Neuralink Just Demonstrated the Future of HUMAN-MACHINE FUSION? (And This Is Just the Beginning)

Amazing Reality

Brain implants that let you control a cursor, AI recreating what people see, AI prescribing acne treatments, virtual cells, and a robot learning to work on its own—this episode is packed with wild tech. Plus, meet the cats using GPS and AI to help find missing pets!  Subscribe for more fascinating tech stories. Tell us in the comments what you liked most about the video!  #AI #Technology #Robots #Science #Innovation  👉 This channel was created in collaboration with https://www.youtube.com/@realidadeimpressionante   • The Inspired Business Technology de RomanSenykMusic  0:01 - Neuralink’s Brain-Computer Interface 5:52 - Reconstructing Images From Brain Activity 9:15 - AI Prescribes Acne Treatment 11:55 - Building Virtual Cells 14:25 - Kawasaki’s Autonomous Humanoid Robot 16:53 - Stretchable Battery Materials 18:24 - Cats Help Find Missing Pets

Transcript

00:01And in today's episode, among other things, you'll see Neuralink has just reached a very

00:06important milestone with its brain implant. And what they showed might be showing us the

00:11first steps of a true merger between man and machine. Meanwhile, another artificial intelligence

00:18was able to observe the activity of an entire brain and with that reconstruct what a real

00:24person was seeing at that moment. In the United States, an artificial intelligence has just been

00:30authorized to make a decision that, until now, had to go through a human doctor. Google, Meta and the

00:36US government are investing billions of dollars in a technology that aims to create. Something

00:42pretty strange, okay? Virtual human cells. Yep, let's make sense of all that and a few more developments

00:49in tech.

00:55A person just looks at the computer screen, doesn't touch the mouse, doesn't touch the

01:00keyboard, doesn't touch anything, but the cursor moves. And now, in some tests, it doesn't even

01:06need to travel all that distance to the desired spot anymore. It simply appears right where

01:11the person intends to click. Yeah, my friends, it's evolving pretty fast. This is a technology

01:16Neuralink calls Teleport Cursor or Teleported Cursor, and it's part of an advancement that

01:22might seem smaller at first glance, you know? I know, but it shows how this direct connection

01:26between the human brain and computers is now starting to change forever.

01:32And Neuralink just announced a new record for their brain implant. Using signals recorded

01:37directly from the brain, participants managed to reach 11.32 bits per second account controlling

01:42the cursor bits per second account. In this case, it doesn't just mean the physical speed

01:47at which the cursor moves across the screen, alright? This metric considers both the speed

01:51and accuracy with which a person can select the item they actually want on the screen.

01:56And according to Neuralink, the result already surpasses the performance of most able-bodied

02:02people in this specific type of task. But it's good to make it very clear that this comparison

02:07hasn't been independently verified yet. But maybe this record isn't even the most important

02:11part of this story because there is a problem that plagues practically every brain-computer interface.

02:16Our brain, the biological brain, doesn't produce the exact same signals every day.

02:21And the Neuralink implant records tiny electrical pulses that are produced by neurons in the cortex.

02:27Motor cortex, the software then learns to recognize certain patterns in that activity and link them to

02:33the intention of making a movement. That is exactly how someone can think about moving the cursor and the

02:38computer turns that intention into a command. It's just that these signals, as I mentioned,

02:42change every day. Slight changes in the recorded neurons can make what worked yesterday no longer

02:48be interpreted the same way today. That's why the systems need to be recalibrated. And that's

02:52precisely where one of the most interesting breakthroughs in this story happened. Neuralink has

02:57gathered over 50,000 hours of neural activity generated by the participants in its clinical trials.

03:02And they use this massive dataset to train what they describe as a foundation model for the brain

03:08computer interface. Instead of just teaching the system to recognize specific commands,

03:13the idea is to have the software learn more general representations of this brain activity.

03:18In practice, it's as if the AI is now beginning to learn something much deeper about how different

03:23intentions to move appear in the system through signals from the human brain. And the initial results

03:29are quite significant. Some of the new decoders kept working for weeks, instead of just a few days,

03:33as was previously the case. For certain participants, the time required for recalibration dropped from

03:40roughly 10 minutes every day to just 10 minutes a week. And that changes the way we look at this

03:45new

03:45technology quite a bit. Because there is a huge difference between an experiment that works under

03:50carefully prepared conditions and a brain-computer interface that you can simply use. Using it day to

03:56day without constantly having to recalibrate it. But Neuralink wants to go way beyond that. The company now

04:03aims to combine neural data from different participants to try and build even more generalized

04:07models. And in the future, this could allow systems capable of performing a new task with very little

04:13calibration, or perhaps eventually, of course, without needing task-specific calibration at all.

