Picture source @ Visual Chinese Text | Fixed Focus, author | Liming, editor | Wei Jia OpenAI is becoming more and more closed, but its opponent Meta (former Facebook) has won a lot of praise through open source. On January 18, Mark Zuckerberg announced that Meta is planning to bu

Picture source @Visual China

text | Ding Jiao, author | Liming, editor | Wei Jia

openai is becoming more and more closed, but its opponent meta (former facebook) has won a lot of praise through open source.

On January 18, Mark Zuckerberg announced that meta is planning to build its own AGI (general artificial intelligence), which will reach or exceed the level of human intelligence in most fields. At the same time, he emphasized that he promised to open this technology to the public so that everyone can benefit from it.

is contrary to openai's refusal to open their algorithms, routes, and papers, while meta chooses to open it to the end. The large model llama that shocked the industry before, the well-received llama 2, and the llama 3 under development are all open source. They have become the objects of research, study and even plagiarism by many AI companies.

Through open source, meta has broken the monopoly of large models to some extent and become the third force besides openai and Google. Some people regard Zuckerberg as a hero in the big model era. Some people left a message under Zuckerberg’s post, suggesting that meta be renamed openai and openai be renamed closeai.

In fact, meta has open sourced almost all the infrastructure it developed in the past, such as react for web and app development frameworks, and the machine learning framework pytorch. Chatgpt was developed based on pytorch, and openai also runs on pytorch. According to meta chief scientist Yang Likun, except for Google, the whole world runs on pytorch because Google has its own system.

It's just that these things have been covered up by meta's negative news in the past - interfering in the electoral process, providing false information, endangering the health of young people. In Silicon Valley, scolding Facebook used to be a kind of "political correctness", and people thought it was a company with corrupt values. company of.

's reputation has changed drastically now. Meta seems to have become the hope of the entire AI circle for a while. Its stock price has risen by 160% in 2023, second only to Nvidia among the top five technology giants in the United States. It has recently returned to a trillion-dollar market value. Stock prices hit new highs. Openai, which once emerged as a challenger, just experienced a palace battle not long ago, exuding an increasingly strong smell of copper.

openai goes to the left and meta goes to the right. The two companies, which did not have much overlap in the past, are taking completely different paths.

openai, moving towards closure

Before discussing the situation of these two companies, let us first pull back the time ten years ago.

In 2013, there was an AI company called DeepMind, which attracted the attention of Google and Facebook (which had not yet been renamed Meta).

was a newly emerging company at that time, and both Google and Facebook wanted to acquire it. In the end, Google won the bidding and acquired DeepMind in January 2014. Three years later, the AI ​​robot AlphaGo developed by DeepMind defeated the Korean Go champion Lee Sedol 4:1 and became famous in one fell swoop.

When Google acquired deepmind, two people felt very uncomfortable. One is Zuckerberg, who had already realized the huge prospects of AI at the time; the other was Musk, who did not want AI to be controlled by large companies like Google.

So Zuckerberg decided to do it himself. He spent a lot of money to invite Yang Likun, an authority figure in the field of computer science from New York University, to launch the fair project and establish an AI laboratory at Facebook. Musk and Altman founded openai to fight Google.

Judging from the situation at the time, Zuckerberg was doing it for his own company, and Musk was doing it for all mankind. Musk believes that a small number of AI systems controlled by large companies are very unsafe, and the only way is to let as many people as possible have AI. Therefore, OpenAI was positioned as open source from the beginning and is a non-profit organization.

Musk said at the time: "We hope to have a Linux-like version of AI that is not controlled by any individual or company."

One is private and the other is public. The two companies chose different paths from the beginning.

The Linux mentioned by Musk is a free and open source operating system. In the early days of the Internet, when computers first became popular, there were not many operating systems available. Either the source code was strictly kept secret by software manufacturers, or they charged very high fees. A European college student developed a prototype of the Linux operating system and then disclosed his code to the outside world for free.Because

is open source, programmers around the world have joined in to improve the code, and finally created the Linux operating system, which is completely free to use. This greatly accelerated the popularity of computers. The Android system we are familiar with today is based on the Linux kernel, and most mobile phones in the world run on this system.

Musk’s idea is very simple. The AI ​​era also needs such an open source and open operating system, and the market cannot be monopolized by giants. The start-up funds of

openai came from donations from a group of technology giants. It did follow the open source path at the beginning. The gpt-2 released by in 2019 is a large open source model. In the year of , some people still used gpt-2 for "rights". "Game" to rewrite the ending of the script.

But after the release of gpt-2, openai gradually became closed. It later formed a for-profit subsidiary and received billions of dollars in investment from Microsoft.

After that, gpt-3 released in 2020, gpt-3.5 upgraded in 2022, and gpt-4 released in March 2023, are all closed source models. At the beginning of , openai also published papers, and later even the model Specific parameters will no longer be announced.

From the perspective of the competitive landscape of the entire large model industry, the dual giant combination of "openai-Microsoft" and "deepmind-Google" has officially formed.

