Latest / Elon Musk Podcast / Elon Musk's Neuralink 2025 summer update

Transcript

Elon Musk's Neuralink 2025 summer update · Elon Musk Podcast

0:01
Hey everybody. Welcome back to the Elon Musk
0:04
Podcast. This is a show where we discuss
0:07
the critical crossroads, the Shape, SpaceX, Tesla X, The
0:11
Boring Company, and Neuralink. I'm your host, Will Walden.
0:19
Hello everybody. My name is Alex Toddling.
0:21
I'm the second participant in the Neuralink study, but I'm
0:25
here to count us down to the demo in 54321.
0:46
Hi everyone, welcome to the Neurolink presentation.
0:53
This is an update for the progress of the neural link
0:57
team. It's been an incredible amount
0:58
of progress. This is we're going to start off
1:03
high level, generally describing what neural links doing.
1:08
And then we're going to get have a very deep technical dive so
1:11
you can actually get an understanding of what exactly
1:14
we're doing at a granular level and what we can do to enhance
1:18
human capabilities and ultimately build a great future
1:22
for humanity. So that's a that's a neural
1:26
inspiring. It's funny thing that me talking
1:29
right now is a bunch of neurons firing that then result in
1:34
speech that you hear that cause neurons to fire in your brain.
1:42
Yeah, part of part of this presentation is about
1:45
demystifying the the brain. It is a remarkable organ.
1:50
I mean, we are the brain basically when you say you, that
1:54
really is you're the brain. Like you can, you can get a, a
1:59
heart transplant, you can get a kidney transplant, but I don't
2:02
know anyone who's gotten a brain transplant.
2:05
So you are your brain and your experiences are these neurons
2:13
firing with the trillions of of synapses that somehow lead to
2:20
conscious comprehension of the world.
2:23
This is something that we have only begun to understand.
2:28
We're really just barely at the beginning of understanding of
2:30
what is the nature of consciousness.
2:34
And I've thought a lot about what, what is consciousness?
2:36
What is it? Where does consciousness arise?
2:43
Because if you start at the beginning of the universe,
2:46
assuming physics is true, the standard model of physics is
2:49
true, then you have this Big Bang.
2:54
You know the matter condensing into stars, those stars
2:58
exploding. A lot of the the atoms that are
3:01
in your body right now were ones at the centre of stars.
3:05
Those stars exploded, recondensed.
3:08
Fast forward 13.8 billion years and here we are.
3:15
And somewhere along that very long journey, to us at least,
3:22
consciousness erodes or the the molecules started talking to
3:27
each other. And it begs the question of what
3:35
is consciousness is, is everything conscious?
3:39
Maybe it's hard to say. We're along that line, that
3:43
there's no sort of discreet point where consciousness didn't
3:48
exist and then suddenly does exist.
3:51
It seems to be maybe you have a condensation of matter that has
3:56
a, a density of like, we don't know what the real, the real
4:01
answer is. We don't know what consciousness
4:02
is. But with the neural link and the
4:07
progress that the company's making, we'll begin to
4:10
understand a lot more about consciousness and what does it
4:14
mean to, to be along the way, We're, we're going to solve a
4:23
lot of, a lot of brain issues where the brains get injured or
4:31
damaged in some way or didn't develop in quite the right way.
4:35
But there's, you know, there's, there's a lot of brain and spine
4:40
injuries that go so along the way.
4:42
And I do want to emphasize that this is all going to happen
4:45
quite slowly, meaning you'll, you'll see it coming.
4:48
Sometimes people think that suddenly there will be vast
4:51
numbers of neural links all over the place.
4:55
This, this is not going to be sudden.
4:57
You'll be able to watch it happen, you know, over the
5:00
course of several years and, and we go through exhaustive
5:06
regulatory approvals. So this is not something that
5:09
we're just doing there by ourselves without government
5:13
oversight. We're we work closely with the
5:16
regulators every step of the way.
5:18
We're very cautious with, with neural links in humans.
5:23
That's the reason we're not moving faster than we are is
5:26
because we're, we're taking great care with, with each
5:28
individual to make sure we, we never miss.
5:31
And so far we haven't. And I hope that continues into
5:34
the future. Every single one of our implants
5:36
in humans is working and working quite well.
5:39
And you'll get to hear from some of the people that have received
5:43
the implants and cured in their words.
5:47
So what we're we're we're creating here with a neural link
5:49
device is a generalized input output technology for the brain.
5:57
So it's how do you get information into or out of the
6:03
brain and do so in a way that does not damage the brain or,
6:09
you know, cause any negative side effects.
6:12
So it's a very hard problem and generally the the reactions I've
6:17
seen to this range from it's impossible to it's already been
6:21
done before those those people should meet.
6:25
Actually the reality is that there actually have been limited
6:32
range of computer interfaces for several decades.
6:38
On a very basic basis, just what we're doing with neural link is
6:43
dramatically increasing the bandwidth by mayors magnitude.
6:49
So you can, you can a human bandwidth output is less than
6:54
one bit per second over the course of a day.
6:57
So there's 86,400 seconds in a day.
7:00
It's very rare for a person to do more than 86,400 bits of
7:04
output per day. You'd have to be really talking
7:07
a lot or typing all day and you might exceed that.
7:12
So what we're talking about here is, is going from maybe one bit
7:16
per second to ultimately megabits and then gigabits per
7:20
second and the ability to do conceptual consensual telepathy.
7:29
Now the the input to the brain is much higher because of,
7:33
especially because of vision. Depending upon how you count it,
7:37
it, it might be on the order of a megabit or in the megabit
7:41
range for input primarily due to sight.
7:47
So, but even for input, we, we think that can be dramatically
7:51
increased to, to the Gigabit plus level.
7:57
And, and, and a lot of the, the thinking that we do is which we
8:02
take a concept in our mind and we compress that into a small
8:06
number of symbols. So when you're trying to
8:08
communicate with somebody else, you're actually trying to model
8:12
their mind state and, and, and then take perhaps a quite a
8:17
complex idea that you have, maybe even a, a complex image
8:21
or, or scene or kind of mental video and try to compress that
8:25
into a few words or a few keystrokes.
