Neuralink 2025 夏季进展

0:03Hello everybody, my name is Alex Todling.

0:06I'm the second participant in the Neuralink Study,

0:09and I'm here to count us down to the demo.

0:13In five, three, two, one.

0:31Hi everyone.

0:35Welcome to the Neuralink presentation.

0:38This is an update for the Progress from the Neuralink Team.

0:42It's been an incredible amount of progress.

0:44This is, we're gonna start off high level.

0:49that generally describing what Neuralink's doing,

0:52and then we're going to have a very deep technical dive.

0:55So you can actually get an understanding of what exactly

0:58we're doing at a granular level, and what

1:01we can do to enhance human capabilities,

1:04and ultimately build a great future for humanity.

1:09So that's a neuron-spiring.

1:11It's funny to think that me talking right now

1:14is a bunch of neurons firing that then result in speech

1:20that you hear, of course, neurons

1:23to fire in your brain.

1:28Part of this presentation is about demystifying the brain.

1:33It is a remarkable organ.

1:35I mean, we are the brain, basically.

1:36When you say you, that really is you're the brain.

1:41like you can get a heart transplant,

1:44you can get a kidney transplant,

1:46but I don't know anyone who's gotten a brain transplant.

1:50So you are your brain, and your experiences

1:56are these neurons firing with trillions of synapses

2:04that somehow lead to conscious comprehension of the world.

2:08This is something that we have only begun to understand.

2:12we're really just barely at the beginning of understanding

2:15of what is the nature of consciousness.

2:18And I've thought a lot about what is consciousness?

2:20Where does consciousness arise?

2:27Because if you start at the beginning of the universe,

2:30assuming physics is true,

2:31the standard model of physics is true,

2:34then you have this big bang,

2:40the matter condensing into stars,

2:42those stars exploding.

2:43A lot of the atoms that are in your body right now were ones at the center of stars.

2:50Those stars exploded, re-condensed, fast forward, 13.8 billion years, and here we are.

2:59And somewhere along that very long journey to us, at least, consciousness arose.

3:08Or the molecules started talking to each other.

3:14And it begs the question of what is consciousness?

3:23Is everything conscious?

3:24Maybe.

3:25It's hard to say where along that line there's no sort of discrete point where consciousness

3:32didn't exist and then suddenly does exist.

3:36It seems to be maybe you have a condensation of matter that has a density of, like we don't

3:44know what, the real answer is we don't know what.

3:46consciousness is. But with the Neuralink and the progress of the company's making will

3:54begin to understand a lot more about consciousness. And what does it mean to be? Along the way

4:06we're going to solve a lot of brain issues where the brains get injured or damaged in

4:16in some way or didn't develop in quite the right way.

4:20But there's a lot of brain and spine injuries

4:25that build so along the way.

4:26And I do want to emphasize that this is all

4:29going to happen quite slowly.

4:31Meaning you'll see it coming.

4:32Sometimes people think that suddenly there

4:35will be vast numbers of neural lengths all over the place.

4:40This is not going to be sudden.

4:41You'll be able to watch it happen over the course

4:45of several years.

4:47And we go through exhaustive regulatory approvals.

4:52So this is not something that we're just doing by ourselves

4:56without a government oversight.

4:58We work closely with the regulators every step of the way.

5:02We're very cautious with the neurolinks in humans.

5:08That's the reason we're not moving faster than we are,

5:10is because we're taking great care

5:11with each individual to make sure we never miss.

5:16And so far we haven't,

5:17and I hope that continues into the future.

5:19Every single one of our implants in humans is working

5:23and working quite well.

5:24And you'll get to hear from some of the people

5:26that have received the implants and cared in their words.

5:31So what we're creating here with a Neuralink device

5:35is a generalized input output technology

5:40for the brain. So it's how do you get information into or out of the brain and

5:49do so in a way that does not damage the brain or you know cause any negative

5:55side effects. It's a very hard problem and generally the reactions I've seen to

6:02this range from it's impossible to it's already been done before. Those people

6:08should meet, actually.

6:12The reality is that there actually

6:14have been limited brain to computer interfaces

6:19for several decades on a very basic basis.

6:24Just what we're doing with Neuralink

6:27is dramatically increasing the bandwidth by Mayors

6:33and magnitude.

6:33So a human bandwidth output is less than one bit per second

6:40over the course of a day.

6:41So there's 86,400 seconds in a day.

6:44It's very rare for a person to do more than 86,400 bits

6:49of output per day.

6:50You'd have to be really talking a lot or typing all day,

6:53and you might exceed that.

6:57So what we're talking about here is going from maybe one

7:00bit per second to ultimately megabits,

7:03and then gigabits per second.

7:07And the ability to do conceptual telepathy.

7:14Now, the input to the brain is much higher,

7:17especially because of vision.

7:19Depending upon how you count it, it

7:22might be on the order of a megabit or in the megabit range

7:27for input primarily due to sight.

7:29But even for input, we think that can be dramatically increased to the gigabit plus

7:38level.

7:42And a lot of the thinking that we do is we take a concept in our mind and we compress

7:49that into a small number of symbols.

7:52So when you're trying to communicate with somebody else, you're actually trying to model

7:56their mind state and then take quite a complex idea that you have, maybe even a complex image

8:05or scene or kind of mental video and try to compress that into a few words or a few key

8:11strokes. And it's necessarily going to be very lossy. Your ability to communicate is

8:16very limited by how fast you can talk and how fast you can type. And what we're talking

8:22about is unlocking that potential to enable you to communicate, like I said, thousands,

8:28perhaps millions of times faster than is currently possible.

8:32This is an incredibly profound breakthrough.

8:36This would be a fundamental change to what it means to be a human.

8:43So we're starting off with reducing human self-rank or addressing issues that people

8:51have, if they've been in an accident or they have some neural disease that's degenerative

8:58so they're losing capability to move their body or some kind of injury essentially.

