The sword points to AI, Musk secretly hides Chen Cang

Author｜Tu Ming, Li Han, Zhu Yue

Editor｜Wang Bo

Source: Jiazi Guangnian

Although many people believe that Musk will form an AI company, no one knows when that day will come.

Just like more than 2,200 years ago, King Yong Zhanghan knew that the Han army was building a plank road, but he didn't know that the Han army would quietly bypass the plank road. As a result, Chencang was attacked and King Yong was defeated.

Musk "entered the AI circle and pointed at the big model" so suddenly. At 0:23 on July 13, Beijing time, Musk announced on Twitter without warning: "xAI was established, in order to understand reality."

Before announcing the establishment of xAI, Musk had been preparing for it for at least 4 months. As early as March this year, Musk revealed that he was forming a "TruthGPT" project with the purpose of "searching for the truth to the greatest extent."

Dramatically, on March 23 this year, Musk also jointly issued an open letter saying: For the sake of human safety, he called on all artificial intelligence laboratories in the world to suspend the training of artificial intelligence systems that are more powerful than GPT-4. 6 months.

It has been less than 4 months since the publication of the open letter, and the "voice of suspension" is still ringing in my ears, but Musk has poached global experts including DeepMind, OpenAI, Google Research, Microsoft Research, and the Department of Computer Science at the University of Toronto. At the foot of the top AI team, an 11-member "all-star founding team" was formed.

**While building plank roads, he advised all major AI companies to "stop doing it, it's too dangerous"; Musk, who didn't follow the routine, blew up the entire AI world. **

1. Double-sided Musk

When Silicon Valley bigwigs talked about the "Tao Te Ching", Musk seemed to understand the "Thirty-Six Strategies".

In the process of acquiring Twitter, Musk used tactics such as "crossing the sky", "taking advantage of the sheep", and "playing hard to catch".

**In the field of AI, Musk has always been the one with the loudest voice in the "AI threat theory". **

Since 2014, Musk has spoken out many times, reminding the society to be vigilant against AI, and calling on the society to strengthen the supervision of AI. According to incomplete statistics from China Business Journal, Musk has made dozens of public statements about "beware of AI" in the past ten years, and even bluntly stated that "artificial intelligence is more dangerous than nuclear weapons."

For example, in 2018, Musk announced his withdrawal from the OpenAI board of directors, citing differences with OpenAI management on AI security issues and commercialization methods. Musk has publicly criticized OpenAI for "not being transparent enough and being too commercial in the pursuit of advanced artificial intelligence", worrying that GPT-4 will spread false information and show bias.

At the beginning of this year, ChatGPT exploded, and AI security issues once again attracted attention. On March 22, the non-profit organization "Future of Life Institute" issued a jointly signed open letter, calling on all artificial intelligence laboratories around the world "for the sake of human safety, it is best to suspend the training of artificial intelligence systems that are more powerful than GPT-4. Suspended for at least 6 months", Musk's name is among them.

In Musk's eyes, the biggest threat to AI is "out of human control." For this reason, most of its suggestions on AI revolve around restricting the development of AI and strengthening the supervision of AI.

**However, Musk is also one of the most determined promoters of AI technology on the planet, and one of the craziest AI pragmatists. **

As early as 2013, Musk proposed the Autopilot project within Tesla, hoping to develop a highway automatic driving system with Google. Since Autopilot recruited its first employee, Drew Gray, in October 2013, Tesla has been on the road to mass-produced autopilot.

It took him less than a year to develop the first-generation hardware of Autopilot, and he was still committed when Google chose to stop the semi-autonomous driving project, and two years later, he opened the Autosteer automatic driving function to the car, pushing Tesla to become the world's largest automaker at that time. The only technology company that has achieved full-stack self-development and self-production in the core field of autonomous driving.

Musk firmly promotes the application of AI in the automotive field, which does not mean that Tesla's self-driving technology is absolutely safe.