8:27
And it's necessarily going to be very lossy.
8:30
Your ability to communicate is very limited by how fast you can
8:34
talk and how fast you can type. And what we're talking about is
8:38
unlocking that potential to enable you to communicate, like
8:43
I said, thousands, perhaps millions of times faster than is
8:46
is currently possible. This is an incredibly profound
8:50
breakthrough. This would this would be a
8:53
fundamental change to what it means to be a human.
8:59
So we're, we're starting off with reducing human suffering.
9:03
So or, or addressing issues that people have, say if they've been
9:07
in an accident or they have some neural disease that's
9:13
degenerative so they're losing capability to move their body or
9:17
some some kind of injury essentially.
9:20
So enabling our first product is called telepathy and that
9:24
enables someone who has lost the ability to command their body to
9:30
be able to communicate with the computer and move the mouse and
9:33
and actually operate a computer with roughly the same dexterity,
9:37
ultimately much more dexterity than a than a human with working
9:41
hands. Then the our next product is, is
9:46
blind sight, which will enable those who have total loss of
9:50
vision, including if they've lost their eyes or the optic
9:52
nerve or maybe have never seen were blind, even blind from
9:56
birth to be able to see again, initially low resolution, but
10:00
ultimately very high resolution and, and then in multiple
10:04
wavelengths. So you could be like Geordie La
10:05
Forge in Star Trek and you can see in radar, you can see an
10:09
infrared, ultraviolet, superhuman capabilities,
10:13
Severnetic enhancement essentially.
10:17
And then along the way, this should help us understand a lot
10:22
more about consciousness. What does it mean to be a
10:25
conscious creature? We'll understand vastly more
10:30
about the nature of consciousness as a result of
10:32
this. And then ultimately, I think
10:34
this helps mitigate the civilizational risk of
10:37
artificial intelligence. We're, we are actually already,
10:45
we're already sort of have 3 layers of thinking.
10:48
There's the limbic system, which is your kind of your instincts,
10:52
the cortic, your cortical system, which is your higher
10:56
level planning and thinking. And then the tertiary layer,
10:59
which is the computers and machines that you interact with,
11:03
you like your phone, your all the applications you use.
11:07
So people actually are already a Cyborg.
11:10
You can maybe have an intuitive sense for this by how much you
11:15
miss your phone if you leave it behind.
11:19
Leaving your phone behind is like, it's almost like missing
11:21
limb syndrome, but your phone is somewhat of an extension of
11:26
yourself, as is your computer. So you, you already have this
11:30
digital tertiary layer, but the bandwidth between your cortex
11:34
and your digital tertiary layer is limited by speech and by and
11:39
by how fast you can move your fingers and how fast you can
11:41
consume information visually. So, so, but I think it's
11:48
actually very important for us to address that input output
11:52
bandwidth constraint in order for the collective will of
11:56
humanity to match the will of artificial intelligence.
12:03
That's my intuition at least. So let's see and, and what what
12:09
this presentation is mostly about is attracting smart humans
12:18
to come and work with us on this problem.
12:21
So this is not a presentation to raise money or anything like
12:25
that. We're actually, you know, very
12:28
well funded. We have a lot of great
12:30
investors. Some of the smartest people in
12:33
the world are invested in neural link, but we we need smart
12:38
humans to come here and help solve this problem.
12:41
So with that, let's let's proceed.
12:55
Hey everyone, my name is DJ, my Co founder and president of
13:01
Neuro Link. And as Elon mentioned, well,
13:04
actually we're standing in the middle of our robot space.
13:07
We have a stage set up, but you know, this is actually where
13:12
some of the next generation most advanced surgical robots are
13:14
being built. So welcome to our space.
13:24
It's important to highlight that this technology is not being
13:26
built in the dark. This is not a secret lab where
13:30
we're not sharing any of the progress.
13:32
In fact, we're actually sharing, you know the progress very
13:35
openly and as well as also telling you exactly what we're
13:38
going to be doing. And we're hoping to progress on
13:41
that as as diligently and as safely and as carefully as
13:46
possible. So to start off, 2 years ago
13:50
when we did our previous fundraising round, we outlined
13:53
this path and timeline to 1st Human.
13:57
And we currently have a clinical trials in the US for a product
14:00
that we call Telepathy, which allows users to control phone or
14:04
computer purely with their thoughts.
14:06
And you're going to see how we do this and what the impact that
14:09
this has had. And not only have we launched
14:13
this clinical trial, but as of today, we have not just one, but
14:18
7 participants and we have an approval.
14:27
And we also have an approval to launch this trial in Canada, UK
14:32
and the UAE. So I guess before we dive into
14:43
what this technology is and what we built, but I wanted to
14:47
quickly share a video with you guys of when our first five
14:50
participants met each other for the first time, so.
14:53
Here you go. All right, we have everyone
14:56
together. What's up guys?
14:58
Thanks everybody for joining. Definitely want to introduce all
15:01
of you. Yeah, I'm Nolan, AKA P1.
15:04
My name's Alex. I am the second participant in
15:07
the Neural link study. I am Brad Smith, the ALS Cypler.
15:12
P. Three, my name is Mike.
15:16
G4 ALS like Francis. Yeah, I'm RJ, I'm P5, and I
15:27
just, yes, I'm trying to do this one to the team here.
15:30
So yeah, appreciate it. And all them Trailblazer, you
15:33
know, somebody's got to get first, man.
15:35
That was you. Appreciate that.
15:36
What's been your favorite thing you've been able to do with the
15:39
neural link so far? I've just had a good time being
15:41
able to use it as I travel flying and draw a little
15:45
mustache on a cat. Had a lot of fun doing that.
15:48
I mean, I've just had a good time playing around with it.
15:50
Oh, you know what? I do know what my favorite BCI
15:53
feature is? Probably not a feature, but I
15:56
just I love web grid more than I love anything in my life
16:00
probably. I think I could play that game
16:03
non-stop forever. Has to be Fusion 360.
16:07
Being able to design parts, design the hat logo with the
16:11
BCI. That's what's up.