9:05So enabling our first product is called telepathy and that enables someone who has lost the

9:12ability to command their body to be able to communicate with the computer and move the

9:17mouse and actually operate a computer with roughly the same dexterity, ultimately much

9:22more dexterity than a human with working hands.

9:29Then our next product is blind sight, which will enable those who have total loss of vision,

9:35including if they have lost their eyes or the optic nerve, or maybe have never seen,

9:39were even blind from both to be able to see again.

9:43initially low resolution but ultimately very high resolution and and then in multiple wavelengths.

9:49So you could be like Jordy LeFarge in Star Trek and you can see in radar, you can see an infrared,

9:53ultraviolet, superhuman capabilities, cybernetic enhancement essentially.

10:01And then along the way this should help us understand a lot more about consciousness. What

10:08What does it mean to be a conscious creature?

10:13We'll understand vastly more about the nature of consciousness

10:16as a result of this.

10:17And then ultimately, I think this helps

10:19mitigate the civilizational risk of artificial intelligence.

10:27We are actually already sort of have three layers of thinking.

10:33There's the limbic system, which is your instincts,

10:36the cortical system, which is your higher level of planning

10:41and thinking, and then the tertiary layer, which

10:43is the computers and machines that you interact with,

10:47like your phone, all the applications you use.

10:51So people actually are already a cyborg.

10:54You can maybe have an intuitive sense for this

10:57by how much you miss your phone if you leave it behind.

11:03Leaving your phone behind is almost like missing

11:05syndrome. Your phone is somewhat of an extension of yourself as is your computer. So you already

11:14have this digital tertiary layer, but the bandwidth between your cortex and your digital tertiary

11:20layer is limited by speech and by how fast you can move your fingers and how fast you can

11:26consume information visually.

11:29So, but I think it's actually very important for us to address that input output bandwidth

11:37constraint in order for the collective will of humanity to match the will of artificial

11:43intelligence.

11:48That's my intuition at least.

11:50So let's see.

11:53And what this presentation is mostly about is attracting smart humans to come and work

12:04with us on this problem.

12:06So this is not a presentation to raise money or anything like that.

12:11We're actually very well funded.

12:13We have a lot of great investors.

12:17of the smartest people in the world are invested in Neuralink.

12:21But we need smart humans to come here and help solve this problem.

12:25So with that, let's proceed.

12:40Hey, everyone.

12:41My name is DJ.

12:43I'm a co-founder and president of Neuralink.

12:46And as Elon mentioned, well, actually,

12:49we're standing in the middle of our robot space.

12:52We have a stage set up, but this is actually

12:55where some of the next generation, most advanced

12:58the vertical robots are being built.

12:59So welcome to our space.

13:09It's important to highlight that this technology is not being

13:11built in the dark.

13:13This is not a secret lab where we're not sharing any of the

13:16progress.

13:16In fact, we're actually sharing the progress very openly, as

13:21well as also telling you exactly what we're going to be doing.

13:24And we're hoping to progress on that as diligently and as

13:29safely and as carefully as possible.

13:31So to start off, two years ago, when we did our previous fundraising round,

13:37we outlined this path and timeline to First Human.

13:41And we currently have a clinical trials in the US for a product that we call telepathy,

13:46which allows users to control phone or computer purely with their thoughts.

13:51And you're going to see how we do this and what the impact that this has had.

13:56And not only have we launched this clinical trial, but as of today,

14:00we have not just one, but seven participants.

14:04And we have an approval.

14:12And we also have an approval to launch the trial

14:15in Canada, UK, and the UAE.

14:25So I guess before we dive into what this technology is

14:29and what we built, but I wanted to quickly share a video

14:32with you guys of when our first five participants

14:36met each other for the first time.

14:37So here you go.

14:39All right, we have everyone together.

14:41What's up, guys?

14:42Thanks everybody joining. Definitely want to use all of you.

14:46Yeah, I'm Nolan AKA P1.

14:48My name is Alex. I am the second participant in the Neural

14:52Link Study. I am Brad Smith, the ALS cyborg P3.

14:58My name is Mike P4.

15:02I have ALS like fried rice.

15:07Yeah, I'm RJ, I'm B5.

15:09And I guess I'm kind of the newest one of the team here.

15:14So yeah, I appreciate it.

15:16No trailblazer.

15:18You know, somebody's got to get first, man.

15:19That was you appreciate that.

15:21What's been your favorite thing you've been able to do

15:23with the Neuralink so far?

15:24I've just had a good time being able to use it

15:26as I travel, flying and drawing a little mustache on a cat.

15:31Had a lot of fun doing that.

15:32I mean, I've just had a good time playing around with it.

15:34Oh, you know what?

15:35I do know what my favorite BCI feature is.

15:38Probably not a feature, but I just, I love Web Grid more than I love anything in my life, probably.

15:45I think I can play that game nonstop forever.

15:49Has to be Fusion 360 being able to design parts, design the hat logo with the BCI.

15:58That's what's up.

15:59Pretty sweet.

16:00That's sweet.

16:00Yeah.

16:01Yeah.

16:01I have a little Arduino that takes input from my quad stick,

16:09converts it into a PPM signal to go to a RC truck,

16:14little rock crawler.

16:17Well, with the BCI, I room code to drive the plane with the quad stick.

16:26That's awesome.

16:27The best thing I like about our length is being able to continue to provide for my family and continue working.

16:42I think my favorite thing is probably been able to turn on my TV.

16:47Yeah, like the first time in two and a half years I was able to do that.

16:51So it's pretty sweet with all these sheeting all these sets.

16:54That's kind of nice.

16:55Excited to see what BCIs got going on.

16:57I got a question, what's your shirt say?

17:00I said, I do a thing on whatever I want.

17:03Yeah.

