I believe that readers will be greatly shocked by the internal struggle for power within OpenAI a few days ago. OpenAI's "coup" event allowed the world to see for the first time the fierce battle between the technology giants who will determine the future of artificial intelligence development.
But the struggle for rights is by no means the first time in the more than ten years of AI development in Silicon Valley. The New York Times interviewed more than 80 executives, scientists and entrepreneurs for this purpose, telling the story of the ambitions and money of Silicon Valley technology tycoons under the AI wave in the past ten years.
The story begins with the establishment of DeepMind.
2010: DeepMind is born
This year, 34-year-old neuroscientist and later DeepMind founder Demis Hassabis (Hassabis) and two colleagues living in the UK were looking for funds to build "general artificial intelligence" (AGI).
But at the time, few people were interested in artificial intelligence. Because after half a century of research since the Dartmouth Conference, the field of AI has failed to develop anything close to human intelligence.
One day, Hassabis was introduced by a friend to a cocktail party held by Peter Thiel at his San Francisco townhouse. (Peter Thiel is a well-known investor in Silicon Valley who accumulated huge wealth through his early investment in Facebook and his early cooperation with Musk at PayPal.)
Peter Thiel loved chess and had a chess board in his living room. Hassabis had been preparing for the meet for a year and believed chess would be his unique entry point. Because when Hassabis was 13 years old, he was the second-ranked chess player in the world in the under-14 age group.
Large model research test portal
GPT-4 portal (no wall, can be tested directly, if you encounter browser warning points, just advance/continue access):
http://hujiaoai.cn
Hassabis believes many people at the party will try to solicit investment from Peter Thiel. His strategy was to arrange another meeting, and he and Peter Thiel chatted about chess and sought a separate meeting. Not long after, Peter Thiel agreed to invest £1.4 million into their start-up.
Hassabis named the companyDeepMind, a nod to "deep learning," neuroscience, and the Deep Thought supercomputer from the science fiction novel "The Hitchhiker's Guide to the Galaxy."
By the fall of 2010, DeepMind began building their dream machine. They firmly believe that because they understand the risks of AI, they are uniquely qualified to protect the world from being destroyed by AI.
Mustafa Suleyman, one of the three founders of DeepMind, said: "The benefits brought by AI technology are huge. Our goal is not to eliminate or suspend the development of AI, but to mitigate the negative impacts of AI."
After winning the support of Peter Thiel, Hassabis entered Musk's sights.
About two years later, they met at a conference organized by Peter Thiel's investment fund, which also invested in Musk's SpaceX. Hassabis was given a tour of SpaceX headquarters. Afterwards, Hassabis and Musk ate in the canteen under the ceiling with rocket shells hanging, and the two chatted very speculatively.
Musk said that his SpaceX plan is to avoid overpopulation on the earth and doomsday moments such as nuclear war and asteroids hitting the earth, and to preserve the fire of human civilization.
Hassabis reminded Musk that "AI" should be added to the doomsday list - the intelligence of machines may surpass humans in the future, and may even want to eliminate humans.
Musk was speechless in surprise. He had never thought of this danger before. Musk quickly invested $5 million in DeepMind to get a closer look at the technology's development.
After DeepMind received funding, it hired researchers who specialized in neural networks, essentially giant mathematical systems that at the time took days, weeks, and even months to identify patterns in large amounts of digital data. First developed in the 1950s, these systems can learn to handle tasks on their own. For example, they can read handwritten text after analyzing names and addresses scrawled on hundreds or thousands of envelopes.
DeepMind took this concept a step further by building a system that could learn to play classic Atari games like Space Invaders, Pong, and Arkanoid to demonstrate the possibilities of AI.
DeepMind soon caught the attention of another Silicon Valley giant, Google, specifically Google co-founder Larry Page, when he saw a demonstration of DeepMind's machine playing an Atari game. He wants to "join the AI game." But in fact, Musk also discussed the investment with his good friend Larry Page shortly after the meeting with Hassabis, which paved the way for the subsequent breakdown of the relationship between the two.