16:15
Pretty sweet. That's sweet.
16:16
Yeah, yeah, I have a little Arduino that takes input from my
16:23
quad stick, converts it into APPM signal to go to ARC truck.
16:30
Cool Little Rock crawler. Well, with the BCII.
16:36
Wrote code, can you? To drive the plane with the quad
16:41
stick, that's awesome. The best thing I like about an
16:46
airline is being able to continue.
16:51
To provide for my family and continue working.
16:58
I think my favorite thing is probably been able to turn on my
17:02
TV. Yeah, like the first time in 2
17:05
1/2 years I was able to do that. So it's pretty sweet.
17:07
But I like shooting the hobbies. That's that's kind of nice.
17:11
Excited to see what BC is got going on.
17:14
I got a question. What's your shirt say?
17:16
I said. I'd do a thing called whatever I
17:18
want. Now, one of the major figure of
17:32
merits that we have is to keep track of monthly hours of
17:35
independent PCI use. Effectively, are they using the
17:38
PCI and not at the clinic but at their home.
17:42
And what we have noticed and this is a plot of all of the
17:46
different participants, first five participants and their
17:48
usage per month over the course of the last year and a half.
17:53
And we're averaging around 50 hours a week of usage and in
17:57
some cases peak usage of more than 100 hours a week, which is
18:00
pretty much every waking moments.
18:09
So I think it's been incredible to see all of our participants
18:13
demonstrating greater independence through their use
18:15
of BCI. Not only that, we've also
18:20
accelerated our implantation cadence as we've amassed
18:23
evidence of both clinical safety as well as value to our
18:26
participants. So to date, we have 4 spinal
18:29
cord injury participants as well as three ALS participants with
18:33
the last two surgeries happening within one week of each other.
18:39
And we're just beginning. This is just tip of the iceberg.
18:43
Our end goal is to really build a whole brain interface.
18:47
And what do we mean by whole brain interface?
18:50
We mean being able to listen to neurons everywhere, be able to
18:54
write information to neurons anywhere, be able to have that
18:58
fast data wireless transfer to enable that high bandwidth
19:01
connection from our biological brain to the external machines
19:06
and be able to do all of this with fully automated surgery as
19:11
well as enable 24 hours of usage.
19:16
And towards that goal, we're really working on three major
19:19
product types. Elon mentioned earlier that our
19:21
goal is to build a generalized input output platform and
19:25
technology to the brain. So to afford the output portion
19:29
of it, which is extremely slow through our meat sticks as as
19:33
Elon calls them, neat hands that are holding the mics.
19:41
We're starting out with helping people with movement disorders
19:44
either through where they lost the mind body connection either
19:47
through a spinal cord injury, ALS or a stroke.
19:50
Be able to regain some of that digital as well as physical
19:53
independence through a product that we're building called
19:55
Telepathy. And this is our opportunities to
19:59
build a high channel read and output device.
20:03
On the input side of things, there's.
20:06
Opportunities for us to help people that have lost the
20:08
ability to to see be able to regain that site again through a
20:13
product that we're calling blind sight.
20:15
And this is our opportunity to build high channel right
20:18
capabilities. And last but not least, be able
20:22
to also help people that are suffering from neurological
20:27
debilitating dysregulation or psychiatric conditions or
20:30
neuropathic pain. By inserting our electrodes in
20:35
reaching any brain regions to be able to insert them not just on
20:39
the cortical layer but into the sulk eyes as well as deeper
20:43
parts of the brain. The so-called limbic system to
20:47
really enable better opportunities to just regain
20:49
some of that independence. Our North Star metrics is 1.
20:54
Increasing the number of neurons that we can interface with.
20:58
And 2nd, to expand to many diverse area any parts of the
21:02
brain, starting with microfabrication or lithography
21:06
to change the way in which we can actually increase the number
21:09
of neurons that we can see from a single channel.
21:12
And also doing mixed signal chip design to actually increase the
21:17
physical channel counts to increase more neurons that we
21:22
can interface to, to, to sort of allow more information from the
21:27
brain to the outside world. And then you know, everything we
21:30
built from day one of the company has always been read and
21:34
write capable. And with telepathy, our first
21:37
product, the focus has been on the read capabilities or the
21:40
output. And we want to hone in on our
21:43
write capability and also show that through accessing deeper
21:47
regions within the, the, the visual cortex that we can
21:50
actually achieve functional vision.
21:52
Let's. Go.
22:04
So now just to step you through what the product evolution is
22:07
going to look like in the next three years.
22:09
Today what we have is 1000 electrodes in the motor cortex,
22:14
the part of the small part of the brain that you see in this
22:16
animation called the hand knob area that allows participants
22:20
control computer cursors as well as gaming consoles.
22:24
Next quarter, we're planning to implant in the speech cortex to
22:28
directly decode attentive words from brain signals to speech.
22:33
And in 2026, not only are we going to triple the number of
22:42
electrodes from 1000 to 3000 for more capabilities, we're
22:46
planning to have our first blind side participant to enable
22:50
navigation. And in 2027, we're going to
23:01
continue increasing channel counts, probably another triple,
23:04
so 10,000 channels and also enable for the first time
23:10
multiple implants, so not just one in motor cortex, speech
23:14
cortex or visual cortex, but all of the above.
23:20
And finally, in 2028, our goal is to get to more than 25,000
23:25
channels per implant, have multiple of these, have ability
23:29
to access any part of the brain for psychiatric conditions, pain
23:33
dysregulation and also start to demonstrate what it would be
23:37
like to actually integrate with AI.
23:47
And all this is to say that we're really building towards
23:51
set of fundamental foundational technology that would allow us
23:54
to have hundreds of thousands, if not millions of channels with
23:57
multiple implants for whole grain interfaces that could
23:59
actually solve not just these debilitating neurological
24:02
conditions, but be able to go beyond the the limits of our
24:06
biology. And this vertical integration.
24:08
And the talent and team that we have at Neural Link has been and
24:11
will continue to be the key recipe for rapid progress that
24:14
we will be making. Just to recap real quick, Neural
24:17
Link is implanted with precision surgical robot.