17:04Yeah.

17:05Yeah.

17:14Now, one of the major figure merits that we have is to keep track of monthly hours of

17:20independent BCI use.

17:21Effectively, are they using the BCI?

17:24And not at the clinic, but at their home.

17:27And what we have noticed, and this is a plot of all of the different participants, first

17:32participants and their usage per month over the course of the last year and a half.

17:37And we're averaging around 50 hours a week of usage.

17:41And in some cases, peak usage of more than 100 hours a week, which is pretty much every

17:45waking moment.

17:54So I think it's been incredible to see all of our participants demonstrating greater

17:58independence through their use of BCI.

18:03Not only that, we've also accelerated our implantation cadence as we've amassed evidence

18:08of both clinical safety as well as value to our participants.

18:12So to date, we have four spinal cord injury participants as well as three ALS participants

18:17with the last two surgeries happening within one week of each other.

18:24And we're just beginning.

18:26This is just tip of the iceberg.

18:28Our end goal is to really build a whole brain interface.

18:32And what do we mean by a whole brain interface?

18:34We mean being able to listen to neurons everywhere, be able to write information to neurons anywhere,

18:41be able to have that fast data wireless transfer

18:44to enable that high bandwidth connection from

18:46our biological brain to the external machines

18:50and be able to do all of this with fully automated surgery

18:54as well as enable 24 hours of usage.

18:58And towards that goal, we're really working

19:03on three major product types.

19:05Elon mentioned earlier that our goals

19:06to build a generalized input output platform

19:09and technology to the brain.

19:11So to afford the output portion of it,

19:14which is extremely slow through our meat sticks,

19:17as Elon calls them, meat hands that are holding the mics.

19:25We're starting out with helping people with movement disorders,

19:29either through where they lost the mind body connection,

19:31either through a spinal cord injury, ALS, or stroke,

19:35be able to regain some of that digital, as well

19:38as physical independence, through a product

19:39that we're building called telepathy.

19:41And this is our opportunity to build a high-channel read

19:45and output device.

19:47On the input side of things, there's

19:50opportunities for us to help people that have lost the ability

19:53to see, be able to regain that site again,

19:57through a product that we're calling Blind Sight.

20:00And this is our opportunity to build high-channel write

20:03capabilities.

20:04And last but not least, be able to also help people that are

20:08suffering from neurological, debilitating dysregulation

20:13or psychiatric conditions or neuropathic pain

20:17by inserting our electrodes in reaching any brain regions

20:21to be able to insert them, not just on the cortical layer,

20:25but into the salt guys, as well as deeper parts of the brain,

20:29the so-called limbic system, to really enable better

20:33opportunities to just regain some of that independence.

20:36Our north star matrix is one, increasing the number of neurons

20:40that we can interface with, and second,

20:43to expand to many diverse area, any parts of the brain,

20:47starting with microfabrication or lithography

20:50to change the way in which we can actually

20:53increase the number of neurons that we can see from a single

20:55channel, and also doing mixed signal chip design

21:00to actually increase the physical channel

21:03to increase more neurons that we can interface

21:07to sort of allow more information from the brain

21:11to the outside world.

21:13And then, you know, everything we built

21:15from day one of the company has always been

21:17read and write capable.

21:19And with Tel Aviv, our first product,

21:22the focus has been on the read capabilities or the output.

21:25And we want to hone in on our write capability

21:28and also show that through accessing deeper regions

21:31regions within the visual cortex that we can actually achieve functional vision.

21:49So now just to step you through what the product evolution is going to look like in the next

21:52three years, today what we have is 1000 electrodes in the motor cortex, the small part of the

21:59brain that you see in this animation called the hand knob area that allows participants

22:04control computer cursors as well as gaming consoles.

22:08quarter we're planning to implant in the speech cortex to directly decode

22:14attentive words from brain signals to speech and in 2026 not only are we going

22:24to triple the number of electrodes from 1000 to 3000 for more capabilities we're

22:31planning to have our first blindside participant to enable navigation and in

22:442027 we're going to continue increasing channel counts probably another triple

22:49So, 10,000 channels.

22:51And also, enable, for the first time, multiple implants.

22:55So, not just one in motor cortex, speech cortex,

22:59or visual cortex, but all of the above.

23:04And finally, in 2028, our goal is to get to more than 25,000

23:09channels per implant, have multiple of these,

23:12have ability to access any part of the brain

23:15for psychiatric conditions, pain, dysregulation,

23:19and also start to demonstrate what it would be like to actually integrate with AI.

23:32All this is to say that we're really building towards

23:35set of fundamental foundational technology that would allow us to have

23:39hundreds of thousands if not millions of channels with multiple implants

23:42for whole-bin interfaces that could actually solve not just these

23:46devilitating neurological conditions but be able to go beyond the limits of our biology.

23:51This vertical integration and the talented team that we have at Neuralink has been

23:55and will continue to be the key recipe for rapid progress

23:58that we will be making.

24:00Just to recap real quick, Neuralink

24:02is implanted with precision surgical robot.

24:05It's physically invisible.

24:07And one week later, users are able to see their thoughts

24:10transform into actions.

24:12And to share more about what that experience is like,

24:16I'd like to welcome Sahedge to the station.

24:32What's up, guys?

24:33My name is Sahedge.

24:34I'm from the brain computer interface team

24:35here at Neuralink.

24:37And I'm going to be talking about two things today.

24:39The first thing is what exactly is a Neuralink device capable

24:43of doing right now?

24:44And the second one is how does that actually impact the day

24:47to day lives of our users?

24:50Very simply put, what the Neuralink device does right now

24:54is it allows you to control devices simply just by thinking.

24:59Now to put that a bit more concretely,

25:01I'm about to play a video of our first user.

25:04His name is Nolan, if you remember from DJ section.