2012~2013: Google attaches great importance to AI development and "wins" Hinton
2012 was the trough of artificial intelligence. Hinton, later winner of the Turing Award and known as the father of deep learning, studied neural network concepts that were not widely accepted by AI researchers of the same period. This caused them to face a huge challenge. pressure, but they did not shrink from it, but instead strengthened their beliefs.
In 2012, the ImageNet competition provided them with the opportunity to prove themselves - to see who could design the best image recognition algorithm and successfully identify more images.
Hinton took his two students Alex Krizhevsky and Ilya Sutskever, who later became the chief scientist of OpenAI, to invent AlexNet in this competition. They abandoned the traditional hand-designed solution, but used a deep neural network and ran it on the GPU. Train them.
AlexNet broke through 75% accuracy, far surpassing all competitors, demonstrating the great potential of deep learning to the world, and marking the beginning of the deep learning revolution.
Scientists were surprised by Hinton and his students' results in the ImageNet competition, which attracted the attention of Yu Kai, an AI scientist at Baidu at the time. Baidu offered Hinton and his students $12 million to join the company in Beijing, according to three people familiar with the matter.
Hinton did not agree directly. He has spent most of his career in academia and is not particularly motivated by money. Hinton said:“We never knew we were so valuable.”
Hinton and two students consulted with lawyers and experts on acquisitions and came up with a plan: "Establish a company and then organize an auction to sell the company." The auction was held during the NIPS conference at Harrah's in Lake Tahoe. hotel.
Google founder and CEO Larry Page saw deep learning technology in Google Brain and believed Hinton's research was valuable. He gave Google Senior Vice President of Engineering Alan Eustace a blank check to hire any AI experts he needed.
Eustace and Jeff Dean, who was leading Google Brain at the time, flew to Lake Tahoe and treated Hinton and his students to dinner at the hotel's steakhouse the night before the auction to lobby them to join Google.
The next day, Hinton conducted the auction in his hotel room. He rarely sits due to an old back injury. He flipped a trash can upside down on the table, placed his laptop on top, and watched the bids come in over the next two days.
Baidu, Google, Microsoft, and DeepMind made bids respectively. As the price rose, DeepMind quickly withdrew from the competition. The price subsequently rose to $20 million, and then to $25 million, according to documents detailing the auction. Microsoft withdrew when the price exceeded $30 million, but rejoined the bidding at $37 million.
Hinton recalled:“We felt like we were in a movie.”
Then Microsoft withdrew for the second time, leaving only Baidu and Google. They raised their bids to US$42 million and US$43 million. Finally, at $44 million, Hinton and his students stopped the auction. The bids were still rising, but Hinton was determined to join Google.
2013~2014: Google acquires DeepMind
Google's seizure of Hinton is a sign that deep-pocketed tech giants are determined to acquire the most talented AI talent. DeepMind’s Hassabis also noticed this. He has been telling employees that DeepMind will remain an independent company, which he believes is the best way to ensure its technology doesn't turn into something dangerous.
But as the tech giants entered the race for talent, Hassabis decided he had no choice: It was time to sell DeepMind.
By the end of 2012, both Google and Facebook were considering acquiring DeepMind.
Hassabis and his co-founders insisted on two conditions:
-
DeepMind technology cannot be used for military purposes;
-
AGI technology must be overseen by independent committees composed of technical experts and ethicists.
Google offered $650 million, and Facebook's Zuckerberg made a higher offer to DeepMind's founder, but he disagreed with Hassabis' terms.
DeepMind ultimately chose Google.
2013: Zuckerberg failed to acquire DeepMind and established FAIR Labs
After the failure to acquire DeepMind, Zuckerberg decided to build his own AI laboratory.
He hired Yann LeCun, a French computer scientist who later won the Turing Award with Hinton. LeCun had made pioneering contributions to research on convolutional neural networks and other areas.
Hinton A year after the auction, Zuckerberg and LeCun flew to Taihao Lake to attend the same AI conference, Zuckerberg walked around the suite at Harrah's Casino in socks, and he personally interviewed some of the top researchers personnel, who quickly received millions of dollars in salary and stock.