24:21
It's physically invisible and one week later users are able to
24:25
see their thoughts, transform into actions and to share more
24:28
about what that experience is like.
24:31
I'd like to welcome Sahej to the station.
24:47
What's up guys? My name is Sahej, I'm from the
24:50
Brain Computer Interface team here at Neurolink, and I'm going
24:53
to be talking about two things today.
24:55
The first thing is, what exactly is the neural link device
24:58
capable of doing right now? And the second one is how does
25:01
that actually impact the day-to-day lives of our users?
25:05
Very Simply put, what the neural link device does right now is it
25:10
allows you to control devices simply just by thinking.
25:14
Now to put that a bit more concretely, I'm about to play a
25:18
video of our first user. His name is Noland, if you
25:20
remember from DJ section. And what Noland is doing is he's
25:25
looking at a normal off the shelf MacBook Pro.
25:28
And with his Neurolink devices, you're going to see he's going
25:31
to be able to control the cursor simply with his mind.
25:34
No eye tracking, no other sensors.
25:37
And what's special about this particular moment is this is the
25:40
first time someone is using a Neurolink device to fully
25:43
control their cursor. This is not your ordinary brain
25:50
controlled cursor. This is actually a
25:53
record-breaking control literally on day one, beating
25:56
decades of brain computer research.
25:59
And I'm about to show you the clip on day one, Nolan breaking
26:03
the BCI world record. Oh well done man, He's a new
26:20
world record holder. I.
26:31
Thought it was higher. I thought I would have to get to
26:34
5 or something. Oh my gosh, that's crazy.
26:39
It's pretty cool, yeah. Another really fun thing you
26:47
could do with the newer Link device, outside of controlling a
26:50
computer cursor, is you can actually plug it in through USB
26:53
through a lot of different devices.
26:55
And here we actually have Nolan playing Mario Kart.
26:58
Now, what's special about this particular clip is Nolan is not
27:02
the only Cyborg playing Mario Kart in this clip.
27:05
We actually have a whole community of users, as mentioned
27:08
earlier. And this is literally five of
27:10
our first users of Neuralink playing Mario Kart together over
27:14
call now. Yeah, Mario Kart is, it's cool.
27:26
You know, you're using one joystick and then you're
27:28
clicking like a couple buttons to throw items.
27:31
What would be even cooler is what if you could control 2
27:34
joysticks at once simultaneously with your mind?
27:37
What I'm about to show you, and I think this is for the first
27:40
time someone playing a first person shooter game with a brain
27:43
computer interface. This is Alex and RJ playing Call
27:47
of Duty controlling 1 joystick to move and then the other
27:51
joystick to like think, point your gun and then shooting
27:53
people as a button. Here's the larger Here's Alex
27:57
shooting another person. 'S bikes.
28:00
Oh dear God, I know I do and I. Want him to freaking.
28:05
Choose long. When I do, I know he's shot me
28:10
in the face. Now that we have a bit of a
28:14
sense of what the BCI can do, a very important question to
28:18
answer is how does this impact the day-to-day lives of the
28:22
people that use it every day? So I'm about to show you a clip
28:26
going back to Nolan for a second where he talks.
28:29
We simply just asked him randomly during a day how he
28:33
enjoys using the BCI couple months ago and this is his
28:36
candid reaction. I work basically all day from
28:41
when I wake up. I'm trying to wake up at like
28:45
6:00 or 7:00 AM and I'll do work until session.
28:49
I'll do session and then I'll work until, you know, 11:12 PM
28:57
or 12:00 AM I'm, I'm doing like I'm learning my languages, I'm
29:07
learning my math. I'm like relearning all of my
29:09
math. I am writing, I am doing the
29:14
class that I signed up for and I just I wanted to point out that
29:20
like this is not something I would be able to do out like
29:25
without the neuro link. Next I want to talk a bit about
29:28
Brad. You guys may already know him as
29:30
the ALS Cyborg and Brad also has ALS and what separates him from
29:34
other users is he's actually non verbal so he can't speak.
29:38
Why this is pretty relevant is he relies at least before the
29:42
neuro link on an eye gaze machine to communicate and a lot
29:45
of eye gaze machines you can't use outdoors.
29:47
You really need like a dark room.
29:49
So what this means is for the last six years since Brad's been
29:52
diagnosed with ALS, he's really unable to leave his house.
29:57
Now with the Neuralink device, we're going to show you a clip
29:59
of him with his kids at the park, shot by Ashley Vance and
30:03
the team Guys ready. No, absolutely.
30:08
I am absolutely doing more with Neuralink than I was doing with
30:11
eye gaze. I've been a Batman for a long
30:17
time, but I go outside now. Going outside has been a huge
30:21
blessing for me and I can control the computer with
30:24
telepathy. Dad's watching.
30:27
OK, He's. Watching on the camera, did he?
30:30
Lose. One of the arms the last user I
30:32
want to talk about is Alex. You've seen some clips of him
30:35
earlier. What's special about Alex to me
30:38
is he's a fellow left-handed guy who writes in cursive all the
30:41
time. And what he mentioned is since a
30:43
spinal cord injury from like, 3-4 years ago, he's been able
30:47
unable to just, like, draw or write.
30:51
And he always brags about how good his handwriting was.
30:53
So we actually got to put in a test.
30:55
We gave him a robotic arm. And I think this is the first
30:57
time you tried using the robotic arm to write anything.
31:01
And this is a spotted version of writing at the convoy trial and
31:04
drawing something now. Yeah, controlling A robotic arm
31:15
is cool, but this one has a clamp.
31:18
And what would be cooler is if you could decode the actual
31:21
fingers, the actual wrist, all the muscles of the hand in real
31:24
time. Just in the past couple weeks we
31:28
were able to do that with Alex and you're about to see him and
31:31
his uncle playing a game. Rock, paper, scissors shoot.
31:39
Rock, paper, scissors shoot. Rock, paper, scissors shoot
31:46
rock. Paper.
31:48
Scissors. Shoot.
32:09
Cool. Controlling.
32:11
Yeah, that's pretty dope. I don't know.
