I was excited for Neuralink, then I watched the demo
thenextweb.com
thenextweb.com
I don't see how that is going to be done with stick on sensors.
And yes, some implantable BCI devices have better capability. The point is to make this as cheap as possible... so that it doesn't require a brain surgeon to install. Because brain surgeons are expensive.
Brain implant surgery is incredibly invasive and risky even with brain surgeons.
That said, the demo was mostly for attracting talent - lots of money and a great team will likely get them somewhere. In terms of their long term goals I expect the area that will give them the most trouble is reading/writing interesting stuff on the cortex. Neural coding is really hard and poorly understood.
Neuralink has two goals:
1) short term. better BMIs to to treat serious brain diseases with more bandwidth and more and better electrodes. I can see this happening.
2) long term. Invasive BMI for something other than treating diseases. Better electronics is not going to make invasive brain implants safe for consumer use, they just reduce certain risks.
To solve (2) Neuralink must solve fundamental issues in medicine related to body implants and brain surgery. It would be groundbreaking and probably revolutionize medicine even outside brain implants. Neuralink researchers might get Nobel in Medicine.
Breakthrough in brain surgery. To get FDA to approve brain surgery (making holes in the dura mater) even brain surgery to open and close the brain for no good medical reason would be insanely hard. Any form of brain surgery has lots of risks. Bleeding in the brain, seizure, scarring of the brain, immediate infection risk and late infection risk. Inserting lace with thousands of electrodes is huge operation even if you could do with with endoscope.
Breakthrough with implants. With the BMI ther are issues with coagulation and constant low level inflammation. It's both health risk and gradually degrades the effectiveness of electrodes. Electrode must be in contact with tissue and interact with it. There is risks even with teeth implants, artificial knees and hips. They collect bacteria around them and they are just inert objects.
Just like with any other implant, the risk must justify the benefit. Cochlear implants have 3% rate for major complications, like life threatening meningitis or facial paralysis (caused by nerve damage). Potential benefit justifies the risk.
The other organs in the ear are the vestubular organs which give you a sense of balance.
The little wire in the implant has electrodes along it's length that shock the 'damaged' cells in the cochlea, bypassing how those cells sense sound, and just directly stimulating them.
The reason that it's a bit 'crude', is that the shocking is not specific to the nerves, it just shocks a bunch of those sensing cells. The reason for this is a bit involved and is due to the tonotopy (sound-map) of the cochlea.
The implantation is relatively simple for a trained surgeon. It can be done on small children.
https://www.cochlear.com/us/en/home/diagnosis-and-treatment/...
Even if the surgery does not result bacterial infection, insertion injury and the pretense of a foreign object often leads to persistent low level inflammation.
There is so much basic research questions that simple engineering is not going to solve.
Where are all these people apparently doing long term, low latency wireless 1024 channel ephys with stim capability?
> “Neuralink are progressing through the steps that have been taken with previous neural interface technology. They have moved from rodents to a large mammal (pigs), and are apparently seeking FDA approval for a human trial. I had some sympathy with Elon, having also had experience of animal experiments that don’t work perfectly the day we have visitors in the room! But proving the safety of new biomedical implants in animals is always a vital step towards a clinical trial.
> “I don’t think there was anything revolutionary in the presentation, but they are working through the engineering challenges of placing multiple electrodes into the brain. In terms of their technology, 1024 channels is not that impressive these days, but the electronics to relay them wirelessly is state-of-the-art, and the robotic implantation is nice.
> “But the biggest challenge is what you do with all this brain data. The demonstrations were actually quite underwhelming in this regard, and didn’t show anything that hasn’t been done before (e.g. decoding limb position during walking). There is a big difference between recording brain cells and ‘reading thoughts’, especially when it comes to higher-level cognitive functions that we don’t understand as well. The idea of ‘writing to the brain’ is even more questionable – there are fundamental limitations to targeting specific networks of neurons in a meaningful way using electrical stimulation.
> “So in summary I would say this is solid engineering but mediocre neuroscience. Finally, I think it is unfortunate that they are presenting their work in this way, rather than publishing peer-reviewed papers that would allow their claims to be scrutinised, but I guess this is something that we will have to get used to as neural interfaces move from the academic to the commercial sector.”
https://www.sciencemediacentre.org/expert-reaction-to-elon-m...
Well, and this is exactly what needs to be done to further this area as it is the part that's not completely addressed by academia. Any kind of academic research obviously focusses on generating one-time results for publication.
I would love some expert take on this, but my first reaction was „12000 sounds like a lot - they all work? That’s cool“.
Also check out this wait but why article which I think explains the resolution pretty well: https://waitbutwhy.com/2017/04/neuralink.html
My guess is that if Neuralink does nothing more than provide a more sophisticated tool for investigating the brain than the current state of the art then it will have been a success.
A quick google suggests the highest resolution fMRIs currently have about 1 million voxels of spatial resolution at a sample rate of about 1/sec. EEGs can measure electrical activity but only at very low resolution and on the surface.
I've seen single-unit recordings done with electrodes and simultaneous Ca2+ imaging with a 2-photon microscope. You trade real-time speed and single unit specificity for a loss of knowledge about the population.
MEG can give you real-time summed behavior, but the inverse problem is indeterminate so you're less sure of _where_ that activity is coming from.
"what's the good tool to measure brain activity" is very often a function of "what do i want to find out about the brain?"
Also, CIs are not brain implants.
In 2016 I got the chance to do a few month’s work on some cortical recodings in mice with a 512-electrode patch, and when i left the lab was complaining about how to modernize up to 1024 while still keeping their calcium imaging working well.
To do bee stuff.
It just show a scale off human brain. And while its brain size is not equivalent to the intelligence/capabilities, pluggin into our brain is going to be a bit more complicated than 12000 wires.