AI was once derided, but now the richest people in Silicon Valley are spending billions to avoid being left behind.
2015: Musk and Page debate AI at birthday party
In July of this year, in order to celebrate his 44th birthday, Musk and his wife held a three-day party in a wine country resort full of cabins in California. They only invited family and friends to attend, including There is Larry Page, the founder of Google mentioned above.
After dinner on the first day, Musk and Page sat by the fire pit by the pool, and the two friends of more than a decade debated whether AI would ultimately benefit or destroy humanity.
As the night went on, the argument between the two became more and more intense. Page said that humans will eventually merge with AI machines. In the future, there will be multiple intelligences competing for resources, and the best will win.
But Musk says if that happens, we're screwed and machines will destroy humanity.
With a voice of frustration, Page insisted on pursuing his utopia, and in the end, he called Musk a "speciesist," someone who favors humans over future digital life forms.
Musk later said the insult was "the final straw in their relationship."
As the night dispersed, many in the party crowd seemed stunned, even a little amused, thinking it was just another debate that often breaks out at Silicon Valley parties.
Musk and Page
Eight years later, the argument between the two seems prescient. The question of whether AI will benefit the world or destroy it (or at least cause serious harm) has fueled an ongoing debate among Silicon Valley tech founders, the public at large, academics, lawmakers and regulators over whether AI technology should be subject to regulatory controls. .
The dispute pits some of the world's richest people against each other: Musk, Page, Zuckerberg, technology investor Peter Thiel, Microsoft CEO Satya Nadella and the most talked about person in the past month. OpenAI CEO Altman.
All are vying for a share of future AGI that could be worth trillions of dollars and the power to shape it.
At the heart of this competition is a puzzling paradox—those who say they are most worried about AI are the ones most motivated to create them and enjoy the wealth they bring. Each of them believes that only they can keep AI at the forefront and avoid endangering the security of humanity and the planet.
2015: Musk out of DeepMind
When Musk invested in DeepMind, he had already broken his unofficial rule of "not investing in any company that he didn't run himself."
The shortcomings of Mr. Musk's investment decision were already apparent when, just about a month after his birthday spat with Mr. Page, he was once again face to face with his former friend and fellow billionaire.
As mentioned above, Musk told Page when he invested in DeepMind, but Page acquired DeepMind without saying hello. Musk tried every means to disrupt the acquisition, but failed. The result of the game was just Get Google to agree to set up an "ethical safety committee" on AI, of which Musk is a member.
Musk hopes to lead DeepMind’s AI research and development direction in this way to ensure that AI does not surpass humans. But in fact, this committee is just a show-off by Page, and Musk has no substantive power over DeepMind. On August 14, 2015, the DeepMind Ethics Committee held its first meeting. Members met in a conference room outside Musk’s office at SpaceX, with a window that could see Musk’s rocket factory, according to three people familiar with the matter.
Participants included three Google executives: Page, Sergey Brin and Google Chairman Eric Schmidt, as well as PayPal's other founder Reid Hoffman and Australian philosopher Toby Ord who studies "existential risk".
But the meeting was the end of Musk's control of DeepMind. Google acquired the entire DeepMind company, and Musk was out. Although he made money, he was not happy.
DeepMind reported that they are advancing AI research and development work, but realize that AI technology has serious risks. DeepMind co-founder Suleyman gave a speech titled "The Pitchforkers Are Coming" (meaning the AI revolution and risks).
Suleyman told the board that AI could lead to a proliferation of disinformation. DeepMind worries that as AI technology displaces countless jobs in the coming years, the public will blame Google for taking away their livelihoods. DeepMind believes that Google needs to share its wealth with the millions of people who can no longer find work and provide a "universal basic income."
Musk agrees. But it's clear that Google's bigwigs have no intention of redistributing wealth. Google Chairman Schmidt said these AI concerns are grossly exaggerated. Page, in his characteristic whisper, agreed, arguing that AI will create far more jobs than it replaces.
Eight months later, DeepMind achieved a breakthrough that shocked the world - AlphaGo came out, defeating the world's top humans Go player Lee Sedol.