32:18
And controlling A robotic hand on screen is obviously not super
32:22
helpful for most people. Fortunately, we have connections
32:26
with Tesla, who have the optimist hand, and we're
32:30
actually actively working on giving Alex an optimist hand so
32:36
that you could actually control it in his real life.
32:39
And here's actual replay of the end of that video using Alex's
32:43
neural signals on an Optus hand. Sean, if you want to play that,
32:53
Yeah. Actually, let me maybe add a few
32:59
things to that, which is so as we advance the neural link
33:05
devices, you should be able to actually have full body control
33:11
and sensors from an optimist robot.
33:14
So you could basically inhabit an optimist robot.
33:17
So not just the hand the whole, the whole thing.
33:21
So you could like basically mentally remote into an optimist
33:27
robot and and be kind of cool. The future's going to be weird,
33:33
but but but pretty cool. And then now I was.
33:41
Another thing that could be done also is like for people that
33:43
have say, lost a limb, lost an arm or a leg or something like
33:47
that, then we think in the future we'll be able to attach
33:53
an optimist's arm or legs. And so you kind of like, I
33:57
remember that scene from Star Wars where Luke Skywalker gets
34:02
his hand, you know, chopped over the lightsaber and he gets kind
34:05
of a robot hand. And I think that's the kind of
34:08
thing that we'll be able to do in the future working with the
34:11
newer Lincoln Tesla. So, so it goes far beyond just
34:16
operating a robot hand, but replacing limbs and, and having
34:19
kind of a whole body robot experience.
34:21
And then I think another thing that will be possible, I think
34:25
it's very likely in the future is to be able to bridge the, the
34:30
where the damaged neurons are. So you can take the signal from
34:33
the brain and, and transmit that signal past where the neurons
34:38
are damaged or strained to the rest of the body.
34:40
So you could reanimate the body so that if you have a neural
34:44
link implant in the brain and then one in the spinal cord,
34:48
then you you can actually bridge the signals and you could walk
34:52
again and have full body functionality.
34:55
Obviously that's what people would prefer.
34:57
To be clear, we realized that that would be the preferred
35:00
outcome and and so that even if you have a broken neck or you
35:06
could. So we believe I'm, I'm actually
35:09
at this point I'd say fairly confident that at some point in
35:11
the future we'll be able to restore full body functionality.
35:23
Yes. So hello, hello everyone.
35:24
My name is Nir and I'm leading the BCR application group.
35:28
And I think the videos just that Sahel just shared with you, I
35:32
probably watch them maybe thousands of time, but still I
35:36
get the goosebump every time I watch them.
35:38
And I think this is one of the cool perks here at New Orleans
35:41
when you get a job is that you might get goosebump every week
35:44
or maybe every few days in good weeks.
35:46
And, and this is really fun as an engineer, it's really cool
35:53
because you can build a new feature.
35:54
You can build a new machine learning model and use software
35:57
feature and test it on the same day with the participant and get
36:00
feedback. And you already saw with our
36:04
first device Telepathy that we can address a very diverse needs
36:08
of the different users that we have for moving a cursor to
36:13
playing games, to move a robotic arm with multiple fingers.
36:17
And we could not have done it without the neural link device.
36:20
The neural link device gives us something that no other device
36:23
can give us, which is in a single neuron recording from
36:27
thousands of channels simultaneously.
36:29
The telepathy products is basically recording the neural
36:32
activity from the small area in the motor cortex that involve an
36:36
execution of hand and arm movements.
36:39
But if we go only about two or three inches below, there's
36:42
another brain area that's involved in execution of speech.
36:46
And with the same device, with the same machine learning model
36:49
architecture, the same software pipeline, the same surgical
36:52
robot, we can have a new application and we can do it
36:55
very quickly. It's really interesting that if
36:57
we can decode someone intention to speak silently and non vocal
37:03
communication, we can use that to revolutionize the way we
37:06
interact with computers, with technology and with information.
37:10
Instead of typing with your finger or like moving the mouse
37:14
or talking to your phone, you'll be able to interact with
37:17
computer with the speed of thoughts.
37:19
It will make this interaction much more, much faster and much
37:22
more intuitive. The computers will understand
37:25
what you want to do. And we can also expand that to
37:28
AI. We can now build an interface
37:31
with AI that you will be able to achieve information, will be
37:35
able to store our thoughts anywhere, anytime, privately and
37:40
silently. Again, because we build a
37:42
fundamental technology, a platform, and we do everything
37:46
in house. We own the entire stack from
37:49
neurons to pixels on the user's computer.
37:52
Now I'll pass pass it to RUSE to talk about UI for Visa.
38:03
Thank you, NIA. Each spike that our implant
38:06
detects goes on a fairly remarkable journey to ultimately
38:09
form a pixel on a participant's display.
38:12
And that experience starts with, of course, unboxing, the very
38:18
first time that a participant pairs to and meets their
38:21
implant, this invisible part of their body, and sees their own
38:26
spikes materialize across the display.
38:30
From there, they'll go into body mapping and actually imagine
38:34
moving their arm again and get a feel for what feels natural to
38:38
them and what doesn't. And they'll take that into
38:41
calibration, using one of those motions to actually move a
38:47
cursor again, iteratively refining their control as they
38:52
go throughout this process, until finally they're teleported
38:57
back to their desktop and can experience the magic of neural
39:02
control for the very first time. And our control interfaces is
39:07
where the OS integration that we do really shines, letting us
39:12
adapt both control and feedback for every interaction.
39:17
So for familiar interactions like scrolling, we can surface
39:21
an indicator over the scrollable parts of the display, add a
39:25
touch of gravity to automatically pop a
39:27
participant's cursor onto that indicator as they approach, show
39:32
the actual velocities that we decode inside of it, and add a
39:36
bit of momentum to those velocities to carry them forward
39:39
as they glide across the page. There are also unique
39:43
interactions that we need to solve for in this space.
39:46
For example, when a participant is watching a movie or just
39:49
talking to somebody next to them, the brain is very active
39:53
still, and that activity can actually induce motion in the
39:56
cursor, distracting them from that moment.