04:19And this is where a particularly interesting idea comes up, you know? Neuralink talks about the future

04:24possibility of creating something similar, similar to an API for the motor cortex. In computing, an API works

04:31as a kind of common layer that allows different systems to talk to each other. Applied to the

04:35brain, the idea would be to develop a sufficiently general representation of movement intention to

04:41turn it directly into digital commands. You think about performing an action, the brain produces the

04:46signals related to that intention, and the implant records those signals, and the computer, using

04:51artificial intelligence right there, interprets what you want to do. It's important to make it very clear,

04:56of course, that we are still a long way from a universal brain-machine interface. Okay, the results

05:02released now come directly from Neuralink, and haven't yet been presented as a peer-reviewed scientific

05:07study. Furthermore, the immediate goal remains restoring autonomy to people with paralysis, allowing

05:13them to control computers and other devices using only brain activity. But yeah, it's hard to ignore the

05:19direction this technology is pointing towards, isn't it? First, we manage to record signals directly from the

05:25brain, then we learn to turn these signals into cursor movements, and now large models are already

05:31starting to learn patterns from tens of thousands of hours of human neural activity, making this

05:36communication much faster, more stable, and also relying on far less calibration. And maybe that's

05:43precisely how a true integration between humans and computers machines, you could say, begins to develop

05:49very slowly, right? Number one, Neuralink tries to turn brain signals into commands for a machine.

05:57Researchers are now doing the opposite. Using artificial intelligence, they can already observe the brain

06:02activity of a real person and reconstruct the images of what they are seeing in a very simple way.

06:08According to them, the work was developed by researchers from the Wiseman Institute of Science,

06:13and they did this using functional magnetic resonance imaging, or fMRI. During the experiment,

06:19volunteers viewed different images while the equipment was recording changes in activity across

06:23various regions of their brains. Based on these signals, the artificial intelligence needed to solve

06:29an extremely difficult problem, figure out which image could have triggered that pattern of neural

06:33activity. To do this, the researchers split the task into two parts, right? One system tried to

06:40extract information related to the image's structure, like shapes, positions, and composition,

06:45and the others sought to interpret higher level information about what appears in the scene.

06:49Then, this information is combined and used by an image generation model to produce a reconstruction

06:55of what the person might have been looking at. And the results are impressive, all right? Precisely because

07:01the artificial intelligence doesn't receive the original photograph, it receives the signals produced

07:06by the brain. While someone is looking at that photograph, in some cases, the reconstruction

07:11manages to preserve the very important features of the original scene, you know? But let's be very clear,

07:16okay? For now, we are far from a machine capable of reading perfectly what's inside a person's head.

07:23And the errors make that quite clear. In one of the examples described by the researchers,

07:27a person was looking at a dog inside a bathtub. The artificial intelligence managed to reconstruct it,

07:32but it was a goat inside a bathtub. It swapped the dog for a goat. But that's already pretty weird,

07:38isn't it? It understood a good part of the structure and even the meaning of the scene,

07:42but got the animal's identity completely wrong. There's yet another problem. Training systems like

07:47this requires massive amounts of brain data and getting thousands of high-resolution fMRI scans.

07:53It is way more complicated than just gathering millions of photos from the internet. That's why the

07:59researchers also created the reverse process. They show an image to an AI model and ask it to estimate

08:04which brain activity pattern that image over there would likely produce in a human. This synthetic

08:09brain data can then be used to help train systems that try to work in reverse from the brain back

08:16to the

08:16image. That's what we are seeing here, for now. All of this depends on massive MRI machines and

08:22experimental conditions that are extremely controlled. It also doesn't mean that someone can put a person

08:28inside a machine and simply figure out whatever thought they happen to be having at that moment. What

08:34researchers are demonstrating is something much more specific, the ability to reconstruct visual content

08:40represented in brain activity. See, place the experiment right there, next to what we just saw with

08:46neural link, and a rather curious situation emerges. On one hand, machines are learning to turn our

08:52intention into digital commands. On the other, they are already beginning to transform our brain activity

08:58into new images. Images that we are probably seeing. And bit by bit, that barrier that throughout human

09:05history kept what goes on inside a person's brain private. Meaning, our thoughts are our own. It's now

09:12starting to open up small windows to the outside as well, isn't it? That's dangerous. And in the United

09:17States, an artificial intelligence has just received authorization to take over a task that until now

09:24had to be handled by a human doctor. First, prescribing medication to patients. Yeah, that's what an AI can

09:30already do over there. The state of Utah authorized a pilot program by NOLA Health which will allow them

09:35to test an artificial intelligence system capable of evaluating patients with uncomplicated acne and

09:41selecting a treatment on its own, considered much better suited to that particular patient.