25:06And what Nolan is doing is he's looking at a normal off the shelf

25:11MacBook Pro.

25:13And with his Neuralink device, as you're going to see,

25:16he's going to be able to control the cursor simply

25:18with his mind, no eye tracking, no other sensors.

25:21And what's special about this particular moment

25:24is this is the first time someone

25:26is using a Neuralink device to fully control their cursor.

25:32This is not your ordinary brain-controlled cursor.

25:36This is actually a record-breaking control,

25:39literally on day one, beating decades

25:42of brain computer research.

25:43And I'm about to show you the clip

25:45on day one, Nolan breaking the BCI world record.

25:51Whoa!

25:52Oh!

25:53Oh!

25:54You need to build records.

25:54Yeah, yeah!

25:55Oh, shit.

25:56Yeah!

25:57Yeah!

25:58Oh, my gosh!

26:01Oh, well, Doug, man.

26:02How's that?

26:03Hey, put it on there.

26:04He's a new world record holder.

26:06I thought it was higher.

26:11I thought I would have to get the five or something.

26:13Oh my gosh, that's crazy.

26:16It's pretty cool.

26:19Yeah, another really fun thing you could do with the

26:33Neuralink device outside of controlling a computer cursor is you could actually plug it

26:37plug it in through USB through a lot of different devices.

26:40And here, we actually have Nolan playing Mario Kart.

26:43Now, what's special about this particular clip is,

26:46Nolan is not the only cyborg playing Mario Kart in this clip.

26:50We actually have a whole community of users,

26:52as mentioned earlier.

26:53And this is literally five of our first users of New Orleans

26:57playing Mario Kart together over call.

26:59Yeah.

27:00Yeah.

27:06Yeah.

27:07Mario Kart is cool.

27:10You're using one joystick.

27:12and then you're clicking like a couple of buttons

27:14to throw items.

27:15What would be even cooler is what if you could control two

27:18joysticks at once simultaneously with your mind?

27:22When I'm about to show you, and I think this is for the first

27:24time someone playing a first person shooter game with a

27:27brain computer interface, this is Alex and RJ playing

27:31Call of Duty, controlling one joystick to move and then the

27:35other joystick to like point your gun and then shooting

27:37people as a button.

27:39Here's Alex.

27:41Here's Alex shooting another person.

27:43Right.

27:44Oh dear God.

27:46Oh, that was kind.

27:48I don't know what to do, but I want him to freak out you.

27:51What am I going to do?

27:52Are you going to do that?

27:53Alex, I got you.

27:54I know it is shot in the face.

27:58Now that we have a bit of a sense of what the BCI can do,

28:01a very important question to answer is how does this impact the day-to-day lives

28:06of the people that use it every day?

28:09So I'm about to show you a clip going back to Nolan for a second, where he talks.

28:14We simply just asked him randomly during a day how he enjoys using the BCI a couple months

28:19ago.

28:20And this is his candid reaction.

28:22I work basically all day from when I wake up, trying to wake up at like 6 or 7 a.m.

28:31I'll do work until session, I'll do session, and then I'll work until, you know, 11, 12

28:40PM or 12 AM.

28:44I'm doing like, I'm learning my languages, I'm learning my math, I'm like relearning all

28:53of my math.

28:55I am writing, I am doing a class that I signed up for.

29:02I wanted to point out that this is not something I would be able to do without the NERL.

29:11Next, I want to talk a bit about Brad.

29:13You guys may already know him as the ALS cyborg.

29:16Brad also has ALS, and what separates him from our other users is he's actually non-verbal,

29:20so he can't speak.

29:22Um, why this is pretty relevant is he relies, at least before the

29:26near link on an eye gaze machine to communicate.

29:29And a lot of eye gaze machines, you can't use outdoors.

29:32Uh, you really need like a dark room.

29:34Um, so what this means is for the last six years since Brad's been

29:37diagnosed with ALS, he's really unable to leave his house.

29:41Now with the near link device, we're going to show you a clip of him

29:44with his kids at the park, Shopeye, Ashley Vance and the team.

29:48You guys ready?

29:51I'm so speaking.

29:52I am absolutely doing more with Norlinx than I was doing with EyeGaze.

29:57I have been a Batman for a long time, but I go outside now.

30:03Going outside has been a huge blessing for me,

30:06and I can control the computer with telepathy.

30:09Dad's watching. Look, he's watching on the camera.

30:12Is he really one of the arms?

30:15The last user I want to talk about is Alex.

30:18You've seen some clips of him earlier.

30:20What's special about Alex to me is he's a fellow left-handed guy who writes

30:25in cursive all the time. And what he mentioned is since his spinal cord injury

30:29from like three, four years ago, he's been able, unable to just like draw or write.

30:35And he always brags about how good his handwriting was. So we actually got to put

30:39it in the test. We gave him a robotic arm. And I think this is the first time you tried using

30:42the robotic arm to write anything. And this is a sped up version of writing at the convoy trial

30:48and drawing something.

30:49Now, yeah, controlling a robotic arm is cool, but this one has a clamp.

31:02And what would be cooler is if you could decode the actual fingers, the actual wrist,

31:07all the muscles of the hand in real time.

31:09Just in the past couple of weeks, we were able to do that with Alex,

31:14and you're about to see him and his uncle playing again.

31:18Rock paper scissors shoot. Rock paper scissors shoot. Rock paper scissors shoot.

31:30Rock paper scissors shoot. That was scissors thought.

31:35That was his bullshit.

31:37He was so scared.

31:42Is that more?

31:46Can I do it?

31:48Whoa!

31:50Cool.

31:53Controlling.

31:55Yeah, that's pretty dope.

31:57I don't know.

32:03And controlling a robotic hand on screen is obviously not super helpful for most people.

32:09Fortunately, we have connections with Tesla

32:13who have the Optimus hand and we're actually actively working on giving Alex an Optimus hand so that you can actually control it in his real life.

