news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

New Intelligence Report

Editor: Editorial Department

【New Wisdom Introduction】Musk excitedly announced that Neuralink has successfully implanted a brain-computer interface in the second human being, and all 400 electrodes went smoothly. In the latest podcast interview, Musk and his core team revealed the work of the Neuralink team and their vision for the future of mankind. He said that eight more implants will be completed this year.

The first patient implanted with a Neuralink chip has achieved mind control and mechanical ascension.

Musk once again excitedly stated that Neuralink has successfully installed a brain-computer interface on the second patient, working with 400 electrodes, and everything went very smoothly.

Noland, the first patient, controls a computer to play chess using only his thoughts

These are the remarks made by Musk, his core team and the first patient, Noland Arbaugh, on the podcast of famous technology anchor Lex Fridman last Friday (local time).

The 8.5-hour conversation was extremely information-dense and full of Musk's usual "wild fantasies" that combine genius and madness.

According to reports, the second patient also suffered a spinal cord injury similar to the first patient (paralyzed by a diving accident, less than 100 electrodes). However, the specific details have not been disclosed.

In the short term, Neuralink's priority is to address basic neurological injuries in the spinal cord, neck, or brain.

We all know that Neuralink’s first product is “Telepathy”, which helps people with damaged neurons restore their body functions.

Currently, they are developing a second product called "Blindsight" to enable blind people to see.

Initially, visual resolution was low, depending on how many neurons could be implanted.

Over time, Musk believes that Neuralink patients will have higher resolution than the human eye and may even be able to see in different wavelengths, just like Geordi La Forge in Star Trek.

During the interview, Musk expressed his highest expectations and evaluation for the future relationship between Neuralink and mankind - improving the symbiosis between AI and humans.

He said that within a year or two, people with Neuralink implants will be faster and more sensitive, surpassing professional gamers.

First patient Noland plays Mario Kart

Neuralink is already able to communicate at 8 bits per second, and in the next five years, it may reach 1M bits per second, which is faster than any human can type or speak.

"Unless we upgrade ourselves with Neuralink, AI will think so fast that communicating with humans will be boring, like talking to a tree."

Musk said he hopes to implant chips in eight more patients before the end of the year.

In addition, what other highlights did Musk and the Neuralink team members bring to the interview?

400 electrodes inserted into the brain, second Neuralink patient

Musk: "I don't want to jump to conclusions, but the second implant seems to be going very well. The signal is very strong, there are a lot of electrodes, and it works very well."

The successful second chip implantation is undoubtedly another major breakthrough for Neuralink.

In January this year, Musk's brain-computer interface company Neuralink successfully implanted the first brain-computer interface chip into a human patient.

As mentioned at the beginning, the second subject had 400 electrodes implanted in his brain.

The first patient, Noland Arbaugh, and three Neuralink team members detailed how the implant and robot-led surgery work.

Before Arbaugh received the implant in January, he could only use the device by tapping on a tablet using a stick he controlled with his mouth.

After the chip was implanted in his brain, he can now move the cursor, surf the Internet, play games and post messages just by thinking in front of the computer screen.

It can be said that the brain-computer interface has changed Arbaugh's entire life.

He has a degree of independence again and is less dependent on caregivers.

Arbaugh also experienced some problems with the device in the early days after surgery, when the implant's thin wires retracted, leaving fewer electrodes available to measure brain signals.

Neuralink said it has now restored the implant’s ability to monitor Arbaugh’s brain signals, by modifying the algorithm to make it more sensitive.

Arbaugh improved on his previous world record by being able to control the cursor with thought alone, “with only about 10 to 15 percent of the electrodes working,” Musk said on the podcast.

Long-term vision: from 10 to hundreds of millions

Lex Fridman: “Do you think there will be hundreds of millions of people in the world with Neuralink in the next few decades?”

Musk: “Yes”

Many years ago, Musk was thinking about this question: "What will prevent the combination of human collective will and artificial intelligence?"

He found an answer to this problem - the low data rate of human beings.

If AI speaks at 1Mb/s and humans can only respond at 1bit/s, the scene would be as absurd as talking to a tree. We need to biologically allow humans to keep up with AI.

Therefore, the more we can increase the input and output rate of humans, the more opportunities we will have in a world filled with AGI.

Musk believes it's possible to increase human output rate by 3 or 6, or even more orders of magnitude, better than what we have now.

The increase in output rate will be achieved by increasing the number of electrodes, the number of channels, and implanting multiple Neuralinks.

This is also Neuralink’s long-term vision, to expand human communication bandwidth and promote the symbiosis of artificial intelligence and humans.

