal testing. Testing on animals has similar regulations to human regulations - which brings me to the last hurdle:
- Public Relations - people understand clean electric cars. The public find space travel inspiring. They don’t like animal testing and they probably aren’t entirely thrilled with the idea of augmenting themselves with robots, so look out for the inevitable smear campaigns when the aforementioned competition no longer want to play fair.
Also, in previous businesses (automotive, space, energy) the competition had paved the way, but Neuralink is going to be educating the market for most of their products. To do this the public will need to directly understand the benefit of the technology, which usually requires a “killer app” approach (although perhaps Neuralink shouldn’t use the term “killer app”).
Types of interfaces
There are several avenues which Neuralink can take to reach their goal. I’ll categorise neural interfaces into three sections: non-invasive, semi-invasive, and invasive. This is a scale of “do you need to insert something into your body”. I’ll also talk about bi-directional interfaces (from the brain to machine and vice versa) and single-directional interfaces.
The EEG I mentioned in the introduction falls into this category because it doesn’t require any surgery. You merely attach a series of electrodes to the outside of your head.
Other existing non-invasive interfaces include bone conduction which is used to conduct sound through the bones of the skull and transcranial magnetic simulation which uses magnetic stimulation to improve symptoms of depression. These are both single directional interfaces however (machine=>brain) and while these might be appealing, the brain already has high bandwidth input (the eyes). The most appealing problem to solve is the brain=>machine interface because that’s the one with the lowest bandwidth.
Non-invasive interfaces aren’t entirely accurate but they are available to everyone with little regulation. For that reason they are probably the best at generating the large data sets that will be required to solve the secondary goal (understanding the information that the brain produces). This approach is already underway by some enterprising companies and open source initiatives like OpenBCI. I believe there is still a lot to squeeze out of the non-invasive domain - most of the devices are clunky or don’t work so well if the sensors don’t have a wet connection to the scalp - these are easy problems to solve with a focused and financially-backed team.
An invasive device is one that is surgically implanted into the brain. The best example of this is deep brain stimulation (DBS), where “rods” emit an electrical impulse into specific parts of the brain to steady body movements in patients with Parkinson’s Disease.
A cochlear implant is another example. Similar to the bone conduction interface above, it is used to give hearing back to deaf people by by-passing the ear altogether. Since it is set in the skull, there is a great opportunity to add “hot-swappable” functionality to the device so that the implant can continue to improve when new sensors are produced. Cochlear implant patients are a great opportunity for Neuralink.
Invasive devices will probably give the highest quality readings of the brain due to their proximity - essentially the difference between a holding a vibrating phone in your hand and having it in your pocket. The brain is infinitely more subtle than a vibrating phone so you need more sensors closer to truly capture all the various “data” that the brain is emitting. Despite this fact, I think invasive devices are too heavily regulated to keep pace with AI. Also, due to their “permanent” nature there isn’t an easy upgrade path. You can’t just pull some rods out of a brain and put in version 2.0 of the rods due to the risk of infection and permanent damage (i.e. death).
Long term, semi-invasive devices will be the best option for neural interfacing, however there could be a while away. Nobody has truly cracked this yet, but there has been some progress in this domain. Two such interfaces are Neural Lace and Neural Dust.
Neural Lace is a mesh that grows into your brain, which has surprisingly already been tested on mice. It is so thin that it can be injected. The brain is hugely complex however so I have a feeling that something that is as “permanent” as neural lace is too risky. There will be a more achievable interim step where the lace is planted on the outside of the skull, but under the skin. This will provide cleaner readings than an EEG. Ultimately however it suffers from the same problem of invasive procedures where surgery is required - high risk and slow progression.
This is the most promising solution in my opinion and I see it coming in several stages.
First will be a type of semi-permanent dye that can be injected and then scanned by an external device (similar to an MRI). This probably won’t have much of a bi-directional application however but will help to understand what is going on in the brain on 3 dimensions. Since the dye is a liquid it won’t suffer from “pixelation”. Neural Lace will have a defined set of sensors, for example 1000, and therefore can only gather data from 1000 points. This is a bit like measuring the wind by watching a tree - you’re much better off to throw flour into the air and observe the wind in more granular detail.