32:22And here's actual replay of the end of that video using Alex's neural signals on an Optimus hand.

32:29Sean, if you want to play that?

32:37Yeah.

32:38Actually, let me add a few things to that, which is...

32:46So, as we advance the neural link devices, you should actually have full body control

32:55and sensors from an Optimus robot.

32:58So you could basically inhabit an Optimus robot.

33:02So not just the hand, the whole thing.

33:05So you could basically, mentally, remote into an Optimus robot and be kind of cool.

33:15The future's gonna be weird.

33:16But, but, but pretty cool.

33:23And then, now I was, another thing that could be done also

33:27is like for people that have say lost a limb,

33:30lost an arm or a leg or something like that,

33:32then we think in the future we'll be able to attach

33:37an Optimus arm or legs.

33:39And so you kind of like, I don't know if you're seeing

33:42from Star Wars where Luke Skywalker gets his hand, you know,

33:47chopped over the lightsaber and he gets kind of a robot hand.

33:51And I think that's the kind of thing

33:52that we'll be able to do in the future working

33:56with the Neuralink and Tesla.

33:58So it goes far beyond just operating a robot hand

34:02but replacing limbs and having kind of a whole body robot

34:04experience.

34:05And then I think another thing that will be possible,

34:09I think, is very likely in the future,

34:12is to be able to bridge the, where the damaged neurons are.

34:15So you can take the signal from the brain

34:18and transmit that signal past where the neurons are damaged

34:23or strained to the rest of the body.

34:25So you could reanimate the body.

34:27So that if you have a neural link implant in the brain

34:30and then one in the spinal cord,

34:33then you can actually bridge the signals

34:35and you could walk again and have full body functionality.

34:40Obviously, that's what people would prefer

34:42to be clear.

34:43We realized that would be the preferred outcome.

34:45And so that even if you have a broken neck or you can still

34:52we believe, I'm actually at this point,

34:54I'd say fairly confident that at some point in the future

34:57we'll be able to restore full body functionality.

35:07Yes, hello, everyone.

35:09My name is Nia, and I am leading the BC application group.

35:12And I think the video is just that I just shared with you.

35:17I probably watch them maybe thousands of time,

35:20but still I get the goose bump every time I watch them.

35:23And I think this is one of the cool perks here at Neuralink

35:25when you get the job is that you might get goose bump

35:28every week, or maybe every few days in good weeks.

35:30And this is really fun.

35:35As an engineer, it's really cool because you can build

35:38a new feature, you can build a new machine learning model

35:40and your software feature and test it,

35:43understand that they were the participant and get feedback.

35:46And you already saw with our first device, telepathy,

35:50that we can address a very diverse needs of the different users

35:55that we have for moving a cursor, to playing games,

35:59to move your body calm with multiple fingers.

36:02And we could not have done it without the New Orleans

36:04device.

36:05The New Orleans device gives us something

36:06that no other device can give us, which

36:09is a single new recording from thousands of channels

36:13simultaneously.

36:14that telepathy products is basically recording the neural activity from the small area and

36:19the motor cortex that involve an execution of hand in arm movements. But if we go only

36:24about two or three inches below, there's another brain area that's involved in execution of

36:29speech. And with the same device, with the same machine learning model architecture, the

36:34same software pipeline, the same surgical robot, we can have a new application and we

36:39can do it very quickly. It's really interesting that if we can decode someone's

36:44intention to speak silently and on vocal communication, we can use that to

36:49revolutionize the way we interact with computers, with technology, and with

36:53information. Instead of typing with your finger or like moving the mouse or

36:58talking to your phone, you'll be able to interact with computer with a speed of

37:02thought. It will make this interaction much faster and much more intuitive. The

37:08computers will understand what you want to do.

37:11And we can also expand that to AI.

37:13We can now build an interface with AI

37:16that you will be able to achieve information,

37:19will be able to store our thoughts anywhere, anytime,

37:23privately and silently.

37:26Again, because we build a fundamental technology,

37:28a platform, and we do everything in-house,

37:31we own the entire stack from neurons

37:34to pixels on the user's computer.

37:36Now I'll pass it to Ruth to talk about UI for this year.

37:48Each spike that our implant detects

37:51goes on a fairly remarkable journey

37:53to ultimately form a pixel on a participant's display.

37:57And that experience starts with, of course, unboxing

38:02the very first time that a participant pairs to and meets

38:05their implant, this invisible part of their body,

38:10and seize their own spikes materialized across the display.

38:15From there, they'll go into body mapping

38:17and actually imagine moving their arm again

38:20and get a feel for what feels natural to them

38:22and what doesn't.

38:24And they'll take that into calibration.

38:28Using one of those motions

38:30to actually move a cursor again.

38:34Iteratively refining their control

38:36as they go throughout this process

38:38until finally they're teleported back to their desktop

38:44and can experience the magic of neural control

38:47for the very first time.

38:49And our control interfaces is where the OS integration

38:54that we do really shines,

38:56letting us adapt both control and feedback

39:00for every interaction.

39:02So for familiar interactions like scrolling,

39:05we can surface an indicator over the scrollable parts

39:07of the display, add a touch of gravity

39:10to automatically pop a participant's cursor

39:13onto that indicator as they approach,

39:16show the actual velocities that we decode inside of it,

39:19and add a bit of momentum to those velocities

39:22to carry them forward as they glide across the page.

39:26There are also unique interactions

39:28that we need to solve for in this space.

39:30For example, when a participant is watching a movie

39:33or just talking to somebody next to them,

39:36the brain is very active still.

39:38and then activity can actually induce motion in the cursor,

39:41distracting them from that moment.

39:43So when a participant wants to just get their cursor

39:45out of the way, they can push it

39:47into the edge of the display to park it there.