The game, broadcast live over the Internet, attracted 200 million people around the world. However, most researchers believe that it will take another 10 years before AI has the creativity to achieve this feat.
Rationalists, effective altruists, and others worried about the risks of AI claim that AI’s triumph confirms their fears.
An AI security researcher at DeepMind wrote in a blog: "This is yet another indication that AI is developing faster than many experts expected."
DeepMind's founders are increasingly concerned that Google could misuse their invention. In 2017, they tried to break away from Google. But Google responded by increasing the salaries and stock award packages of DeepMind’s founders and their employees, and DeepMind chose to stay. But the ethics committee never held a second meeting.
2015: Dissatisfied with DeepMind’s exit, Musk took the lead in creating OpenAI
Musk was convinced that Page's optimistic view of AI was completely wrong and was angry about losing DeepMind, so he wanted to start building his own AI lab.
Also in 2015, Musk had dinner with Sam Altman, then the head of the famous technology incubator Y Combinator, and several researchers in a private room at a hotel in California.
Sam Altman
Musk believes that the safe development of AI must be beyond the control of individuals or companies, and this can only be achieved by the competition of a large number of independent AIs. The best way to fight against the current powerful Google and DeepMind is to establish a company with open source technology. It will be responsible for generating the most advanced technology and open source it to all potential competitors, so that hundreds of companies that are destined to emerge in the future will defeat Google. and DeepMind.
Everyone suggested naming the company after Musk’s open source conceptOpenAI. OpenAI was officially established at the end of 2015, with Musk, Hoffman and Thiel and others pledged to invest $1 billion in OpenAI, and Altman, who was only 30 years old at the time, was responsible for operating OpenAI.
In order to be at the forefront of OpenAI research, Musk deliberately poached Ilya Sutskever from Google. Yes, it was the research scientist Google "acquired" in the Hinton auction mentioned earlier, and the man who later became the chief scientist of OpenAI. .
Initially, Musk wanted to run OpenAI as a nonprofit as a way to move away from the monetary incentives driven by Google and other companies. But when Google and DeepMind shocked the technology world with "AlphaGo", Musk began to change his views on the way OpenAI operates. He desperately wanted OpenAI to invent something that would capture the world's attention and close the gap with Google, but as a nonprofit, it fell short of the task.
In 2018, Musk quit OpenAI
According to people familiar with the matter, in the second half of 2017, Musk hatched a plan to seize control of OpenAI from Sam Altman and other founders, turn it into a commercial operation, and then cooperate with Tesla. and relies on supercomputers being developed by Tesla.
When Altman and others objected, Musk backed out and said he would focus on Tesla's AI efforts. Musk announced his departure to employees in February 2018 during a meeting on the top floor of OpenAI’s offices. When Musk said OpenAI needed to speed up its progress, a researcher at the meeting retorted that Musk was being reckless.
Musk called the researcher a "stupid" and left angrily with abundant financial resources. Naturally, the US$1 billion originally promised to invest in OpenAI will not be fulfilled.
Musk and Altman no longer have friendly relations.
From 2019 to 2021, OpenAI "embraces" Microsoft, and Amodei establishes Anthropic
OpenAI was suddenly in desperate need of new funding. Altman flew to Sun Valley for a conference and met Microsoft CEO Nadella, and collaboration seemed natural. Altman knows Microsoft CTO Kevin Scott. Microsoft acquired LinkedIn from OpenAI board member Hoffman. Nadella told Scott to get the deal done, and it was done in 2019.
OpenAI, which formed a for-profit company under its original nonprofit umbrella, received $1 billion in new funding, and Microsoft has a new way to build AI into its massive cloud computing service.
But not everyone inside OpenAI is happy.
When OpenAI was founded, Dario Amodei, a researcher associated with the Efficient Altruism community, was unhappy with the Microsoft deal because he believed it steered OpenAI toward commercialization rather than AI safety. He and other researchers went to the board to try to oust Altman, according to five people familiar with the matter. After that failed, they left.