Musk was very dissatisfied with this. He said that OpenAI was established as an open source company, but now it has become a closed source, profit-seeking company controlled by Microsoft. This was not what he wanted at all.

meta has not been idle during this period. In addition to tossing the metaverse, meta's AI team has been studying large language models and released some open source projects. Everyone does their own thing without interfering with each other.

A few weeks before OpenAI released chatgpt, Meta released a similar chatbot called Galactica, specifically designed to write scientific papers. Who would have known that this product not only failed to cause a stir, but also attracted a lot of criticism, with netizens denouncing that it would undermine scientific publishing. So much so that the meta people canceled the demo and couldn't sleep.

At that time, people were concerned about whether Meta's Metaverse project was about to go out of business and where the layoffs were. No one cared about how its AI was doing. As for openai, everyone thinks it represents a new force and is more tolerant to it.

According to Yang Likun, chatgpt came out two weeks later and was regarded as the arrival of the savior. For a long time afterwards, OpenAI was a popular player in the global technology community, ascending to the altar and taking away all the attention. People are talking about how OpenAI broke through Google's blockade and suppressed the giant.

In the United States, which advocates individual heroism, the script of young entrepreneurs challenging the old forces has always been full of topicality. As a result, the competition among the top big model players has turned into a battle between openai and Google.

But Zuckerberg, who has always been aggressive, will not stand idly by. Meta's secretly developed llama large model is on the way.

meta is the hope of the whole village?

In February 2023, 3 months after chatgpt was released, the first version of llama was open source. At first, this model could only be used for research. In July, the upgraded llama 2 supports free commercial use. meta has released a large number of details such as model training data, training methods, data annotations, etc., and all starting codes are open source. The performance of

llama 2 is very outstanding, surpassing all open source large models. Some people have found that its maximum parameter version has smaller parameters than gpt-3, but has better results.

Scientist Andrej Karpathy, who followed Musk from OpenAI to Tesla and was later poached back by OpenAI, regards the release of llama 2 as an important day in the development of artificial intelligence and large models, because this is "the first time anyone can The most powerful language model that can get model weights."

For a time, the entire AI circle looked at meta with admiration. When a group of giant companies launched a large model arms race and tried to achieve market monopoly through technological closure, meta used llama to tear a hole. What openai didn't do, meta did. Wang Sheng, a partner of

Inno Angel Fund, jokingly called Yang Likun "klaus lecun" (Yang Likun's English name was yann lecun, and klaus fuchs provided information about the Manhattan Project to the Soviet Union during World War II). Wang Sheng joked: "We can't just let the United States have atomic bombs."

Wang Sheng found that after the emergence of chatgpt, the technology circle was enthusiastic and confused about large models.Many investors are just watching the fun, "because first they don't understand, second they don't dare to invest, third they can't invest. A new technological paradigm has emerged, and no one knows where the boundaries of its capabilities are."

Wang Sheng said "Ding Jiao" said that llama's open source has a great impact on the industry. "meta's large model technology accumulated over the past few years was originally a secret that was not passed on. As soon as it was open sourced, many perceptions were quickly leveled."

Before, domestic teams developed large models and could only continue to try from scratch. Wrong, once something gets stuck, even if it is not too difficult from a technical point of view, it may take a year or two to solve it by yourself, and you need to step through many pitfalls yourself. Now meta directly brings out its experience and data, giving the industry very important guidance in terms of ideas and methods.

"We have learned a lot about in this process. If had not been open sourced by llama, the current level of domestic large models would have been much worse, and we would have been significantly left behind." Wang Sheng said.

This is why more than 200 large models suddenly emerged in China. Wang Xiaochuan's Baichuan Intelligence released a large model just two months after the company was founded, which actually borrowed from llama. The large-scale model launched by Kai-fu Lee's Zero-One-Wuxi also uses llama's open source model architecture. The most intuitive impact of

llama open source on openai is that a bunch of Chinese companies have jumped out to "beat" gpt. In their caliber, they have surpassed the gpt model in many indicators, although many of the lists are created by brushing up. The deeper impact of

is that the power of the large-scale open source community is growing rapidly, and programmers around the world can contribute to open source. They have developed various open source data sets, iterated on more new models, and narrowed the gap with large closed source models. So Google engineers said that Google has no moat, and neither does OpenAI. The image of

meta in the technology circle has become taller. Looking back at the development of meta, has actually always had an open source tradition.

As early as 2016, the meta team developed and open sourced the deep learning framework pytorch. It and Google's tensorflow have become the two mainstream frameworks in the field of deep learning.

is different from openai, which has only begun to secretly promote chip manufacturing projects. Meta designed a self-developed AI inference chip three years ago and adopted an open source architecture.

has launched many open source models in terms of language translation. The nllb model released two years ago is one of Wikipedia's translation suppliers. The seamlessm4t released in 2023 can translate nearly a hundred languages. It also released the world's largest open source model. Multimodal translation dataset. After comparison, some researchers at

believe that meta is more inclined to trust, accountability and the democratization of artificial intelligence through open source.