39:59
So when a participant wants to just get their cursor out of the
40:01
way, they can push it into the edge of the display to park it
40:05
there. And of course we add gravity to
40:07
sort of hold it still, but they can push it out with either just
40:12
a firm push or in this case, a gesture.
40:16
And of course, it's when it goes without saying that all of these
40:19
control interfaces are designed hand in hand with our
40:21
participants. So huge shout out to both Noland
40:24
and Brad for helping us design these two.
40:27
And those control interfaces, of course, extend typing.
40:30
We have a great software keyboard that does everything
40:33
you'd expect it to, popping up when a participant clicks on a
40:35
text field, giving them feedback about the click on the surface
40:39
of the key, and supporting both dictation and swipe.
40:49
Hi everyone, I'm Harrison and ML Engineer here at Neural Link.
40:52
And I must say, being an ML engineer at Neural Link is a bit
40:55
like being a kid in a candy store.
40:59
When you think of the inputs to most ML systems out there, you
41:02
might think of pixels of tokens or of a user's Netflix watch
41:06
history. The input to our systems is a
41:09
little different. It is pure raw brain power.
41:13
And when we think about the ML systems we can build here at
41:15
Neuralink, really we're limited by our imagination and our
41:18
creativity. There's no reason our ML systems
41:21
can't do anything that the human brain can do, such as
41:25
controlling a phone, typing, or even gaming.
41:29
Right here to my left is actual footage of Alex, one of our
41:32
participants, playing a first person shooter against RJ,
41:36
another one of our participants. Now, for those unfamiliar with
41:39
first person shooters, this is not a trivial feat.
41:42
It requires 2 fully independent joysticks or 4 continuous
41:46
degrees of control, as well as multiple reliable buttons.
41:51
Now contrary to popular belief, the neural link does not simply
41:55
read people's minds, it's simply reading neuronal activations
41:59
corresponding to motor intent. So one of the fun challenges
42:02
with this project was figuring out which motions were going to
42:05
be mapped to the joystick. We started with the typical left
42:09
thumb and right thumb, but quickly found that the dominant
42:11
hand overshadowed the non dominant hand.
42:14
My personal favorite is we had one of our participants imagine
42:16
walking for the left joystick and aiming for the right
42:19
joystick. So in game, they were simply
42:21
doing naturalistic motions like you might do in virtual reality
42:25
in Ready Player 1, and that was really cool to watch.
42:28
What we ended up on was the thumb for the left joystick and
42:32
the wrist for the right joystick.
42:34
And I challenge the audience to try to replicate their emotions.
42:37
I'm really in all of them being able to pull this off.
42:40
I want to talk a bit about the progress to our cursor
42:42
calibration experience. To my left, here you can see RJ
42:45
completing his first ever cursor calibration with a redesigned
42:49
Open the Flow, where he first gather information about his
42:51
intent and how to map the neural activity to the first time he
42:54
controls a cursor, to the final product where he has smooth and
42:57
fluid control of his computer. And most remarkably, this
43:02
experience took only 15 minutes from start to finish, 15 minutes
43:06
from not 15 minutes from no control to fluid computer use.
43:17
Contrast that to a year and a half ago with P1, where that was
43:20
multiple hours to get to the same level of control and
43:23
several engineers standing around a table pulling their
43:25
hair out. There was virtually no need for
43:27
neural link engineers to even be at this session.
43:30
This was basically an out-of-the-box experience for
43:32
our participants. And even more remarkably, we're
43:38
continuing to smash day one records, with RJ being able to
43:41
achieve 7 BPS on his very first day with a neural link.
43:49
Now such an effective and efficient calibration process is
43:52
only made possible by high fidelity estimations of a user
43:56
intention or labels. And to briefly illustrate just
44:00
how challenging of a problem that is, this is an animation of
44:03
myself trying to draw circles on my desktop with a mouse.
44:07
Now the task was simple, draw uniform circles at a constant
44:10
speed, repeatedly. And as you can see by that
44:13
animation, I am horrible at that.
44:16
Even though my intent was pretty obvious, unambiguous, the
44:19
execution was really poor. There is a ton of variation in
44:23
both speed and the shape itself. To visualize this a little
44:27
differently, each row here is one of those circles unwound in
44:31
time with synchronized starts, and you can just see how much
44:34
variation there is in the timing of each circle as well as I'm
44:36
doing at any given point in time.
44:40
Orthogonal to the labeling problem is neural non
44:43
stationarity, or the tendency of neural signals to drift over
44:46
time. And I think that's honestly a
44:48
beautiful thing, right? If you if your neural signals
44:51
didn't drift, you couldn't grow. When you wake up the next day,
44:54
you're not the same person you were the day before.
44:56
You've learned, you've grown, you've changed, and so too must
44:59
your neural data change. This animation.
45:02
Here is a simple illustration of the learned representation by
45:05
the decoder and how it drifts the further away we get from the
45:08
day it was trained on. This is one of the key
45:10
challenges we need to solve here at Neurolink to unlock fluid and
45:13
product level experience for our users.
45:23
Hey everyone, Hey everyone, my name is Joey.
45:31
Blind Sight is our project to build a visual prosthesis to
45:35
help the blind see again. Users would wear a pair of
45:39
glasses with an embedded camera and receive an implant in their
45:45
visual cortex. Scenes from the environment are
45:49
recorded by the camera and processed in the patterns of
45:52
stimulation delivered to the brain, causing visual perception
45:58
and restoring functionality. Now blind sight will be enabled
46:02
by placing our implant into visual cortex.
46:04
This is a new brain area for us, and this brings new
46:07
opportunities and challenges. So the surface of the brain for
46:12
visual cortex represents just a few degrees of angle in the
46:16
center of the visual field. Larger fields of view are
46:20
represented deep within the cortical folds of the calcarine
46:23
fissure. Our threads are able to access
46:27
these deeper structures, providing the possibility of
46:30
restoring vision over a functional, useful visual field.
46:34
So the N1 implant has had experimental stimulation
46:37
capabilities for quite some time, but our new S2 chip is
46:41
designed from the ground up for stimulation.
46:44
It provides over 1600 channels of electrical stimulation, high
46:49
dynamic range recording capabilities and a wide range of
46:53
micro stimulation currents and voltages.
