Chapter 317 Large Model of the Human Brain
When Lin Jia met with car company executives one after another in Haikou, Chen Yuanguang had already gone to Yanjing to make a detailed report.
There is a loud voice within Lao Zhong about open source technology.
The reason is simple. Both external forces and internal private enterprises hope that technology can be open source.
If we simply adopt a cooperative approach, there are great variables in who comes first and who comes last, who can cooperate and who cannot.
Although Robin publicly stated that Baidu and Guangjia Aerospace are strategic partners, he actually has no idea whether they can take the lead in this round of autonomous driving.
Even if you say it conclusively during an interview.
Compared with the black box, which has great uncertainty, private enterprises naturally prefer open source technology, which means that the technology itself is also in their hands.
Bureaucrats who hold this view also have strong theoretical support. Internet technology has always been open source, and Amerikan’s route and source code for artificial intelligence technology have never been hidden.
In addition, Chen Yuanguang believes that technology should be open source, and we should listen to the opinions of professionals. Is there anyone more convincing than Chen Yuanguang, who leads technology research?
This kind of view has always been there, but I don't dare to express my views publicly. At most, I only say a few words within a small circle.
When everyone knew Chen Yuanguang’s public opinions to Yanjing, everyone’s discussion began to get louder.
Also dare to speak out through the South China Morning Post.
Whether it is the South China Morning Post or the United Daily News, it can be seen as an unofficial communication channel, hoping to create public opinion from the outside to achieve its own goals internally. However, as the national strength of the East University increases, this The effect of this method is getting worse and worse.
Lin Jia held a meeting in Haikou and Chen Yuanguang held a meeting in Yanjing.
"Hello everyone, I think you all know something about the robot driver in Shenhai recently. There are various opinions on the Internet.
The robot is called Tiedan, and it is not a Terminator as foreigners say. .
We have been communicating with relevant parties internally from the beginning of the project, from testing on the test road to communicating with the Ministry of Industry and Information Technology, the Ministry of Transport and other departments to obtain their approval.
Including the subsequent official operation as an online car-hailing service on urban sections, we also communicated with relevant departments. Tiedan even obtained what should be the only robot driver's license in the world
This was all done with the traffic management department. Communicate fully.”
Chen Yuanguang said that he wanted to introduce the ins and outs first.
“Then I want to talk about the underlying technology. It is not actually driverless. To be precise, it is a large model of the human brain.
Our internal abbreviation is HBM.
Everyone is familiar with ChatGPT in previous years. This type of model is collectively called LargeLanguageModel, and the Chinese abbreviation is large model
To be more precise, it should be large language model
.
Input language, output language, although its output content subsequently evolved from language text to images, tables and even videos.
The essence of these outputs is still data, which is nothing more than a change from the arrangement and combination of structured data to unstructured data. The input to the large model is also data, which is existing data on the Internet
The HBM we are doing this time is to input human brain waves to train the large model. It uses the machine body as a carrier to ultimately achieve Real World Impact.
Take Tiedan as an example. We have collected more than 100,000 hours of brainwaves from online ride-hailing drivers, and then fed this data to HBM. After self-training, it removes impurities from the data. , starts output through the machine body.
First run on the test site, then run on the test road section, then run in the prescribed area, and finally run without restrictions.
Data is also used to train HBM, but it is not text data, but the human brain.
You can see that the data is abstracted from the adult brain and finally input into HBM.
So driverless driving is only one of its applications, just like large models outputting text is only the first application, and soon they began to output images, videos, and tables.
Similarly, we will have many application scenarios for HBM in the future. Simply using it for autonomous driving is not cost-effective.
Unmanned driving is just a prototype similar to ChatGPT’s earliest appearance. ”
Everyone here looked at each other in confusion. Even the regulatory authorities related to HBM did not know much about the underlying technology. Everyone thought it was just autonomous driving technology.
Looking now, this technology is much more valuable than they thought driverless cars would be.
