Original Source: Blue Word Project

Image source: Generated by Unbounded AI

On July 13, Elon Musk officially announced the establishment of xAI, an artificial intelligence company, and will personally lead a core team of 11 people.

The most conspicuous backbone is Igor Babuschkin, who was recruited more than three months ago. When he was working at DeepMind, he was one of the backbones of the AlphaStar project in which the AI model of StarCraft defeated the human world champion. ; was a member of the ChatGPT project when employed by OpenAI.

In addition to Babushkin, a former employee of both DeepMind and OpenAI, three of xAI's current teams are from DeepMind, three from Google's research team, and one from OpenAI.

This may be the most appropriate arrangement of fate: if it weren't for Musk's special style, both DeepMind and OpenAI, the top two heavyweight players in the AI world, would have the opportunity to have the surname Ma, and the world's richest man would not have to open another trumpet to win the AGI (General Artificial Intelligence) Smart).

01 Fall out with the founder of Google

Since the launch of ChatGPT at the end of November 2022, Musk has only said one good thing in public: "AI is really getting better and better these days."

Since then, on a weekly basis, he has repeatedly thrown out the view that "AI is the greatest danger to mankind" and the accusation that "OpenAI is too wrong".

There is a joke in the market: Musk is destined to be against AI, and he will offend whoever he deals with with AI companies. He often makes friends the earliest, but misses every time.

DeepMind was acquired by Google, and the opportunity also happened on Musk’s private jet, but it buried the root of the breakdown of his friendship with Google founder Larry Page.

DeepMind's foundation and foundation is to start from the direction of reinforcement learning, which is easier to implement, to improve the efficiency of AI playing video games.

The AI products developed by DeepMind in the early years can outperform human players in playing "Breakout" (Breakout) and "Space Invasion".

In mid-2013, DeepMind sent the demonstration video of AI playing "Pinball Breaking Bricks" to all angel round investors - to hand over the performance and raise additional money to the beneficiary fathers. Those who received the video included Paypal co-founder Luke Nosek and Musk.

When Nosek was chatting with the owner of the DeepMind video on Musk's private jet, Page, who happened to be present, also heard it, and had the idea of acquiring DeepMind.

Page is not someone who wants to buy a company by listening to gossip. When he was still in charge of Google, the policy of aggressively entering the AI industry had been set.

At the end of 2012, Google bought DNNresearch, the company of Geoffrey Hinton, a leading figure in the AI industry. In 2013, Page held a meeting with the then Google executive team on a small island in the South Pacific, and said on the spot that "Let's really go big". According to the participants, everyone was shocked and "taken a step backwards".

At this point, let Page know of DeepMind's achievements, and the acquisition is inevitable. After Page took Alan Eustace, the then head of Google's technology department, and others to London to meet with the DeepMind team, one of the founders of DeepMind, Dennis Hassabis, also brought his colleagues Fly to Google's California headquarters for an interview.

In the continuous interaction, the prospect of acquisition is getting brighter.

Google's AI developers recognized the authenticity of DeepMind's achievements, and reinforcement learning just filled the gap in Google's AI research and development territory at that time.

And Hassabis also found that Google's technical strength and financial resources are irresistible: Google's self-developed Torch software can greatly improve the efficiency of writing AI codes, and once such a company with hundreds of billions of assets turns its attention to only a few Ph.D. Small companies, small companies face the reality that either all employees are poached with high salaries, or the company is completely bought out. Now that the future path options have been locked, it is natural to choose the more favorable one.

In early 2014, Google overwhelmed Facebook's bid to buy DeepMind at a price of 400 million pounds (about 660 million U.S. dollars at the time). The transaction amount ranked among the top ten acquisitions of startups in the world in 2014.

But before that, Musk got involved.

At the end of 2013, Musk made a last-minute offer to acquire DeepMind, but lost out to Google. Musk wanted to steer the lab's research and told colleagues and those at DeepMind that Page was not a reliable person to oversee the creation of advanced AI, according to people familiar with the matter.

Considering the situation at the time, Musk's move was really weird.

In early 2013, sales of the Tesla Model S were sluggish due to frequent system failures and backward sensor design, and Tesla was on the verge of bankruptcy. In the first week of April, Musk had already asked Page for help, and begged his old friend to spend 6 billion US dollars to buy Tesla, of which 5 billion US dollars should be given to the factory immediately, and he only wanted to continue to lead the company for 8 years or to The third-generation car leaves the factory. If it weren’t for Tesla’s $11 million profit in the first quarter of 2013, which eased the bankruptcy crisis, Musk would have been too busy to take care of himself.

** In the first half of the year, I was still begging my old friend to rescue my own property, but in the second half of the year, I stabbed my old friend, saying that his level is not reliable. The style of the richest man in the world is really astonishing. **

On the evening of July 18, 2015, the Musks hosted Page at a banquet held in California. In the middle of the night, people at the banquet chatted about whether AI is terrible or not. Musk insisted that AI is likely to become a "terminator", while Page believed that paranoid doubts about AI will hinder technological progress, or make it worse. The militaries of various countries have become leaders in AI research.