46:56
We can achieve these capabilities because we are
46:59
vertically integrated and we designed this custom ASIC in
47:02
house. Similarly, we design and
47:05
fabricate our electrode threads in house, and here you can see
47:08
one of our standard threads designed for recording in an
47:12
electron micrograph for blind sight.
47:15
Our requirements are a little different, and our vertical
47:18
integration allows us to rapidly iterate on the design and
47:21
manufacturing of these threads for this new purpose.
47:25
So here I'm using Red Arrows to highlight the electrode
47:28
contacts, which are optimized for stimulation.
47:32
And as you can see, they're a little bit larger, which results
47:35
in a lower electrical impedance for safe and effective charge
47:38
delivery, which is important for blind sight.
47:43
Now, how can we calibrate our implant for blind sight?
47:46
So here's one way we stimulate on the array, picking say three
47:50
different channels. The user perceives something,
47:54
say 3 spots of light somewhere in their visual field and points
47:57
at them. We track their arm and eye
48:00
movements and repeat this process for each of the channels
48:03
on the array. And here's what a simulated
48:08
example of a blind sight vision could look like after
48:10
calibration. Now I showed you how for blind
48:24
sight, we need to insert threads deeper into the brain than we
48:26
have previously, and doing this requires state-of-the-art
48:30
medical imaging. So we worked with Siemens to get
48:34
some of the best scanners on Earth.
48:36
We built out our imaging core from scratch in the past year.
48:40
Actually, it was faster than that.
48:41
It was about four months from dirt to done.
48:45
Since bringing the scanners online, we've scanned over 50
48:47
internal participants, building out a database of human
48:50
structural and functional anatomy.
48:52
What can we do with the imaging information from these scanners
48:56
so medical imaging can be used for surgical placement?
48:59
It lets us parcel it out brain regions by their function and we
49:02
use our imaging capabilities to refine the placement for
49:05
telepathy. It also gives us the capability
49:08
to target new brain regions for future products such as blind
49:11
sight or speech prosthesis. And we're working towards more
49:14
capabilities. So one click, automated planning
49:17
of surgery from functional images to robot insertion
49:19
targets. Here you can see a screen
49:21
capture from one of our in house tooling to do end to end
49:24
surgical planning. You can see a region of motor
49:27
cortex known as hand knob and the thread trajectory plans that
49:31
will be sent directly to the robot.
49:34
This is a really incredible degree of automation that's only
49:37
possible because we're controlling the system from one
49:39
end to the other. My name is John and I lead the
49:46
robot mechanical team. This is our current R1 robot.
49:50
It was used to implant the first 7 participants.
49:53
This robot works really well, but it has a few flaws.
49:56
One of which is the cycle time is rather slow.
49:59
So to insert each thread it takes in a best case scenario 17
50:03
seconds. And many cases external
50:05
disturbances cause us to have to retry to reinsert, grasp that
50:09
thread and then reinsert it. To scale our number of neuron or
50:13
neurons access through higher channel count, increased numbers
50:16
of threads, we need to have a much faster cycle time.
50:19
So let me introduce our next generation robot, which is right
50:23
here. Through rethinking the way that
50:34
we hold the implant in front of the robot, by holding it
50:36
directly in front on on the robot head, we will achieve an
50:39
11 times cycle time improvement. So each thread takes 1 1/2
50:43
seconds. We also scale up a lot of
50:46
surgery. Workflow process improvements
50:49
through deleting the separate operator station and implant
50:52
stand. Now the outside of the robot
50:55
looks pretty similar between the two, but it's what's inside that
50:58
really counts. Each system has been redesigned
51:01
from the the ground up with the focus on reliability,
51:04
manufacturability, serviceability and using a lot
51:07
of our vertical integration techniques.
51:09
It's enabled us to have a lot more control of the system end
51:11
to end. Now that fast cycle time doesn't
51:15
mean much if it's not compatible with a significant portion of
51:18
the human population. Prior to each surgery we scan
51:22
our participants anatomy and ensure that they will be
51:24
compatible with the robot and vice versa.
51:27
Unfortunately, the robot isn't compatible with everyone so we
51:30
had to extend the reach of the needle in the next generation
51:33
robot and now we're compatible with more than 99% of the human
51:36
population. We've also increased the depth
51:38
of the needle can insert threads.
51:40
Now we can reach more than 50mm from the surface of the brain,
51:43
accessing and enabling new indications.
51:46
We have to produce a ton of custom sterile components for
51:49
each surgery. We actually supply more than 20
51:51
of these parts. Many of these parts are made
51:54
through traditional CNC manufacturing capabilities,
51:56
which we do just on the other side of this wall actually, and
51:59
some custom developed processes like this femtosecond laser
52:02
milling used to manufacture the tip of the needle.
52:05
Now these processes take quite a bit of time, effort and cost.
52:09
So let's take a look at how we're going to reduce costs and
52:12
time for one of the components. So the current needle cartridge
52:16
has a total cycle time of about 24 hours and the machine
52:20
components cost about $350.00. The final assembly is performed
52:24
by a set of like highly skilled technicians.
52:26
They have to glue 150 Micron diameter Canyon onto this wire
52:30
EDM machined stainless steel base plate.
52:32
They have to Electro Polish a 40 Micron wire into a sharp taper
52:36
and then they have to thread that 40 Micron wire into the a
52:39
60 Micron hole in the Canyon. This is done manually and then
52:43
they finally have to laser Weld all the components together.
52:48
Next generation needle cartridge takes only 30 minutes of cycle
52:51
time and $15 in component. We were able to delete the wire
52:55
EDM machined base plate and the Kenya gluing step by switching
52:58
to an insert molded component. So we get a box of these base
53:01
plates with the Kenya was already installed for like 1000
53:04
of them for like a couple 5-10 dollars a piece.
53:07
We also deleted the Electro polishing step with the revised
53:10
needle tip geometry, which is also compatible with inserting
53:13
the threads through the dura. We have a few revised
53:17
manufacturing techniques to delete the manual threading
53:19
through a basically a funnel. Rather simple, but it has been a
53:22
big impact. And then we're able to delete
53:24
the laser alding through using crimping.