If it is just driverless driving, each company is not far behind L4.
It is entirely possible to make up for it through hardware.
Like the optical satellite networking that many new energy car companies are working on, the shortcomings of the algorithm are made up from the hardware level, and the gap from L4 is already very small.
Therefore, open-sourcing this technology will not cause much loss to Lao Zhong, and it can also force Amerikan, Europe and Neon, which were not so keen on new energy, into the new energy vehicle track.
Oil trucks have natural flaws in driverless driving.
But if it is like what Chen Yuanguang said, then fundamentally everyone's views need to be reshaped.
"Yuan Guang, I am a complete layman on artificial intelligence technology. Although I have listened to many lectures by experts like you, my professional understanding is still far behind yours.
I I would like to ask, what are its application scenarios? Can you briefly explain it?
The other thing is what it can do in the military field. "
Chen Yuanguang said: "HBM can evolve.
Different types of work have different technical difficulties, such as construction workers and textile industry workers. It is fully capable of such simple mechanical repetitions.
The driver's gold content is slightly higher, and now it seems that it is quite competent.
For jobs such as electricians and fitters that require higher precision, their competence requires not only the evolution of algorithms, but also the evolution of hardware, such as the improvement of the accuracy of cameras responsible for vision and the accuracy of force sensors on fingers. Improve and so on.
Including the improvement of its brain computing power.
This will be an overall improvement.
For me, its biggest use in the short term is as construction workers on the moon, responsible for building lunar bases.
In the medium term, our space station will be dominated by robots, responsible for the repair of space mining equipment and the maintenance of the space station.
In the military, I think it is not cost-effective. The cost of robots is high, the efficiency of performing tasks is not high, and the robustness is seriously insufficient. In my opinion, it is far inferior to mechanical dogs and drones.
A few can be used for emergency rescue work, but the essence is cost. In terms of cost, it is too expensive to replace drivers. "What does robustness mean?" "
"Sorry, this refers to the system's ability to survive under abnormal circumstances, which can be understood as stability.
In short, such precision instruments are not very stable on the battlefield. "Chen Yuanguang said.
"Yuanguang, I originally supported open source technology, but after listening to it, I feel that this technology has huge potential.
My thoughts have been seriously shaken. I think many colleagues here may have similar thoughts to mine.
I hope you can help us all clarify the benefits of open source technology. "
Chen Yuanguang smiled: "This is also the biggest purpose of coming here this time.
Many things cannot be explained clearly through video, and interviews are the best way.
I would like to talk about it first. In the past, artificial intelligence research focused on making machines simulate humans as much as possible.
Computers have advantages in many aspects, the most typical of which is signal transmission speed.
The signal transmission of human neurons is an electrochemical process, and its speed is 100m/s, while the electrical signal transmission speed in silicon-based chips is close to 70% of the speed of light, which is 20 million meters per hour. Seconds, the electrical signal transmission speed of topological semimetals is even more amazing, and can be close to 90% of the speed of light.
The error probability of human neurons in the signal transmission process is one percent, the error of silicon-based chips is one in 4.2 billion, and the error of topological semi-metal chips is even smaller.
The chip has very obvious advantages in the speed and accuracy of information processing.
In the past two years, Intel had a neuromorphic project called Halapoint, which used 1.15 billion digital neurons to simulate the human brain.
Even if so many digital neurons are used, and even though silicon-based chips have natural advantages over human neurons, Intel's Halapoint can still only handle computing problems and does not perform well in neuromorphic computing.
On the contrary, a project called Brainoware that Harvard University did at about the same time performed better at simulating the human brain.
Harvard's project combines human brain cells with silicon-based chips to build new hardware they named Brainoware.
They first used human multifunctional stem cells to cultivate brain organoids, and then part of the entire brainoware used traditional computer hardware and part of it used this brain organoid.