Looking back at openai, still maintains a leading position in technology, but refuses to be open.

"No one knows how far OpenAI has progressed, including the gpt-5 it is developing, and whether it will continue to use the transformer is unknown. Now everyone is following the technical route of gpt-3.5, assuming it quietly changes direction. , then everyone will fall into a trap. Because no one knows what the correct route is." said an AI company founder.

At the same time, it tries to block competitors to maintain its first-mover advantage.

ByteDance was deactivated by OpenAI in December 2023 because it violated OpenAI's terms of service when it called OpenAI's API to develop its own large language model. Behavior that goes against business ethics is definitely wrong, but the outside world has drawn attention to OpenAI's business terms, which prohibit customers from using the content output by gpt to develop any AI models that may bring competition to OpenAI.

It is worth noting that Microsoft also has similar terms. They value commercial competitive advantage. Chen Ran, founder and CEO of

artificial intelligence company Opencsg, believes that the market space of open source is far larger than that of closed source, but the future must be open source and closed source at the same time. “It doesn’t matter whether openai is open source or not. ."

is open source and closed source, who is right and who is wrong?

Is openai blinded by profit, or is meta selfless and wants to contribute to all mankind?

Perhaps, they are not fundamentally different.openai is not that selfish, and meta is not as noble as netizens advocate. The difference is only in the route and strategy.

In terms of contribution to the entire AI industry, openai is undoubtedly the largest. After all, this AI wave was started by chatgpt. It lit a fire that accelerated the popularity of large models. In fact, in the past two years, the large model open source community basically revolved around gpt-3.

openai is not like giants like Google and Meta that have strong capital accumulation. The funds donated by the big guys in the early years have been almost spent. Without money, you can't do research and development, and you can't use love to generate electricity. So openai embraced Microsoft, and Altman tried his best to make money for openai.

Chen Ran believes that the success of openai is actually the success of the business model. The essence is that investors see huge business value and profit model. The direction of competition between open source and closed source must be innovative competition in money-making models.

meta makes the technology open source, and it does not just make wedding dresses for others. open source is a strategy that can attract more developers to help it iterate technology and fix loopholes for free. As the saying goes, everyone adds fuel to the flames, but the core technology is still in the hands of meta, and it will not attack its opponents when it is time. Be soft-handed. Just like Google, its Android is an open source system, but domestic mobile phone manufacturers still have many restrictions when building their own systems.

Before the big model became popular, meta fell into the deep pit of the metaverse and was unable to extricate itself. It spent money, lost money, laid off employees, and saw no hope. The big model is the life-saving straw, and chatgpt is the light that illuminates the way forward.

took a deeper look and found that the closed source openai and open source meta of fundamentally have different judgments on the value of AI.

If AI is compared to nuclear weapons in the future world, openai believes that it is very dangerous and must be cautious and cannot be abused. meta believes that nuclear weapons cannot only be in the hands of a few players, but should be open to research.

The debate about the pros and cons of AI is an old topic in the technology circle. Musk believes in the "AI danger theory". He is worried that one day the intelligence and consciousness of machines will surpass humans, and they may replace humans and destroy human civilization. So he took the lead in establishing openai, but later openai did not follow the path he envisioned.

is based on the vision of "preventing AI from replacing humans", and two major factions have emerged in the industry. One group believes that a large number of independent AI systems are more secure than a small number of AI systems controlled by large companies, and AI research should be open source and open; the other group believes that research should be closed, which can mitigate security risks and prevent criminals from abusing the obtained code.

openai belongs to the latter. Those who support it believe that the open source strategy increases risks. For example, the CEO of an AI company in London believes that meta is the most irresponsible artificial intelligence participant. He asked: "Should we make the design of nuclear weapons transparent?"

meta Regarding open source as the optimal strategy, Yang Likun believes that machines will eventually be smarter than humans. Humans cannot prevent bad people from obtaining it. AI must be open source. Only by allowing more people to participate in it, will the ultimately developed system be safer. . He believes that Musk’s AI threat theory is just too much science fiction.

Yang Likun's former partner, Jeffrey Hinton, who won the Turing Award with him, holds a different view and believes that AI will pose a serious threat to mankind. His ideas may have had an impact on his student Ilya. Ilya is the co-founder and chief scientist of OpenAI, and almost determines OpenAI's technical route. He was also the key figure in the dismissal of Altman in the OpenAI palace dispute. Ilya has been committed to pursuing safe and reliable AGI. It is generally believed in the industry that the dispute between commercialization and AGI's routes triggered the conflict.

So, openai and meta have different beliefs. openai regards agi as the ultimate goal and believes that it will be the first to achieve it; meta regards the open platform as the best path and tries to formulate open source standards. Of course, they all hope to get the commercial benefits they deserve in this process.

In the long run, the development of AI will be in a spiral, and whether it is open source or closed source, it will inject momentum into it. Only time will tell who is really worried about the future of all mankind.