39:50And of course, we add gravity to sort of hold it still,

39:53but they can push it out with either just a firm push

39:57or in this case, a gesture.

40:00And of course, it's with it goes without saying

40:02that all of these control interfaces

40:04are designed hand in hand with our participants.

40:07So huge shout out to both Nolan and Brad

40:09for helping us design these two.

40:11And those control interfaces, of course, extend typing.

40:15We have a great software keyboard that does everything

40:17you'd expect it to, popping up, one of our distinct clicks

40:20on a text field, giving them feedback about the click

40:22along the surface of the key and supporting

40:25both dictation and swipe.

40:33Hi, everyone.

40:34I'm Harrison, an ML engineer here at Neuralink.

40:37And I must say, being an ML engineer at Neuralink

40:39is a bit like being a kid in a candy store.

40:44When you think of the inputs to most ML systems out there,

40:47you might think of pixels, of tokens,

40:49or of a user's Netflix watch history.

40:51The input to our systems is a little different.

40:54It is pure raw brain power.

40:57And when we think about the ML systems

40:59we can build here at Neuralink,

41:01really we're limited by our imagination and our creativity.

41:03There's no reason our ML systems can't do anything

41:06that the human brain can do,

41:08such as controlling a phone, typing, or even gaming.

41:14Right here to my left is actual footage of Alex,

41:16one of our participants playing at first person shooter

41:20against RJ, another one of our participants.

41:22Now for those unfamiliar with first person shooters,

41:25this is not a trivial feat.

41:26It requires two fully independent joysticks

41:29or four continuous degrees of control

41:32as well as multiple reliable buttons.

41:36Now contrary to popular belief,

41:38the Neuralink does not simply read people's minds.

41:41It's simply reading neuronal activations

41:43corresponding to motor intent.

41:45So one of the fun challenges with this project

41:48was figuring out which motions were

41:49going to be about to the joystick.

41:51We started with the typical left thumb and right thumb,

41:54but quickly found that the dominant hand overshadowed

41:56the non-dominant hand.

41:58My personal favorite is we had one of our participants

42:00imagine walking for the left joystick

42:02and aiming for the right joystick.

42:04So in-game, they were simply doing naturalistic motions,

42:08like you might do in virtual reality in Ready Player One.

42:11And that was a really cool to watch.

42:13What we ended up on was the thumb for the left joystick

42:16and the wrist for the right joystick.

42:18And I challenged the audience to try to replicate their motions.

42:21I'm really in awe of them being able to pull this off.

42:24I wanna talk a bit about the progress

42:26to our cursor calibration experience.

42:28To my left here, you can see RJ

42:30completing his first ever cursor calibration

42:33with a redesigned open-loop flow

42:34where he first gathered information about his intent

42:36and how to map the neural activity

42:38to the first time he controls a cursor,

42:40to the final product

42:41where he has smooth and fluid control of his computer.

42:44And most remarkably, this experience

42:47took only 15 minutes from start to finish,

42:5015 minutes from not, 15 minutes from no control

42:58to fluid computer use.

43:01Contrast that to a year and a half ago with P1,

43:04where that was multiple hours to get to the same level

43:06of control and several engineers standing around a table,

43:09pulling their hair out.

43:10There was virtually no need for

43:12neural and engineering to even be at this session.

43:14This was basically an out-of-the-box experience

43:16for our participant.

43:21And even more remarkably, we're continuing

43:23to smash day one records with RJ being able to achieve

43:277 BPS on his very first day with an early.

43:34Now, such an effective and efficient calibration process

43:37is only made possible by high fidelity estimations

43:40of a user intention or labels.

43:43And to briefly illustrate just how challenging

43:45of a problem that is, this is an animation of myself

43:48trying to draw circles on my desktop with a mouse.

43:51Now the task was simple, draw uniform circles

43:54at a constant speed repeatedly.

43:57And as you can see by that animation, I am horrible at that.

44:00Even though my intent was pretty obvious, unambiguous,

44:03the execution was really poor.

44:05There was a ton of variation in both speed and the shape

44:09itself.

44:11To visualize this a little differently,

44:13each row here is one of those circles unwound in time

44:16with synchronized starts.

44:17And you can just see how much variation there is in the timing

44:19of each circle, as well as what I'm doing at any given point in time.

44:25Orthogonal to the labeling problem is neural non-stationarity,

44:28or the tendency of neural signals to drift over time.

44:32And I think that's honestly a beautiful thing, right?

44:34If your neural signals didn't drift, you couldn't grow.

44:37When you wake up the next day, you're not the same person

44:40you were the day before.

44:41You've learned.

44:41You've grown.

44:42You've changed.

44:43And so too, must your neural data change.

44:46This animation here is a simple illustration

44:48with the learned representation by the decoder,

44:50and how it drifts the further away we get from the day

44:52it was trained on.

44:54This is one of the key challenges we

44:55need to solve here at Neuralink to unlock fluid and a product

44:57level experience for our users.

45:08Hey, everyone.

45:12Hey, everyone.

45:13My name is Joey.

45:16Blind Sight is our project to build a visual prosthesis

45:19to help the blind see again.

45:22Users would wear a pair of glasses with an embedded camera

45:27and receive an implant in their visual cortex.

45:32Scenes from the environment are recorded by the camera

45:35and processed in the patterns of stimulation,

45:38delivered to the brain, causing visual perception

45:43and restoring functionality.

45:45Now, blind sight will be enabled by placing

45:47her implant into visual cortex.

45:49This is a new brain area for us,

45:51and this brings new opportunities and challenges.

45:54So the surface of the brain for visual cortex represents just a few degrees of angle in the

46:00center of the visual field.

46:03Larger fields of view are represented deep within the cortical folds of the calcurine

46:07fissure.

46:09Our threads are able to access these deeper structures, providing the possibility of restoring

46:15vision over a functionally useful visual field.