They built a three-layer computing framework, divided into input layer, reservoir layer and output layer, in which brain organoids were used in the reservoir layer.
The organoid receives signals through an input layer, which converts them into signals for electrical stimulation. The brain organoid acts as an adaptive database, mapping these signals to the output layer. In the output layer, neural activity representing the reservoir state is recorded and decoded to provide readouts for applications such as classification, identification and prediction.
Brainoware's physical reservoir properties, including nonlinear dynamics, spatial information processing, and fading memory, were tested by evaluating responses to stimuli of varying pulse times and voltages. The system is then applied to real-world tasks such as speech recognition and nonlinear chaos equation prediction.
In speech recognition tasks, Brainoware needs to identify the speaker's voice in a pool of speakers. A total of 240 isolated audio clips of Japanese vowels pronounced by eight different male speakers were used to train the system.
In the end, they achieved the same results while spending less than 10% of the training time on traditional hardware.
Okay, that’s it for the two examples.
These two examples illustrate that current foreign hardware has natural drawbacks, and the HBM model has very poor adaptability to traditional silicon-based chips.
Of course, I have not yet developed brain organoids that can be commercialized on a large scale to replace silicon-based chips.
But what I can tell you is that if foreign countries want to use the HBM model, then they must buy topological semi-metal chips produced by Dongda, which is equivalent to us being stuck on the hardware side.
Having said that, even if the technology is not open source, considering that to make progress in the HBM model, we need to cooperate with leading domestic technology companies. The more people involved, the greater the risk of technology leakage.
We might as well open source it directly and block the supply from the upstream hardware side.
Sign a technology open source agreement with all participating countries and organizations. All technological progress obtained around the HBM model needs to be open source and cannot be used in the military field.
If you do not comply, then we have an agreement as a basis and can openly refuse to supply topological semi-metal chips.
To put it simply, the technology will be leaked sooner or later, but we have absolute control from the hardware side and open source it to other countries. We can jointly promote the development of HBM technology globally and help Dongda chip companies open up the world. Market gate.
Of course, I think there are many things that can be traded, and we can talk about them slowly. ”
The Brainoware mentioned above is an article published in the electronic sub-journal of Nature in December 23.
Composite machines built by combining biology and machinery have extraordinary potential in the calculation of nonlinear equations and speech recognition.
Perhaps mechanical ascension will be the privilege of a minority group in the future.
After Chen Yuanguang elaborated, the voices supporting open source prevailed.
“I think Yuan Guang said it very well. This is a very valuable bargaining chip for us. We can take it out, but we have to exchange it for something valuable enough.”
< br>“I also support our chip industry’s entry into overseas markets just mentioned by Yuanguang. I think this is the strategy they have to adopt if they want to use HBM technology.
Just like we had to buy NVIDIA graphics cards in the past because there was a lack of alternatives on the market.
And we have to talk about some substantial benefits back. "
"I agree with everyone's views just now. I have a question to ask you. That is, you just mentioned the Harvard research and the device that combines human brain tissue and computer hardware. What you mentioned is that there is currently no such device. Get it out.
I would like to ask, should we take this technical route in the future?
Does this technical route have potential? ”
This question was not only curious to him, but also to many people present.
The combination of machinery and body may have a serious impact on ethics.
“This technical route is very promising, but the cost is too high. There are big problems in terms of the generation and maintenance of organoids, the power consumption of the entire device, and the efficiency management of the data level.
We can sponsor some pre-research projects through the Natural Fund, but there is no need to be a pioneer on this technical route," said Chen Yuanguang.
By the end of the meeting, everyone's opinions gradually became unified.
"Yuan Guang, if the HBM model can only use topological semi-metal chips, then I think open source technology is completely negotiable.
Next, specific companies have been arranged to find experts for verification and verification of authenticity. , write a detailed report and submit it.
This is not to distrust you, but this matter is related to a very important work direction for us, and we must do things rigorously from a work perspective. ”
(End of this chapter)