09:45The process starts on a cell phone. The patient, who must be an adult, fills out a questionnaire that

09:50takes about 10 to 15 minutes, providing information about their medical history and condition. Then,

09:56they use their own smartphone to take detailed pictures of their face. The artificial intelligence

10:00then analyzes this information, assesses the condition, and can choose a treatment from a specific

10:05list of medications, which was previously defined by a human doctor, there are important limitations.

10:11Okay? The system was initially authorized only for mild and moderate acne cases, and it can't just

10:17pick any available medication. It also cannot prescribe oral medications or treatments like isotretinoin,

10:22which is used in more severe medical cases of acne. If the AI encounters any situation outside the

10:28established parameters or doesn't have sufficient confidence in its decision, the patient must then be

10:33referred to a human professional. But perhaps the most interesting part lies precisely in how Utah

10:40plans to gradually give this AI more autonomy. For the first 100 patients, all prescriptions generated by the

10:47system still need to be approved by a human doctor before they are actually sent to each patient. After that,

10:54a new phase begins. The artificial intelligence will finally be able to issue the prescription first,

10:59while doctors review the system's decisions afterward. If all results remain within the established criteria,

11:06supervision can decrease even further, moving to audits of just a single sample of these cases.

11:11In other words, the goal of the experiment is to reach a situation that is quite unusual, you know?

11:16An AI analyzing patients, choosing treatment, and issuing an initial prescription without having to wait for

11:22human doctor approval. What is being tested is not just whether artificial intelligence can already

11:28recognize a skin condition. What is really happening now is watching how far we human beings can allow

11:35an AI to go in making medical decisions, even before a human needs to intervene in all of this.

11:41Today, my friends, the limit is a restricted list of acne treatments. But if this model works safely,

11:47which is what these guys want, the next question practically appears on its own, right?

11:51What will be the next medical decision that we will be willing to hand over to a machine?

11:56And a new initiative, involving some of the biggest tech companies in the world,

12:00wants to take artificial intelligence to a completely different scale.

12:05Google, Meta, and the United States government are now participating in an effort to create

12:10virtual models capable of predicting the behavior of real cells.

12:14The initiative is led by the Chan Zuckerberg Biohub and also brings together organizations like Google DeepMine,

12:19DeepMining, and Eisenberg Labs, combining funding, infrastructure, computational resources, and other

12:25contributions. The effort involves about 1.8 billion dollars. But according to these guys, the goal isn't

12:32simply to build a database much larger than what they already have. The idea is to generate the massive

12:37amount of biological data needed to develop what researchers call virtual cells. Today, to figure out how

12:45a cell will respond to a specific change, it is often necessary to physically run multiple experiments.

12:51They modify the gene, apply a specific substance, or introduce some kind of disruption. Then, they observe

12:58what happens to those cells. And a virtual cell would try to play a part in those processes.

13:04Inside a computer, the researcher could simply input the initial state of a cell and ask the model,

13:09what would likely happen if a certain gene were turned off, or a protein were altered, or some

13:15other kind of intervention were performed there. The artificial intelligence would then try to predict

13:21the cell's response before the actual experiment was carried out in the real world. And this could

13:27allow scientists to test huge amounts of hypotheses virtually and take to the lab only those deemed the

13:32most promising. But there's a huge problem with this whole thing. Okay?

13:37Real cells are insanely complex. Even today's massive databases, containing information on hundreds of

13:44billions of them, might not be enough to teach an artificial intelligence to predict their behavior

13:49in a truly reliable way. It is estimated that models of this type may need to learn from billions,

13:54and maybe even trillions, of cells. That is why a key part of the project would be to generate new

14:01biological data sets on a much larger scale than what currently exists. But if or

14:06when this technology reaches the expected level, it could indeed change one of the most time

14:10consuming stages of biological research. Because in the future, before testing an idea on an actual

14:15cell, scientists might perhaps first figure out what will likely happen. Just by testing that same

14:21idea on its virtual counterpart, that would be interesting. And after about a decade of developing

14:27its own humanoid robot model, Kawasaki is now preparing a very significant shift.