At that time, Musk tried his best to argue with Page, asking his old friend to talk about why he can rest assured that AI will not become a "terminator" and destroy human beings. Page was so angry that he said "you speciesist (specieist)" and left.

Musk is so good at being a man, he really has no friends. In an exclusive interview on Fox TV in April this year, Boss Ma admitted that "Page and I are no longer close friends."

Back then, when Musk went to the Bay Area, he always sent an email to find Page to rent a house. Page also publicly said on TED Talk: "If I am killed by a bus, my property will not be donated to charity organizations, but will be donated to Elon. ."well.

02 Bedding from OpenAI Volume

After DeepMind turned against his old friend, Musk turned against his partner because of OpenAI’s entrepreneurship.

As a well-known entrepreneur, Musk is world-renowned for his perseverance and perseverance, but this character is also another expression of domineering and arrogant.

If the entrepreneur has never been successful, then this temper may not grow. However, Musk’s entrepreneurship has fallen behind and has repeatedly risen. It is inevitable that he will firmly believe in "the only one in the world" and "how can my industry not listen to my command". "Co-founders" are employees, not partners, subordinates, not peers in the eyes of such people.

In 2015, Musk co-founded OpenAI with Sam Altman (Sam Altman), aiming to fight against Google and DeepMind. In 2018, Musk left OpenAI.

After ChatGPT became popular, the current and former employees of OpenAI are no longer silent, and have released information to the commercial media. In fact, Musk failed to showdown and seize power, and was unanimously rejected by the entire OpenAI team from entrepreneurial partners to ordinary employees.

Some media picked up from the public evidence documents of the legal proceedings involving Musk, and as late as 2017, Musk still regarded OpenAI as a milestone achievement worthy of contribution.

He told his brother Kimball Musk (Kimball Musk) in a text message on his mobile phone: "OpenAI and Neuralink (Note: The neurotechnology and brain-computer interface company founded by Musk) are both crucial to the bright future of mankind. Intuition tells me , and the proportion of time spent on them should be greatly increased.” At that time, the two offices were rented by Musk in the same one in San Francisco.

In addition to the name of OpenAI being taken by Musk and the start-up funds raised by Musk, the current "father of GPT" and chief scientist Ilya Suskever was also dug from DeepMind by Musk.

According to the world's richest man, Satskevier hesitated many times whether to switch to OpenAI. At that time, DeepMind's CEO Hassabis and Musk competed with each other to persuade him. Satskevier is responsible for the story of the GPT series and the DALL-E series of large models.

However, a former employee of OpenAI broke the news, saying that Musk was neither ignorant nor impatient at the time, and set an unrealistic research and development timetable for OpenAI, requiring that a major breakthrough leading to AGI must be made before then, otherwise The reputation of the unit cannot be hanged. Musk was still conducting surveys within OpenAI every day, wondering when employees thought AGI could be made.

There is such a founder among the founders, and the company is already in a bad mood. If this man still initiates coercion against other business partners and demands monopolization of power, the final outcome of this man will not be difficult to imagine.

｜The scene where Musk and his wealthy peers do not make an appointment, but only hold hands and hug, will only happen in AI fictional pictures

At the beginning of 2018, Musk did this: asked other co-founders to take full control and operate OpenAI himself, on the grounds that OpenAI’s research and development progress had fallen behind Google.

Before Musk forced the palace, OpenAI was a non-profit organization with no CEO and no formal management structure. In 2017-2018, there were about 50 employees. Although it is a small company, entrepreneurial partners such as Sam Altman and Greg Brockman unanimously rejected Musk's statement.

Everyone knows the result of forcing the palace. In 2018, Musk withdrew from OpenAI for the public reason of "avoiding conflicts of interest". After that, there was a long attack on OpenAI.

It cannot be said that there is no conflict of interest. After all, when Musk left, he poached Andrej Karpathy, a technical talent, to Tesla to be in charge of the AI and autopilot department. However, the former OpenAI employees broke the news that no one in the unit believed this was the main reason. Musk’s parting speech at OpenAI also repeated this excuse, which caused a bad response among OpenAI employees at the time: the company even counted the bosses as No. 52. Man, who are you fooling?

Ironically, Musk's resignation from OpenAI has become the long-term cause of GPT's launch.

Boss Ma was the financial pillar of OpenAI at the beginning, promising to give OpenAI full access to the $1 billion in funding promised by early backers. This promise has not been canceled even after Musk left, but the actual amount received has been controversial. In 2023, Musk claimed on his Twitter that $100 million was in place, but some sources said it was less than $50 million.