Just in February of this year, the US Highway Traffic Safety Administration (NHTSA) forced a recall of more than 362,000 Tesla vehicles equipped with fully automatic driving software on the grounds of safety issues. A month later, the San Francisco Federal Court also filed a lawsuit, arguing that Tesla's concealment of technical risks "created serious risks of accidents and injuries."

This does not affect Musk's love for AI. So far, almost all companies under Musk, such as aerospace company SpaceX, Suncity, neurotechnology research company Neuralink, and Twitter, are using or even relying on AI technology.

On the one hand, it criticizes AI in a high-profile manner and questions its safety; on the other hand, it supports the promotion of AI and applies it to various products. Musk has shown a complete "contradictory sense" on the issue of AI. **The subtext behind this sense of ambivalence may be, "AI is very dangerous. I don't worry about developing it by others, so I have to use it myself." **

Some people also feel this "contradictory sense". Xiaobing CEO Li Di recently accepted an exclusive interview with Xinhua News Agency's "New Science and Technology Innovation" column. When asked "Why Musk called for stopping the research and development of large models", Li Di said: "I asked" Xiaobing at that time, Ma What is the purpose of SK's move. After investigation, Xiaoice found that Musk himself purchased a large number of GPUs, so he came to the conclusion that "maybe to buy himself more time to prepare".

Li Di himself believes that Musk "really thinks that artificial intelligence is dangerous", but due to more reasons, Musk made a statement "calling to stop the research and development of large models",** "People face this new technology. So cramped, at a loss, at a loss, emotions and emotions are very complicated. But there is no doubt that such calls and appeals are useless." Li Di said in the "New Science and Technology Innovation" column. **

After the official announcement of the establishment of xAI, Musk had a conversation about AI safety with two government lawmakers at the Twitter Spaces event. During the conversation, Musk stated that he wanted to "build a set of safer artificial intelligence tools." He also stated: "xAI will not program human morality into an artificial intelligence, but will seek to create an artificial intelligence capable of maximum curiosity and pursuit of truth."

"I am cultivating the humanity of AI, trying to let AI understand the true nature of the universe. This is the way that is beneficial to human beings, and it can make AI understand the importance of human beings," Musk said. Best solution that came to mind.”

In order to accomplish this goal, at least four months ago, Musk was already in full swing to build his own AI company. According to foreign media reports, at the beginning of March, while registering X Holdings Corp and X Corp, xAI Company also registered simultaneously.

In April, Musk said in an interview with the Financial Times that he had purchased 10,000 GPU processors from Nvidia. At the time, Musk’s explanation for his behavior was, “It seems like everyone and their dogs are buying GPUs, and Tesla and Twitter must be buying GPUs too,” and the industry has speculated since then that Musk is brewing his own artificial intelligence technology. Smart mockup.

Also in April, in a TV program, Musk further disclosed that he was preparing for the TruthGPT project, and publicly expressed his concerns about the "bias" of existing artificial intelligence. He said that the newly established AI project was aimed at fighting against OpenAI's ChatGPT. Looking at it now, the TruthGPT project at that time may be the predecessor of today's xAI.

In terms of financing, there is no doubt about the ability of the world's richest man to find money. Musk has held discussions with some investors in SpaceX and Tesla, inviting investors to inject capital into xAI. "There's a lot of people investing ... it's real and they're excited about it," the source said.

The information about Musk's poaching peers was first reported by foreign media. Since March this year, Musk has been persuading researchers from Alphabet's artificial intelligence research department DeepMind to join xAI. Currently, Igor Babuschkin and Manuel Kroiss, who participated in the research and development of AlphaStar, AlphaCode and other models, have become members of the founding team of xAI.

The establishment of xAI externally aims to challenge all AI technology companies including OpenAI and Claude, and internally aims to further integrate resources under Musk's "X brand". "Brand X" is a "collection" of technology companies that Musk is trying to build recently, aiming to make "X" a universal application. In March of this year, Musk established X Holdings Crop in Nevada, USA, to manage all companies under the framework of the "X brand", such as Twitter, whose business name has been updated to X Crop. According to industry analysts, xAI may use X Crop (Twitter) data for training and use Tesla's computing resources.