53:32
Hi. Hi.
53:38
I'm Julian. I'm one of the leads on the
53:40
implant team. So the way humans communicate
53:44
today, if they want to output information, is by using their
53:47
hands and their voice, as I'm doing right now.
53:50
And if you want to receive information, you use your ears
53:53
and your eyes. And of course, that's how you're
53:55
receiving this very talk. But we've built this implant and
54:00
this implant is very special because it is the first time
54:04
that we're able to add a completely new mode of data
54:07
transfer into and out of the brain.
54:11
If you look at this device in a nutshell, it's really just
54:14
sampling voltages in the brain and sending them over radio.
54:17
But if you zoom out and look at the system from end to end, what
54:21
you actually see is that we're connecting your brain or
54:25
biological neural net to a machine learning model or a
54:29
silicon neural net on the right hand side.
54:32
And I actually think this is really elegant because the
54:36
machine learning model on the right hand side is in fact
54:39
inspired by neurons on the left hand side.
54:42
And so in some sense, we're really extending the fundamental
54:46
substrate of the brain. For the first time, we're able
54:48
to do this in a mass market product that's a very, very
54:51
special piece of hardware. So these are some of the first
55:02
implants that we ever built. There are electrodes that were
55:07
made with our in house lithography tools.
55:10
We have custom ASICS that we also designed in house.
55:13
And this was really a platform for us to develop the technology
55:16
that allows us to sense micro level volts in the brain across
55:20
thousands of channels simultaneously.
55:22
We learnt a lot from this. But as you'll notice in the
55:25
right to images, there are USBC connectors on these devices.
55:30
These were not really the most implantable implants.
55:34
This next set of images are the wireless implants, and it was a
55:38
complete evolution that we went through to add the battery, the
55:42
antenna, the radio, and to make it actually fully implantable.
55:46
Once it's implanted, it's completely invisible.
55:49
It's very compact, it's modular, and it's a general platform that
55:52
you can use in many places in the brain.
55:55
Going from that top row to the bottom row is very challenging.
56:00
The implant you see on the bottom right here is in fact the
56:03
device that we have working in seven participants today, and
56:06
it's augmenting their brain every day and restoring their
56:10
autonomy. But getting to that point
56:12
involved a huge number of formidable engineering
56:15
challenges. We first had to make a hermetic
56:17
enclosure, passing 1000 separate conductors through the enclosure
56:21
of the device. We had to figure out how to make
56:23
charging seamless and work with very tight thermal constraints
56:26
in a very, very small area. And then we also had to scale up
56:30
our testing infrastructure so that we could support large
56:32
scale manufacturing and very safe devices and have confidence
56:35
in our iteration cycle. So what's next?
56:39
We're going to be increasing our manufacturing so that we don't
56:43
just produce, you know, a certain like a small number of
56:46
implants per year, but thousands and then eventually millions of
56:48
implants per year. We're also going to be
56:50
increasing channel count. More channels means more neurons
56:55
are sensed, which means more capabilities.
56:57
In some sense, We often think a lot about the the Moore's Law of
57:01
neurons that we're interacting with.
57:03
And in the same way that Moore's Law propelled forward many
57:06
subsequent revolutions in computing, we think that sensing
57:10
more and more neurons will also completely redefine how we
57:13
interact with computers and reality at large.
57:17
I want to leave you with one final thought.
57:20
When I was a child I used a 56 KB modem to access the Internet.
57:26
If you remember what it's like, you would go to a website.
57:29
You're lucky. You're lucky.
57:32
Bastard. Yeah, When I was a child we had
57:35
acoustic couplers. Oh yeah, OK.
57:37
So just beep. Just beep at each other.
57:39
Yeah, the the first modem was the acoustic coupler, incredible
57:43
device honestly. But then if you, I guess if
57:47
you're my age, you started with 56 K bit modem and you, you
57:55
would go to a website and and like there would be an image and
57:58
it would, it would scroll like slowly it was loading pixel by
58:02
pixel on the screen. So that that's what it's like to
58:05
be bandwidth limited. Now imagine using the current
58:08
Internet with that same modem. It it's like it's inconceivable.
58:12
It would be impossible to do so. What broadband Internet did to
58:17
the 56 KB modem is what this hardware is going to do to the
58:21
brain. We are trying to drastically
58:23
expand the amount of bandwidth that you have access to, to have
58:27
a much richer experience and superhuman capabilities.
58:37
So I guess just to kind of close out and to recap today Neurolink
58:42
is working reliably and has already changed the lives of
58:47
seven participants and making a real impact.
58:51
And our next milestone is to go to market and enable scaling of
58:55
this technology to thousands of people and as well as expand
58:59
functionalities beyond just the movement to enable robotic,
59:03
sophisticated robotic arm control, speech, vision, give
59:07
sight back and even getting to the speed of thought.
59:11
I, I hope you got a good sort of sample of our technology stack
59:15
and the challenges that we have and I'd like to hand over the
59:19
mic to Elon for any closing remarks.
59:24
Well, we're trying to give you a sense of the the depth of talent
59:30
at Neurolink. There's a lot of really smart
59:32
people working on a lot of important problems.
59:36
This is one of the most difficult things to to actually
59:40
succeed in creating and have it work and work at scale and be
59:43
reliable and available for millions of people at an
59:46
affordable price. So super hard problem and would
59:53
like to have you come join and help us solve it.
59:56
Thank you. Hey, thank you so much for
1:00:15
listening today. I really do appreciate your
1:00:17
support. If you could take a second and
1:00:19
hit this subscribe or the follow button on whatever podcast
1:00:22
platform that you're listening on right now, I greatly
1:00:25
appreciate it. It helps out the show
1:00:27
tremendously and you'll never miss an episode.
1:00:29
And each episode is about 10 minutes or less to get you
1:00:33
caught up quickly. And please, if you want to
1:00:36
support the show even more, go to patreon.com/stagezero.
1:00:41
And please take care of yourselves and each other, and
1:00:44
I'll see you tomorrow.