Humans who communicate through Neuralink will have the opportunity to subvert the efficiency of existing language expression, just like people listening to this podcast may use 1.5x or even 2x speed, because 1x speed is too slow. The efficiency of information consumption is much higher than information creation, which is a disadvantage of a low-bandwidth society.

But if we can bridge directly between the brains instead of the mouth, then information can be exchanged like a rar file, achieving a lossless compression process.

To realize this great vision, Neuralink still has a long way to go, and there is still a large gap in human subjects for brain-computer interfaces.

When asked "how quickly the number of human participants will expand," Musk said it would depend in part on the speed of regulatory approval.

They hope to reach their goal of 10 subjects this year, so there are still 8 to go.

Musk believes that in the future, this number will reach hundreds of millions.

The number of electrodes on the chip will increase, and the transmission speed will reach 1Mb/s (perhaps in 5 years), which is faster than anyone can communicate by typing or speaking now.

At that time, humans will achieve "telepathy" through Neuralink.

Interview with DJ Seo

DJ Seo: “Once you penetrate the brain, you enter the arena, so to speak.”

After the conversation with Musk, Fridman also interviewed three Neuralink executives, including co-founder, president and COO DongJin Seo.

Seo received a bachelor's degree in electrical engineering from the California Institute of Technology, and then received a Ph.D. in electrical engineering, computer science, and neuroscience from UC Berkeley. He joined Musk's Neuralink founding team in 2017.

There are two mainstream paths in the field of brain-computer interface (BCI) today - invasive and non-invasive, the main difference being whether the electrodes are surgically implanted below the cerebral cortex. Seo also explained why Neuralink took the invasive BCI approach.

The choice of invasive or non-invasive is ultimately based on the application. I am personally interested in truly understanding and using high-resolution, high-fidelity data to understand the activity at a certain location in the brain.

An analogy may be necessary here, since we are dealing with electrical recordings mediated by charged particles, which is difficult for most people to imagine.

It turns out that a lot of the activity that takes place in the brain, and its frequency bandwidth, is very similar to the audible range of sound waves and normal conversation. Imagine a football game in progress, and you are standing outside the stadium.

You might be able to understand the game and which team is winning based on the cheers and boos of the fans, but you can't get more detailed information like what the score is, what the next play or goal is, or what a particular player or spectator is talking about.

What BCI does is throw microphones into the stadium to get close to the sound source, such as someone talking or a crowded place.

So, the difference between invasive and non-invasive can be understood as where you place the microphone and how you process and use the information.

The information obtained from BCI can give us a glimpse into the mechanisms by which organisms store and calculate information, which include not only electrical signals, but also biological and chemical components, as well as various mechanisms such as vibration, movement or diffusion physics.

What’s even more interesting is that, as physicist Roger Penrose said, there is something “beautiful weirdness” in all these quantum mechanical effects, and consciousness may very well be born from this.

As one of Neuralink's engineering leaders, Seo also responded to the technical safety of brain-computer interfaces - how to ensure that every process is safe?

He said there is an acute safety threat between 0 and 3 months after electrode insertion and at any time beyond 3 months.

The gold standard is to see if there is trauma to the tissue and whether it is associated with any visible behavioral abnormalities. Neuralink has set up a complete pathology department to review pathological slides of relevant specimens, and the FDA also supervises this.

Overall, everything, including surgery, has extremely high standards and is performed in a highly regulated environment, with regulatory agencies reviewing every medical device that comes onto the market.

Talking with Musk

Computing power, data, and big models

Musk：「Play to win, or don't play at all.」

After entering the era of large models, Musk has also frequently exerted his strength in computing power, constantly spending a lot of money to buy chips and build supercomputers. First, he built a 100,000-card H100 cluster for Grok, and then there was.

Dojo D1 supercomputing cluster

In his opinion, not only the amount of computing power reserves is important, but also the rate of increasing training computing power is equally important.

In the podcast, Musk was asked such a question - what factors determine what is a good model, computing power, data, post-training? Or the ability to package products?

His response was that many factors were important.

Just like in an F1 race, how do you answer the question "Which is more important, the car or the driver"?

Anyone familiar with F1 understands that both are important.

Even the best driver will still lose badly if his opponent has only half the horsepower, but if his opponent has twice the horsepower, then a mediocre driver may win.

For AI models, training computing power is similar to the horsepower of a car engine; how to effectively use training computing power and perform efficient reasoning depends on talent. Of course, the amount of data also plays an important role.

Grok’s advantage is that it has real-time access to Twitter data, but in reality, most leading AI companies already crawl all Twitter data.

Even so, we need to realize that the data accumulated by humans so far is very small, adding up to only trillions of tokens. After deduplication and filtering of low-quality information, there is not much left, and the AI ​​model will be exhausted quickly.

But unlike text, Tesla and Optimus have the potential to accumulate huge amounts of data.