Behind xAI's sudden official announcement, there is a whole set of designs by Musk.

2. Digging out 11 big bosses

In the team list announced this time, the 11 members other than Musk are from top technology companies such as DeepMind, OpenAI, Google, Microsoft, and Tesla, and they are all well-known "insiders" in the AI industry. Relying on his own company (Tesla)'s ruthless digging, Musk just dug an "AI team".

Young faces, excellent resume, this team is impressive.

According to xAI's official website, team members have contributed many innovative technologies in the AI field, such as Adam Optimizer, batch normalization, layer normalization, and adversarial sample discovery. These tools and methods have been widely used in the AI industry. Among them, the Adam Optimizer developed by Jimmy Ba is currently one of the preferred algorithms for training deep learning models.

In addition, these members have also led or participated in the research and development of important models in the AI field, such as AlphaStar, AlphaCode, Inception, Minerva, GPT-3.5, and GPT-4. Each model is a legend in the AI world.

Nvidia AI researcher Linxi Jim Fan is very familiar with these "legends". He posted on the workplace social networking site LinkedIn, "The team is an 'all-star founding team', and its high talent density has left a deep impression on me. impression—they have published too many papers to count.”

**It is worth noting that in this team, the proportion of Chinese scientists reached one-third, they are Tony Wu, Greg Yang, Guodong Zhang and Zihang Dai ). **Among them, Young is also the co-founder of xAI.

Greg Yang was born in Hunan. He studied in the Department of Mathematics of Harvard University for both his undergraduate and master's degree, and studied under the famous Chinese mathematician Academician Qiu Chengtong. As an undergraduate, he was nominated for the Hoops Award (for best undergraduate paper at Harvard) and the Morgan Award (given to the most outstanding undergraduate research performer in the Department of Mathematics). After work, he joined Microsoft Research as a senior researcher at Microsoft.

Tony Wu is a Ph.D. from the University of Toronto and a postdoctoral fellow at Stanford University. He is dedicated to machine learning, machine reasoning, and formal reasoning. Previously, he worked at Google N2Formal led by Christian Szegedy. His main research interest is to create "machine mathematicians" that can derive mathematical theorems autonomously.

Guodong Zhang graduated from Zhejiang University with a bachelor's degree and a master's degree from the University of Toronto. He won the 2020 Borealis Artificial Intelligence Scholarship. Mainly engaged in research in the field of machine learning and artificial intelligence, focusing on the training, tuning and alignment of large language models. Worked at the University of Toronto and the Vector Institute. His research on the theoretical basis of machine learning and practical algorithms has been adopted by AI companies such as DeepMind, OpenAI, and Google Brain.

Zihang Dai Graduated from Tsinghua University with a bachelor's degree and a Ph.D. from the School of Computer Science, Carnegie Mellon University. Joined Google in 2020 as a Google Brain Research Scientist, mainly engaged in machine learning and neural network research.

In order to dig this initial team, Musk took great pains. For example, in March this year, Musk revealed that he was recruiting former Google DeepMind members Igor Babuschkin and Manuel Kroiss. At that time, he provided the positions of Twitter Senior Engineering Director and Software Engineering Senior Director. In the list of xAI team members, the two are impressively listed.

Igor Babuschkin has worked at DeepMind twice. Serve as a senior engineer from 2017 to 2022, start deep learning research, and participate in the AlphaStar project as a core member-this model once defeated the human champion in the StarCraft e-sports competition. Then, Babuschkin joined OpenAi as a technician, and returned to DeepMind a year later as a senior research engineer.

Manuel Kroiss As a software engineer, he has worked in technology companies such as Google and Deepmind, and has made outstanding contributions in the fields of reinforcement learning and artificial intelligence.