14:32It wants to make Kaleido, its humanoid robot, increasingly less reliant on a person controlling

14:37what it's doing. This robot model started being built by the company back in 2015 and has already

14:42gone through several generations. The latest version, the Kaleido 9, can walk autonomously and use sensors

14:48like LiDAR to map the environment, locate its own position, and identify people, as well as obstacles

14:55and even stairs around it. But there's a big difference between walking on its own and working on its own,

15:00when Kaleido needs to perform more complex tasks, it still requires a heavy reliance on

15:05teleoperation or even human supervision. And that's precisely what Kawasaki intends to change,

15:11according to the company, right? They announced a collaboration with the Japanese company Noetra

15:16to integrate a physical AI platform into the humanoid robot. And the company says it's already in

15:21talks with potential clients in the automotive and semiconductor sectors to understand where

15:25humanoid robots of this kind could actually be useful to those companies. But it's important to

15:30keep the timeline and expectations realistic, alright? Kawasaki isn't saying that Kaleido has

15:36now become a completely autonomous humanoid robot. The integration of this new technology should begin

15:41around 2030, initially with operations still assisted or partially controlled by humans.

15:48And their real goal is to reach robots fully capable of working autonomously in less structured

15:54environments, a goal the company is planning to achieve around 2040. And it's important to mention

16:00that Kawasaki is also working with Nvidia on physical artificial intelligence technology. The idea is to

16:05enable increasingly complex autonomous movements for these robots. All of this shows one of the biggest

16:10challenges in this race for humanoid robots. Building legs capable of walking, arms capable of carrying

16:16objects, and hands capable of manipulating tools is already extremely difficult, isn't it? But there's an issue

16:22that's even bigger than that. Getting these machines to understand on their own what they really need

16:26to do in a world that completely changes all the time, isn't that right? And maybe that's precisely why

16:33the next big evolution in humanoid robots isn't just about building better bodies, but rather finally

16:39giving those bodies some ability to decide on their own what to do with themselves. And researchers from South

16:46Korea have developed a new material for energy storage devices that can do something quite unusual.

16:53It can be stretched to roughly nine times its original size and still keep working after that.

16:59The developed material is the type of hydrogel electrolyte that incorporates liquid metal particles,

17:04and electrolytes are fundamental components of batteries and other energy storage devices. And it is

17:09through them that ions are able to move between these electrodes during charge and discharge cycles.

17:14The problem is that if we want to build truly flexible electronics or devices capable of keeping

17:20up with human body movements, the materials used to store energy also need to withstand constant

17:26deformations. And that's precisely what the researchers tried to solve. This new hydrogel can reach

17:31approximately 900% elongation without losing its electrochemical function.

17:37Part of this behavior occurs because its structure has reversible bonds, which means that when the

17:42material is stretched, some of these bonds can break to absorb the strain. Afterward, they can form again.

17:49And this combination opens up possibilities for devices that conventional rigid batteries simply struggle

17:55to keep up with. Wearable electronics, sensors that move along with the body, bio-integrated devices,

18:02and even energy storage systems capable of repeatedly bending and stretching throughout their use.

18:07But as electronics begin to leave rigid surfaces behind to be incorporated into clothing, flexible objects,

18:14and even devices that move with our bodies. Yes, batteries will also need to learn to bend along with them,

18:21keeping up with what we're doing. And to wrap up with something a bit more laid back here,

18:26someone decided to answer a question that probably no one thought needed to be asked. What if the best way

18:33to find a

18:33lost cat was to ask another cat for help? That is precisely the idea behind an initiative in Kazakhstan,

18:40which is turning the cats that already roam freely through the neighborhoods into a kind of search team.

18:45Some animals are fitted with tiny cameras and GPS trackers attached to their bodies. After that,

18:51they don't need to learn any special tasks. They simply keep doing what cats already do every night.

18:55They walk the streets, slip into backyards, crawl under cars, explore bushes, alleys, and other places,

19:04where probably no volunteer would ever think to look. As they do this, the cameras record the cats

19:09they come across along the way, and that's where artificial intelligence comes in. The captured images

19:13can be analyzed and compared against a database containing photos and information on missing pets.

19:19When the system finds a potential match, the location recorded by GPS can then help pinpoint where

19:25that cat was spotted. According to the project leaders, the platform has already gathered over 28 000 users

19:32and helped 407 animals return to their owners. It's a pretty unlikely combination of animal behavior,

19:38cameras, GPS, and artificial intelligence to solve a problem that has existed pretty much since we started

19:44keeping pets, isn't it? If you liked this episode, you already know the drill. Leave a like, make sure you're

19:49subscribed.

19:49On your screen, there are two new suggestions, one here and the other one here. Just click and I'll see

19:54you in the next one. Alright? Thanks, I appreciate it. Bye-bye. See you soon. I'll be waiting for you

19:57there.

Category

Transcript

Recommended