Millions of Teslas with cameras, and hundreds of millions (even billions) of Optimus robots will be the largest sources of data.

Especially Optimus, because Tesla can only drive on the road, but Optimus can go anywhere and interact with reality and complete actions.

For example, Optimus can pick up a cup and get feedback to see if it's done the right way, or pour water into a container and see if the water goes into the cup or spills out.

Simple actions like these can be repeated at a billion times scale, generating useful data from reality, as well as cause and effect relationships.

Simple things like picking up a cup (whether it is done in the right way), pouring water (whether the water enters the cup, whether it spills), etc. can be repeated on a billion-fold scale, generating massive amounts of data from reality, as well as causal relationships.

Regarding the accumulation of real data, Musk said: "Reality scales to the scale of reality."

Perhaps the so-called "exhaustion of training data" is a false proposition in Musk's eyes. We don't find enough data because the scale or granularity of viewing reality is still too coarse.

The Five-Step Mantra of Engineering Excellence

Musk: “The most common mistake made by the smartest engineers is optimizing for something that shouldn’t exist.”

Musk, who was an engineer, founded companies such as Tesla and SpaceX. He is now leading the construction of a supercomputing cluster. Many people are curious about how he leads so many engineering teams in different fields and achieves great results again and again in a very short period of time.

Moderator Fridman said that he saw a strong motivation to simplify processes in the supercomputing cluster in Memphis, that is, to continuously improve and iterate after understanding the processes.

Musk agrees: "It's easier said than done to simplify."

As a true believer in first principles, he has a basic "mantra".

First, question the demands. Demands are always stupid to some extent, no matter how smart the person making them is, so start by reducing the number of demands.

It's entirely possible to get the right answer to the wrong question, so try to make the questions themselves as error-free as possible.

Then, the second thing is to try to remove any steps, whether it's a part or a workflow.

This may sound obvious, but people often forget it. If you're not forced to recover at least 10% of your deleted content, you're not deleting enough.

This is a bit counterintuitive because most of the time, people consider success to be when they are not forced to restore deleted content.

This overcorrection is necessary. If you are too conservative and never need to restore any deleted content, it means that there will be a lot of unnecessary things in the system.

For example, if a problem occurs when running an algorithm on a supercomputing cluster, my first reaction is to try to delete it first.

The third thing is to try to simplify or optimize.

These things all sound very simple and obvious, but I have made these mistakes myself more times than I can remember, hence the mantra.

In fact, the most common mistake made by the smartest engineers is to optimize something that shouldn't exist.

The fourth thing is to speed it up, whatever your preset or current speed is, even if you think it's close to the limit, it can be made faster. But don't do this step before trying to delete or optimize.

The final step is automation.

Musk's "Sci-Fi Plot"

Musk: "I often claim that I am an alien, but no one believes me. My green card does say 'Alien Registration Card'."

As we all know, Musk's initial inspiration for the Mars immigration plan came from "The Hitchhiker's Guide to the Galaxy" which he read as a child.

Whether he has seen aliens is a question he is often asked and a topic he enjoys talking about.

This conversation with Lex Fridman was no exception. Musk mentioned Kubrick's films "A Clockwork Orange", "Star Trek", Douglas Adams, Arthur Clarke and "2001: A Space Odyssey".

Musk spent a lot of time talking about "aliens and curiosity" and cited Douglas Adams's point of view:

“Sometimes it’s easy to give the answer, but asking the right question is the hard part. Once you can ask the right question, the answer is just around the corner.”

"SpaceX's goal is to allow life to exist on multiple planets, which takes the Fermi paradox into full consideration."

The Fermi Paradox, coined by physicist Enrico Feimi, represents a conflicting view on the existence of extraterrestrial life. On the one hand, the size and probability of the universe seem to support the argument that intelligent life is common in the universe; on the other hand, there is a complete lack of evidence that intelligent life has ever appeared anywhere other than Earth.

In Musk's view, one of the main obstacles as to why we haven't seen aliens yet is that we are not yet a multi-planet species.

In addition, Musk reiterated his assertion that "declining fertility rates are the root cause of the collapse of civilization" and enthusiastically recommended Will and Ariel Durant's book The Lessons of History.

Trying to alleviate human suffering and expand the capabilities of the human mind through Neuralink, trying to build a colony on Mars, trying to explore the possibilities of artificial intelligence in this world and create billions of robots...

This is what Musk is doing: building the future while inspiring more people to keep building and creating cool stuff, including kids.

As he said at the end of the podcast -

"Go forth and multiply!"

References:

https://x.com/foxshuo/status/1819939215549051029

https://lexfridman.com/elon-musk-and-neuralink-team/

https://lexfridman.com/elon-musk-and-neuralink-team-transcript