Also from the Google Department are Toby Pohlen and Christian Szegedy. Pohlen, he graduated from RWTH Aachen University in Germany with a major in computer science. He is a former Google research engineer and has participated in projects such as AlphaStar League and Ape-XDQfD. He has extensive experience in machine learning and reinforcement learning.

Christian Szegedy is a PhD in Mathematics from the University of Bonn and worked as a Research Scientist at the Cadence Research Laboratory in Berkeley. Since joining Google in 2010, he has devoted himself to research in the fields of deep learning, artificial intelligence, and computer vision for 13 years, and is also the oldest member of the start-up team.

In addition to the above-mentioned members, the other two core members have also made remarkable achievements in the fields of deep learning and artificial intelligence.

For example, Jimmy Ba, who is both an assistant professor of computer science at the University of Toronto and the chairman of CIFAR-AI in Canada, is one of the most well-known members of the xAI team. In 2014, he and Diederik Kingma co-published the most influential deep learning series of papers in the field of AI - "Adam: A Method for Stochastic Optimization". According to data from the academic social networking site ResearchGate, this paper has the most citations in the field of AI, with 95,460 times.

Kyle Kosic is a full-stack website reliability engineer and data scientist. He has worked in technology companies such as OpenAI. He has rich experience in mathematical research and has a strong interest in automation, scalability and distributed computing research. He has extensive experience in machine learning and data analysis.

Dan Hendrycks, the director of the American Artificial Intelligence Security Center, was also invited by Musk to serve as the X.AI security consultant.

Notably, the organization is committed to "reducing societal risks associated with artificial intelligence." Earlier, Dan Hendrycks published an open letter to global leaders, saying that artificial intelligence is comparable to "epidemic and nuclear war" technology, or poses an existential threat to mankind.

Where will xAI lead with the addition of Dan Hendrycks? What kind of AI development concept will he collide with Musk? This can't help but make people look forward to it.

Currently, the recruitment of xAI is also in full swing. It is reported that the xAI team is actively recruiting experienced engineers and researchers in the San Francisco Bay Area.

**3. What does xAI want to do? **

Looking for money, shaking people, bombarding OpenAI, Musk founded xAI in a flash, what exactly does he want to do?

There is only one sentence on xAI's official website, which is "Understand the Universe". Compared with OpenAI's goal - "to build a safe and beneficial AGI" - the goal of xAI is obviously much more abstract.

To further explain this goal, within an hour of Musk’s official announcement of xAI, xAI posted its first tweet, “What is the most basic unanswered question?”

Musk went on to echo in the comments, "What is the most basic unknown question? Once you know what to ask, the answer is often easy, as my idol Douglas Adams said," and attached A screenshot of xAI’s official website introduction, marked the company’s founding date in red, and wrote next to it, “7+12+23=42”, July 12 obviously has a deep meaning, “42 is the ultimate question of life, the universe and all things. Answer".

In the scientific world, 42 is a meme. This comes from the famous science fiction novel "The Hitchhiker's Guide to the Galaxy" by British writer Douglas Adams, which was once the favorite book of young Musk.

In the novel, a group of super-intelligent mice built a supercomputer named "Deep Thought" in order to find out the answer to the ultimate question of the universe. After 7.5 million years of calculations, the computer came up with a single number—42. Interestingly, in order to analyze the ultimate meaning of 42, the mouse civilization spent another 10 million years to calculate. As a result, when the answer was only 5 minutes away, the computer responsible for the calculation task was blown up. Efforts and ultimate answers vanished.

Since then, 42 has had an unusual place in geek and science circles.

This is not the first time Musk has expressed his love for the number "42". The previously exposed Tesla mysterious project is called "Project 42". SpaceX's Starlink system will have a total of 42 engines. Musk was 42 years old when Tesla entered China.

Regarding the relationship between 42 and xAI, under Musk’s Twitter, some netizens analyzed that 42 implies a mathematical meaning, and Musk may solve the problem of AI for Math from the first principle. The attitude of xAI's Lianchuang Yang also confirmed this point of view. He said on Twitter, "Math for Al and Al for math!"