46:18So the N1 implant has had experimental stimulation capabilities for quite some time.

46:24But our new S2 chip is designed from the ground up for stimulation.

46:29It provides over 1,600 channels of electrical stimulation, high dynamic range recording capabilities,

46:36and a wide range of microstimulation currents and voltages.

46:41We can achieve these capabilities because we are vertically integrated and we design this

46:46custom basic in-house.

46:47Similarly, we design and fabricate our electric threads in house.

46:52And here you can see one of our standard threads designed for recording in an electron micrograph.

46:59For blind sight, our requirements are a little different.

47:02And our vertical integration allows us to rapidly iterate on the design and manufacturing of these

47:07threads for this new purpose.

47:10So here, I'm using red arrows to highlight the electrode contacts, which are optimized

47:15for stimulation.

47:17And as you can see, they're a little bit larger, which results in a lower electrical impedance

47:20for safe and effective charge delivery, which is important for blind sight.

47:27Now how can we calibrate our implant for blind sight?

47:30So here's one way.

47:32We stimulate on the array picking, say, three different channels.

47:37The user perceives something, say, three spots of light, somewhere in their visual field,

47:41and points at them.

47:43We track their arm and eye movements and repeat this process for each of the channels on the

47:47And here's what a simulated example of a blind sight vision could look like after calibration.

48:07Now I showed you how for blind sight we need to insert threads deeper into the brain than

48:11we have previously, and doing this requires state-of-the-art medical imaging.

48:16So we worked with Siemens to get some of the best scanners on Earth.

48:20We built out our imaging core from scratch in the past year.

48:24Actually, it was faster than that.

48:26It was about four months from dirt to done.

48:29Since bringing the scanners online, we've scanned over 50 internal participants building

48:33out a database of human structural and functional anatomy.

48:37What can we do with the imaging information from these scanners?

48:40So medical imaging can be used for surgical placement.

48:43It lets us parsulate our brain regions by their function.

48:46And we use our imaging capabilities to refine the placement for telepathy.

48:51It also gives us the capability to target new brain regions for future products such

48:54as blind sight or speech press thesis.

48:57And we're working towards more capabilities.

48:59So one-click automated planning of surgery from functional images to robot insertion targets.

49:05Here you can see a screen capture from one of our in-house tooling to do end-to-end surgical

49:09planning.

49:10You can see a region of motor cortex known as hand knob and the thread trajectory plans

49:15that will be sent directly to the robot.

49:18This is a really incredible degree of automation that's only possible because we're controlling

49:23the system from one end to the other.

49:25My name is John and I lead the robot mechanical team.

49:32This is our current R1 robot.

49:35It was used to implant the first seven participants.

49:38This robot works really well, but it has a few flaws, one of which is the cycle time

49:42is rather slow.

49:43So to insert each thread, it takes, in a best case scenario, 17 seconds.

49:48And many cases, external disturbances cause us to have to retry to reinsert grass-fed

49:54thread and then reinsert it.

49:55To scale our number of neurons access through higher channel count, increased numbers of threads,

50:01need to have a much faster cycle time.

50:03So let me introduce our next generation robot,

50:07which is right here.

50:17Through rethinking the way that we hold the implant

50:19in front of the robot by holding it directly

50:21in front on the robot head, we will achieve an 11 times cycle

50:25time improvement.

50:26So each thread takes one and a half seconds.

50:29We also scale up a lot of surgery,

50:32workflow process improvements through deleting

50:34the separate operator station and implant stand.

50:38Now, the outside of the robot looks pretty similar

50:40between the two, but it's what's inside that really counts.

50:43Each system has been redesigned from the ground up

50:46with a focus on reliability, manufacturability, service

50:49ability, and using a lot of our vertical integration

50:52techniques, it's enabled us to have a lot more control

50:55of the system end to end.

50:58Now, that fast cycle time doesn't mean much

51:00if it's not compatible with a significant portion

51:02of the human population.

51:04Prior to each surgery, we scanned our participants' anatomy

51:08and ensure that they will be compatible with the robot

51:10and vice versa.

51:11Unfortunately, the robot isn't compatible with everyone,

51:14so we had to extend the reach of the needle

51:16in the next generation robot, and now we're compatible

51:19with more than 99% of the human population.

51:21We've also increased the depth that the needle

51:23can insert threads.

51:24Now we can reach more than 50 millimeters

51:26from the surface of the brain,

51:27accessing and enabling new indications.

51:30We have to produce a ton of custom sterile components

51:33for each surgery.

51:34We actually supply more than 20 of these parts.

51:37Many of these parts are made through traditional CNC

51:39manufacturing capabilities, which we do just on the other side

51:42of this wall, actually.

51:43And some custom-developed processes,

51:45like this femtosecond laser milling,

51:47used to manufacture the tip of the needle.

51:49Now, these processes take quite a bit of time, effort,

51:53and cost.

51:54So let's take a look at how we're going to reduce costs and time

51:58for one of the components.

51:59So the current needle cartridge has a total cycle time

52:02of about 24 hours, and the machine components

52:05cost about $350.

52:07The final assembly is performed by a set of highly skilled technicians.

52:10They have to glue a 150 micron diameter cannula onto this wire EDM machined stainless steel

52:16base plate.

52:17They have to electropolish a 40 micron wire into a sharp taper.

52:21And then they have to thread that 40 micron wire into a 60 micron hole in the cannula.

52:25This is done manually.

52:28And then they finally have to laser weld all the components together.

52:33Next generation needle cartridge takes only 30 minutes of cycle time and $15 in component.

52:38We were able to delete the wire EDM machined base plate

52:41and the kenio gluing step by switching to an insert mold

52:44it component.

52:44So we get a box of these base plates

52:46with the kenio was already installed for like 1,000 of them

52:49for like a couple of $5, 10 a piece.