In 2021, Dario Amodei and his sister led a team of about 15 engineers and scientists to create Anthropic, planning to develop AI in a safe and controllable way.
Dario Amodei and his sister
"The co-founders of Anthropic are not trying to oust Sam Altman from OpenAI," an Anthropic spokesperson later said. "The co-founders themselves came to the conclusion that they wanted to leave OpenAI to start their own company and contribute to OpenAI." 's leadership communicated this and within weeks negotiated an exit on mutually agreeable terms."
Two years later, Anthropic received $4 billion from Amazon and $2 billion from Google.
2018-2021, development of GPT-1~3 series models
Led by Ilya et al., OpenAI created the original GPT model in 2018. A year later, they released GPT-2. This new model showed the great potential of large language models, but it also caused dissatisfaction in the AI community because OpenAI is no longer a non-profit entity, and open source works are becoming more and more popular. few.
Yet despite the controversy, OpenAI has achieved significant research results.
The following is a summary of the iterations of the GPT model, each representing a major advance in the field of natural language processing:
GPT-1 (2018):This is the first model in the series. One of its key innovations is the use of an unsupervised pre-training method. After training on a large-scale Internet text data set, it learns to predict words in a sentence based on the previous word context, enabling the model to deeply understand the language structure and generate similar words. Text for humans.
GPT-2 (2019):Builds on the success of GPT-1, training on a larger data set to produce a more powerful model. A major advance of GPT-2 is its ability to generate coherent and fluent paragraph text on a variety of topics, making it a key player in unsupervised language understanding and generation tasks.
GPT-3 (2020):A significant breakthrough in both scale and performance. GPT-3 has 175 billion parameters that shocked the world at the time. It has achieved state-of-the-art performance on many language tasks and is comparable to human levels in capabilities such as question answering, machine translation, and summary generation. Additionally, it demonstrates the ability to perform simple coding tasks, write coherent news articles, and even poetry.
GPT architecture
In 2022, ChatGPT will come out amazingly
In 2020, when OpenAI released GPT-3, OpenAI reached an important turning point. Yet despite its impressive performance, GPT-3 didn’t attract much attention. In order to make GPT-3 available to more people, Sam Altman convinced the other two founders to come up with the idea of creating a user-friendly interface.
After OpenAI received another $2 billion in funding from Microsoft, Altman and another executive, Greg Brockman, visited Bill Gates at his mansion on the shores of Lake Washington outside Seattle. The Microsoft founder is no longer involved in the company's day-to-day affairs but remains in regular contact with company executives.
Over dinner, Mr. Gates told them he doubted large language models would work. He will remain skeptical until the technology completes a task that requires critical thinking—such as passing an Advanced Placement Biology test.
Five months later, on August 24, 2022, Altman and Brockman came again, bringing with them Chelsea Voss, an OpenAI researcher. Ms. Voss won an International Biology Olympiad medal in high school. Nadella and other Microsoft executives were also in attendance.
On a giant digital display outside Gates' living room, the OpenAI team demonstrated a technology called GPT-4.
Brockman took a multiple-choice advanced biology test on the GPT-4 system, and Ms. Voss scored the answers. The first question concerns polar molecules, groups of atoms with a positive charge on one end and a negative charge on the other. The system answered correctly and explained its choice.
"It's just trained to provide answers," Brockman said. “The essence of chat almost magically disappears.” In other words, the AI is doing something that people didn’t really design it to do.
A total of 60 questions were tested, and GPT-4 only answered one incorrectly.
Gates sat in his chair, his eyes wide open. He had a similar reaction in 1980 when researchers showed him the graphical user interface that became the basis of modern personal computers. Gates believes GPT is revolutionary.
By October, Microsoft added GPT technology to online services including its Bing search engine. Two months later, OpenAI officially announced the release of ChatGPT, which is currently used by 100 million people every week.
In 2023, OpenAI will usher in the era of global large models
This year, OpenAI truly opened the era of global large-scale model competition.
Although GPT-4 has been developed as early as 2022, when OpenAI officially released GPT-4, the most powerful model in history, in March 2023, competitors who were busy catching up and copied ChatGPT must have been dumbfounded.