Most of Musk's fear of AI comes from the fear that it will be out of human control, which is closely related to whether AI can give birth to consciousness. In xAI’s reply to tweets, some netizens asked Musk if he believed that AI has consciousness. Musk did not answer directly, but mentioned cells, quarks, leptons and aliens, and even sent out “the meaning of life.” what?" question.

Unknown and vague, one can only speculate as to what the purpose of xAI is. Before the online question-and-answer meeting, a deliberately created mysterious atmosphere was infinitely exaggerated and magnified.

Right now, regarding xAI, the relatively certain topic is "how to build AI security". At the Twitter Spaces event held in the early hours of July 13th, Beijing time, Musk had a conversation with two government lawmakers on AI safety, in which he described plans to build safer artificial intelligence. Musk mentioned that xAI will not use rules to frame AI in the existing way, but will seek to create an artificial intelligence that can satisfy curiosity and pursue truth to the greatest extent.

The development goal of xAI is also determined. Musk has publicly stated many times that he regards OpenAI and DeepMind as competitors, and words such as "subverting the AI industry" also frequently appear in Musk's mouth.

Despite great ambitions, at present, there is still a gap between xAI and OpenAI and DeepMind. Musk called these two companies "two gorillas in AI" and said, "xAI is still just the beginning , it will take some time to reach the scale of competitors.”

On the AGI track, the competition among the various competitors is extremely fierce, and the speed of technological iteration is measured in days.

Just the day before the official announcement of xAI, Anthropic officially released the new Claude2, which showed great improvement in code, mathematics and reasoning. Anthropic, along with OpenAI and DeepMind, are currently ranked among the top three companies in the field of cutting-edge AI models, and are the only start-up company among them that is not deeply bound with major manufacturers.

The co-founder and CEO of Anthropic, Dario Amodei, served as OpenAI's vice president of research and security. It was because of differences with OpenAI on large-scale model security issues that he chose to leave and founded Anthropic. Claude2 follows 3 Hs in terms of model alignment: honest (honesty), helpful (useful), and harmless (harmless).

As xAI's biggest competitor, OpenAI, its co-founder and CEO Sam Altman responded to the open letter in April this year. He said that OpenAI has not yet started training GPT-5, and it will not start for some time. Compared with continuing to expand the scale of the large model, OpenAI chose to focus on the functions of GPT-4.

Shortly after the release of GPT-4, OpenAI launched the Plugin function to help ChatGPT access third-party services. In June, OpenAI released a series of API updates, including more powerful AI models, new function call capabilities, and longer context processing capabilities. On July 6, OpenAI stated that the GPT-4 API will be fully open for use, and all paid API users can directly access it.

In addition to exploring APIs and plug-ins, OpenAI's ambitions don't stop there. OpenAI is preparing to release a model store (marketplace) similar to the App Store. Enterprises can sell AI models customized according to their own needs to other companies, and formally lay out large models in vertical fields on the basis of general large models.

"Wired" magazine pointed out that at this stage xAI seems to lack the cloud computing capabilities that match OpenAI, Microsoft and Google. Compared with the deployment of hundreds of people in artificial intelligence projects by large companies, xAI's research team of only 11 people is not impressive, and only xAI's security consultant Dan Hendrycks has work experience in artificial intelligence risks.

Even so, Musk still brings his own chips. The official website of xAI states that it is currently a company independent of X Corp, but will work closely with X (Twitter), Tesla and other companies, and may use Twitter content as data to train its language model and use Tesla Computing resources, such as Tesla's self-developed supercomputer Dojo, Musk has said that the computer will be opened to other companies as infrastructure.

On the xAI official website, Musk and his team stated that they will hold an online question-and-answer meeting in Spaces, a voice chat room on Twitter, this Friday (July 14) to focus on responding to public questions. The specific time has not been announced. With the continuous disclosure of xAI-related information, the results of Musk's "darkness" this time will soon be made public.