Insufficient money will always be the motivation to stimulate cost reduction and efficiency improvement. OpenAI researchers switched to the Transformer model developed by Google in 2017 as the direction of effort, because the Transformer model has wider applicability, higher parallel computing efficiency, and better GPU acceleration of derivative models. In short, cost-effective than higher.

Even though Boss Ma left OpenAI, his feud with Sam Altman is still hard to get over. According to people familiar with the matter, in early 2023, Musk privately told Altman that he would start a competing AI project. Altman first politely wished Musk all the best, and then Piri poked a sentence: I don’t understand how adding a project can eliminate Musk’s concerns about the rapid development of AI.

Since 2023, OpenAI's unit behavior and Altman's personal words have shown that his old colleagues have been deeply offended by Musk.

In late March, OpenAI board member Shivon Zilis (Shivon Zilis) resigned. The reason is that as a former girlfriend who had a pair of children with Musk and is now a Neuralink executive, Zilis continues to be on the OpenAI board of directors. A conflict of interest will arise.

At almost the same time, in interviews with podcasts and appearances on various occasions, Altman said that Musk is a "jerk" (jerk), "He is my mentor in R&D operations...but he behaves a little better. I don't want to follow suit."

｜The world’s richest man made a recent sampling: "Zuckerberg is a turtle man, I suggest that we two compare who is bigger."

Musk can be said to be the only one in the world who can make Altman, who has always been gentle and quiet in public, so angry. This achievement is not just a feud in the past, the world's richest man has worked hard to increase the rift. In addition to the daily verbal censure of OpenAI, Musk's appearance at the OpenAI headquarters also played a role.

At the end of November 2022, Musk's private reaction to watching ChatGPT's achievements is said to be "furious". In early December, Musk announced that he would tear up the data access agreement between Twitter and OpenAI before his acquisition.

In early 2023, Altman asked Musk to meet at OpenAI's San Francisco headquarters to try to resolve differences. The Musk people are here, but not only him and his bodyguard team, but also his toddler son and his nanny.

When OpenAI employees in the conference room demonstrated how Twitter can be compatible with ChatGPT, Musk and the nanny took turns putting the child on his lap and shaking it. In conversations with more OpenAI employees, Musk repeated his "world as a digital analog" concept before walking away.

Although those present said Musk had a moderate attitude, the demeanor of the world's richest man was closer to a nouveau riche returning home than a cooperative negotiation. If it clears up the grudge, hell.

03 "Top match" new lover, hard to realize old dream

Unable to manage DeepMind and OpenAI, Musk can now finally manage his own unique enterprise xAI. Will he be the first to create AGI? According to the available information, there is hope, but not very great.

The experience of Musk's companies in AI is indeed top-notch in the industry, but the advantages are all decomposed and unilateral. The "Dojo" supercomputer, the D1 training chip, the FSD algorithm, and the "Optimus Prime" humanoid robot, although not number one in the world in terms of performance parameters, are all on par with top competitors.

However, the special point of AGI is that it is "general", that is, it can handle input data of various modalities and unique problems of various natures. This is the characteristic of "general artificial intelligence" that can rival human intelligence. However, Musk's existing advantages are precisely not here.

For example, the architecture of the D1 chip is tailor-made for Tesla's autonomous driving, and when used on a general-purpose multi-modal large model, the memory bandwidth will appear to be limited. The combination of on-chip communication, on-chip memory, and off-chip memory requirements for large models and autonomous driving algorithms varies greatly.

Tesla's algorithm performs well enough to process the input data of visual images, but the effect is limited to translating image data into space vector data, and then making action decisions for the automatic driving system. Transformer-based large language models and multimodal models require very different code architectures.

The "Optimus Prime" humanoid robot can directly follow the action of the FSD algorithm in the video in 2023, because the amount of data and the model of the robot action and the automatic driving system are all focused on the perception and navigation of the surrounding physical environment.

However, due to the technical ceiling, to complete these actions with low energy consumption, high precision, high speed and low latency, the Tesla algorithm can only be applied to models and humanoid robots with extremely limited physical size. In short, the Tesla algorithm cannot drive a truck, nor can it drive a large multi-armed robot.

Even the ideas advertised by xAI are not exclusive. The bosses and backbones of xAI said that the existing large language model is only a high-end differential statistical application and correlation imitation engine, and xAI wants to enable the artificial intelligence model to discover new knowledge and discover facts that people can't understand, rather than just repeating Emulate what is in the training data.

However, the "world model" concept of "Yann Lecun", the soul of Meta's artificial intelligence research department, expressed almost the same opinion three years ago. The I-JEPA model, published in a paper on June 14, has begun to use background knowledge about the world, rather than surrounding pixel data, to fill in the missing parts of the image.

Therefore, although xAI can borrow Tesla's resources and obtain Twitter data, it has definite advantages. However, in terms of completing the vision, the progress of existing competitors has a visible lead. In the arena where the AGI laurels are won, the situation of competing for the throne will not change in the short term.