The next stage will be to inject physical nano-devices which are capable of a bi-directional interface. The difficulty is getting these small enough to allow free flow through the brain, they will have to be on the nano-meter scale.
Neural Dust is the best option because it provides a quicker upgrade path than Neural Lace. Technology is incremental so it makes sense to ensure that the interfaces can be upgraded at any time. With Neural Dust the old versions will be flushed out and shouldn’t require surgery to do so.
Commercial viability and practical uses
The importance of the goal is larger than the risks. I’m glad we have an entrepreneur with a proven track record that recognizes this. Nevertheless, this is a business that has an extremely uncertain path to profitability. Here are some potential paths and applications:
Elon has already mentioned that the initial stages will focus on solving medical issues (which is the same strategy that the main competition is using too). This may include:
- Deep Brain Stimulation for Parkinson’s patients
- Cochlear implants for deaf patients
- Any other terminal or degenerative disease where the risk of procedure is less than the risk of inaction
- Mind-controlled robotic prosthetics are a potential market, however these are better controlled by EMG devices which measure the electrical impulses of muscles rather than the brain
- Controlling wheelchairs for paraplegics - this is a good one since there are a limited set of commands that need to be issued by the brain
- Biofeedback can cognitive enhancement - this is where the brain measurements are displayed back to the participant for them to try to influence the results (sometimes for concentration or meditation, but also can be applied to elite athletes).
- Mood manipulation - the ability to change your mood through electrical impulses (e.g. alpha waves for concentration or delta waves for sleeping)
- Neurogenesis - the ability to regrow brain cells through electrical impulses
Gaming, sports, industrial, and entertainment
The medical path will definitely give the most in-depth data but it won’t give the quantity and diversity required to achieve the other meaningful goal - understanding the brain. For a plethora of data, Neuralink will probably have to focus on consumer markets. Due to the uniqueness of each brain, the electrical impulses emitted from the brain don’t exactly match from one person to another. This means that you need to train the device to fit the task. For example, if the device is required to turn the light on in your house then you would watch the lights training on and off (using a physical switch) and then attempt to do it without the switch. The device measures your brain pattern when you are doing the task physically and when your brain thinks about doing the task it produces a similar pattern, so the device can understand what your brain “looks like” when you try to turn on and off the lights. The device would (hopefully) be self learning, so the more you do a task, the better it gets at interpreting the brain waves for that task. There probably isn’t much money in consumer markets, but they are critical for truly understanding the brain.
- Gaming and movie immersion. The gaming industry is full hackers who are keen to spend their money on immersion into external experiences. They embrace different interfaces with ease and they are constantly pushing the boundaries of new interfaces. A good option would be to ride the Augmented/Virtual Reality wave since it provides another possible interface to hook into. For example, imagine you are playing Counter-Strike and you can “sense” other players behind you through electrical impulses emitted into your brain
- Controlling equipment such as drones or other robotics. Drones are a good starting piece since they have good fail-safes built in and have a limited set of instructions (up, down, left, right, etc).
- Manufacturing - similar to drones, although typically in manufacturing you want little margin for error, so I think this won’t be appealing for a while
- Smart homes - the ability for your home to sense your mood and adjust, and for you to control it without handling a physical device
- Software interfacing - Probably a long way off unless the software is extremely simple. One example might be reminders however. You could “set” a reminder, and then when you go through your list of reminders it could induce some brain waves which would “recall” the reminder.
Finally is the military market. This is the tricky one because although there is definitely a lot of money here, it would almost certainly restrict how the technology (i.e. no sharing). This is a direct contradiction to what Elon believes as he’s already proven with OpenAI and the opensourcing of Tesla’s battery technology. OpenAI was set up specifically to keep the technology out of the hands of the few. However there are many military applications. Unmanned vehicles and drones, cognitive enhancement in soldiers, and general monitoring through nano-technology are a few of the options in this industry.
The future of computing
It’s hard to talk about the future of computing without it sounding like science fiction. This is definitely the most pioneering of Elon’s businesses. A whole new market will need to be created for it to be successful. I hope that the industry can band together in this adventure, because we’re already standing on the shoulders of the tallest giant.