The tone of this article is negative and I instinctively want to defend Neuralink because I hope that we can eventually achieve all of those sci-fi ambitions. But I think that's the wrong approach, and possibly insulting to researchers who have been working on these problems for decades only to be overshadowed by what I have to admit was not the most impressive demo.
However, this event does attract attention, so what I'm really hoping for is an influx of long-read articles and in-depth analyses coming from experts in the various fields surrounding Neuralink's aims as a result of the hightened interest in the topic.
I count body modification as something else.
They don't really have that much competition apart from Soyuz at the moment.
There is huge support in US to have a alternative to Soyuz.
Also NASA is prisoner of government corruption to the point where NASA operations are crippled in their capabilities - ie you get funding only if you keep unnecessary bureaucracy running for sake of some governor who won by promising space jobs.
SpaceX took all motivated and knowledgeable people gave them money to do what NASA was not allowed to do.
I am not trying to take away from their successes. The reusable modules landing in sync are amazing feat of engineering.
And SpaceX has actually working 'product', even more solid than Tesla (as a car its expensive, low volume, with a lot of issues compared to what we come to expect of a car).
Other stuff that Musk is trying to push are fairy tales, hyperloop is bad idea from the ground up. Boring company is boring. Its just bad implementation of a subway. And this neurolink is so far away from being anything usable, they are not even clear what it suppose to be. "Stick in some wires we will see what happens revolutionary tech TM."
But SpaceX business was waiting to happen. Musk was smart enough to throw money at it. So I wouldn't use spaceX as example of some godlike Musk abilities to create something out of nothing.
Software will always be vulnerable and it’s just a matter of time until someone gets eliminated by this vector - or worse, think about a wormable exploit aimed at mass adopted Neuralinks to mass fry people..
[1] https://www.wired.com/story/pacemaker-hack-malware-black-hat...
Terrorism, and that is it as far as I can see. In cyber up until know, criminals and espionage seem like much bigger threats than terrorism. We are introducing a new attack surface here, and it makes sense to consider whether we want to. But it is not an attack surface that seems likely to attract many attackers.
* Change in opinion on a given topic by the intruder group paid off by some government/megacorp/group. i.e., voter/social mindspace on demand.
- Change of opinion of critical lynch-pin individuals.
* Personality graft
* Surveillance
- Thought Policing
- (Corporate) Espionage
etc...
[0] Science Fiction
[1] Ghost in the Shell
Heck, maybe messing with the rate for a short while is harmless but scary, and a good time to show ads for some pills...
So they showed progress and proved they are actually a lab and actually perform research. That being said, there was absolutely nothing in the demo remotely related to a "brain-computer" interface, nor are they likely to get there for many, many years. That's the speed when doing hard science, and pumping more money only has a marginal efect.
Seems a bit harsh.
lets say i have a device that can encode activity of every neuron in my brain simultaneously. i think some variation of the thought "elephant," and the neuralink creates some digital representation of these signals.
then i send that digital encoding to your brain's neuralink.
how on earth are you going to think of ANY elephant, let alone my elephant? the structural nature of our brains is certainly different, the network of activated neurons aren't isomorphic, certainly not identical, right?
i am by no means an expert, but i dont see how you can ever take a digital signal and turn it into a conscious meat thought.
motor signals seem somewhat more tractable
The promise seems more that ideas can be transmitted with higher bandwidth. My signals -> word1, send word1 to you, word1-> your signals for word1 would be likely. Both of us would have to the system for word1, but if we could exchange it much faster than speaking, maybe even reading.
My issue with the system is that the line between reading and writing signals is very thin, and this these are hackable devices. Once the device learns a person's brain signals to make the legs run, it's easy to replay them on a bridge when facing sideways. There's no technology right now that's foolproof, and I'm not sure how any company can boast that it's secure enough — even if encryption is strong laws can still be enacted to force companies to hand over keys or install backdoors.
Both brains sensors are calibrated on the same images, then a common representation can be found as intermediary. It has been done with neural nets translating into a common intermediary language, thus needing N encoders/decoders instead of (N-1)^2 separate networks. In the easiest case it can be a simple linear transform to map one vector space into another.
you'd need to train a bidirectional adapter for every person instead of (N-1)^2 adapters for each pair of people
I'll start listening when they have a working definition of what a "thought" is or what "translating" it into a "universal language" even means.
Yes but he also launches rockets into fucking space; that is to say, he may be selling snake oil, but he’s also selling legitimate magic beans.
It’s the latter that’s causing some concern I’d think.
Autonomous driving is still a relatively young beast. I don't have specific numbers infront of me, but I imagine mile for mile, autonomous vehicles are still safer than humans.
Planes used to fall a lot in the beginning, computer hardware was once finicky and unreliable, even rockets used to blow up before they got it right. Takes time, and sometimes it takes early adopters trust to get to safety.
Saying a private company put that vehicle in space is technically true but also at the same time highly misleading.
Like the author, I would also be upset if I believed Musks promises of a brain-computer interface, only to see them dashed. But Neuralink is already an amazing improvement, and I am still excited to see what will come of it.
As to him making claims about the future, perhaps he believes it. In any case, there is a history of his companies doing “the impossible”; economical reuse of rockets was considered by many to be hot gas by him. Conmen don’t usually deliver to that level.
Separately, you might be interested to know that apparently the build quality has gone up quite a bit recently: https://youtu.be/TOrrdqje9Og
- Coin-sized brain implant, flush with skull, no external wires.
- 1024 channels EEG
- Electrode insertion with minimal brain-matter loss
- Analysis and prediction of brain-activity (node-beep and motion joint positions)
Drama queen.