52:52We also deleted the electro polishing step

52:54with the revised needle tip geometry, which

52:56is also compatible with inserting the threads

52:58through the dura.

53:00We have a few revised manufacturing techniques

53:02to delete the manual threading through basically a funnel,

53:05rather simple, but it has been a big impact.

53:08And then we're able to delete the laser welding

53:09through using crimping.

53:16Hi.

53:23Hi, I'm Julian.

53:23I'm one of the leads on the implant team.

53:26So the way humans communicate today,

53:29if they want to output information,

53:31is by using their hands and their voice,

53:33as I'm doing right now.

53:34And if you want to receive information,

53:36you use your ears and your eyes.

53:38And of course, that's how you're receiving this very talk.

53:42But we've built this implant.

53:44And this implant is very special,

53:46because it is the first time that we're

53:49able to add a completely new mode of data transfer

53:53into and out of the brain.

53:55If you look at this device in a nutshell,

53:57it's really just sampling voltages in the brain

54:00and sending them over radio.

54:02But if you zoom out and look at the system from end to end,

54:05what you actually see is that we're

54:08connecting your brain or a biological neural net

54:11to a machine learning model or a silicon neural net

54:15on the right-hand side.

54:17And I actually think this is really elegant,

54:20because the machine learning model on the right-hand side

54:23is, in fact, inspired by neurons on the left-hand side.

54:27And so in some sense, we're really

54:28extending the fundamental substrate of the brain.

54:32For the first time, we're able to do this

54:33in a mass-market product.

54:35That's a very, very special piece of hardware.

54:45So these are some of the first implants that we ever built.

54:49There are electrodes that were made with our in-house

54:53lithography tools.

54:54We have custom Asics that we also designed in-house.

54:57And this was really a platform for us

54:59to develop the technology that allows

55:01us to sense micro-level volts in the brain

55:04across thousands of channels simultaneously.

55:07We learned a lot from this, but as you'll

55:09notice in the right two images, there

55:11are USBC connectors on these devices.

55:14These are not really the most implantable implants.

55:19This next set of images are the wireless implants.

55:22And there was a complete evolution that we went through

55:25to add the battery, the antenna, the radio,

55:28and to make it actually fully implantable.

55:30Once it's implanted, it's completely invisible.

55:33It's very compact, it's modular,

55:35and it's a general platform

55:36that you can use in many places in the brain.

55:40Going from that top row to the bottom row

55:43is very challenging.

55:45The implant you see on the bottom right here

55:47is in fact the device that we have

55:48working in seven participants today,

55:51and it's augmenting their brain every day

55:53and restoring their autonomy.

55:55But getting to that point involved

55:57a huge number of formidable engineering challenges.

56:00We first had to make a hemetic enclosure

56:02passing a thousand separate conductors

56:04through the enclosure of the device.

56:06We had to figure out how to make charging seamless

56:09and work with very tight thermal constraints

56:11in a very, very small area.

56:13And then we also had to scale up

56:14our testing infrastructure so that we could support

56:16large scale manufacturing and very safe devices

56:19have confidence in our iteration cycle.

56:22So what's next?

56:24We're going to be increasing our manufacturing

56:27so that we don't just produce a certain, like,

56:30a small number of implants per year,

56:31but thousands and then eventually millions

56:32of implants per year.

56:34We're also going to be increasing channel count.

56:37More channels means more neurons are sensed,

56:40which means more capabilities.

56:42In some sense, we often think a lot

56:44about the Moore's law of neurons that we're interacting with.

56:47And in the same way that Moore's law propelled forward,

56:50many subsequent revolutions in computing,

56:53we think that sensing more and more neurons

56:55will also completely redefine how we interact

56:58with computers and reality at large.

57:02I wanna leave you with one final thought.

57:04When I was a child, I used a 56 kilobit modem

57:08to access the internet.

57:11If you remember what it's like, you would go to a website.

57:14You're lucky.

57:15You're lucky bastard.

57:17Yeah, I was a child.

57:19We had acoustic couplers.

57:20Oh, yeah, OK.

57:21So just beep it to just beep it each other.

57:24Yeah, the first modem was the acoustic coupler.

57:27Incredible device, honestly.

57:30But then if you, I guess, if you're my age,

57:32you started with 56 k-bit modem.

57:33Oh, yeah.

57:36And you would go to a website and there would be an image.

57:43And it would scroll.

57:45Slowly, it was loading pixel by pixel on the screen.

57:48So that's what it's like to be bandwidth limited.

57:51Now imagine using the current internet with that same modem.

57:54It's inconceivable.

57:56It would be impossible to do.

57:58So what broadband internet did to the 56-helibut modem

58:03is what this hardware is going to do to the brain.

58:06We are trying to drastically expand the amount of bandwidth

58:10that you have access to to have a much richer experience

58:13and superhuman capabilities.

58:15So I guess just to kind of close out and to recap, today, Neuralink is working reliably

58:29and has already changed the lives of seven participants and making a real impact.

58:36And our next milestone is to go to market and enable scaling of this technology to

58:40thousands of people, as well as expand functionalities beyond just the movement to enable

58:47robotic sophisticated robotic arm control, speech, vision,

58:52give sight back, and even getting to the speed of thought.

58:55I hope you got a good sort of sample of our technology stack

59:00and the challenges that we have.

59:02And I'd like to hand over the mic to Elon

59:05for any closing remarks.

59:07Well, we're trying to give you a sense of the depth

59:14of talent at Neuralink.

59:15There's a lot of really smart people

59:17working on a lot of important problems.

59:20This is one of the most difficult things

59:23to actually succeed in creating and have it work

59:26and work at scale and be reliable

59:28and available for millions of people

59:30at an affordable price.

59:33So super hard problem

59:37and we'd like to have you come join and help us solve it.

59:40Thank you.