However, in the face of the huge power gap, what has been brought to the market is investment fanaticism and large-scale arms competition.
It is worth mentioning that Meta, as a disruptor, took the lead in open-sourcing the LLama model, which is a great benefit to the domestic "Battle of 100 Models".
After all, the phenomenon of "as soon as foreign countries open source, domestic research and development will be independent" is still the phenomenon.
Except for the independent research and development of a few companies and research institutions in China, everyone must know how the large models of the remaining companies come from. There are currently far more than a hundred large models "independently developed" in China.
This year, OpenAI defeated the efficient altruistic Anthropic and also defeated Google with the optimist Page.
What has Musk done in 2023?
The first is Musk’s acquisition of Twitter after 22 years and changing its name to X. The other is Musk’s affairs with Space thing.
We won’t go into detail about these today, because the pedestrian robot that I’m most looking forward to doesn’t seem to have a big breakthrough?
But in terms of AI, how can Musk, who missed DeepMind and OpenAI, stop here?
So you can see that in July this year, Musk officially announced the establishment of a new company, xAI, saying that its ultimate goal is to explore the nature of the universe. The team of the new company is very luxurious, almost all of which are from well-known researchers such as OpenAI, Google, DeepMind, and Microsoft.
Just last month, Musk finally launched the first AI model-Grok.
As for the effect of Grok, I won’t mention it because there has been no actual test.
But Lao Ma is indeed a Lao Ma, and he is the number one person in Silicon Valley in terms of marketing (unless you tell OpenAI that the coup incident was also a marketing incident).
Musk and Altman currently dislike each other.
First, Ultraman posted a post to tease Musk, using GPTs to create a model with the same name as Grok:
After that, Musk used his ultimate move-"Geme".
Musk seems not to care about anything, but also seems to care about everything.
In November 2023, OpenAI’s week-long regime struggle
Regarding this part, the author will not spend more time talking about it. I believe that everyone is tired of eating melon and has become numb.
Readers who are still interested can read Xi Xiaoyao Technology’s past articles:
-
"Father of ChatGPT" Ultraman was fired from OpenAI! ! Is the Steve Jobs story repeating itself?
-
Negotiations broke down! Altman will no longer serve as OpenAI CEO...
-
OpenAI founders Altman and Greg Brockman will join Microsoft
-
On the first anniversary of ChatGPT, Ultraman officially announces OpenAI’s new actions!
The only thing worth saying is that Nvidia is the biggest winner.
In 2024, can "Google + DeepMind" still catch up with "Microsoft + OpenAI"?
As a company with 200,000 employees, Google has always maintained a leading position in the field of AI and has released a series of landmark works in AI history such as AlphaGo and Alphafold. This year, it was given over by OpenAI, a start-up company with only 800 people. A slap in the face.
After the advent of ChatGPT, Google's series of activities gave people the feeling of being "one step behind". First, an error in the answer demonstrated at the Bard press conference caused Google's market value to evaporate by more than 710 billion yuan, and then there were constant reports of company employees and external users Questioning Bard's ability "can only benchmark against GPT-3.5 but not in the same dimension as GPT-4."
In April this year, Google realized that it had fallen behind and began to try to break the situation. Google announced that it would merge its two well-known research institutions, Google Brain and DeepMind, to form a new department, Google DeepMind.
Google DeepMind took more than half a year, used tens of thousands of Google TPU AI chips, and estimated that it would cost hundreds of millions of dollars to break the boat. Gemini was originally scheduled to be released next week.
However, just yesterday, The Information just broke the news that Google has postponed the release of the Gemini large model.
Obviously, Google’s Top Gun and Last Stand seem to have suffered a serious setback today. According to The Information, the reason why Google delayed the release of Gemini seems to be because Gemini cannot reliably handle ‘non-English’ tasks.”
This reason sounds really disappointing.
Finally,On the first anniversary of the release of ChatGPT, does Google DeepMind still have a chance?
AI has been turbulent in Silicon Valley for ten years. Does China still have a chance?
Note: This article is based on the translation, processing and editing of the New York Times article.