At the end of the year, Zuckerberg holds 3.35 million H100! His new goal: AGI, and it will be open source!

Wallstreetcn
2024.01.19 04:38
portai
I'm PortAI, I can summarize articles.

The AI department undergoes a major restructuring, accelerating the training of Llama 3. By the end of the year, we will have 350,000 H100 cards in hand, and Meta Platforms is sprinting full speed ahead in the AGI race.

As we all know, the artificial intelligence rising star OpenAI sees achieving artificial general intelligence (AGI) as its historical mission. Alphabet's long-standing tech giant, DeepMind, has also set the same goal. Now, Meta Platforms, which previously bet heavily on the metaverse and suffered huge losses, has also joined the AGI army.

"We need to build general (artificial) intelligence," said Mark Zuckerberg, CEO of Meta Platforms, in a recent interview with tech media The Verge.

According to The Verge article, although Zuckerberg has not yet set a specific timetable for achieving AGI, or even provided a clear definition, based on the company's recent actions, it can be seen that Zuckerberg's goal is to achieve AGI and may even eventually open-source it.

Major restructuring of the AI department, accelerating the training of Llama 3, and Meta Platforms racing into the AGI track with 350,000 H100 by the end of the year

"The AI research group is being transferred to the company's direct product department, attracting key technical talents."

Firstly, Meta Platforms' product department will undergo a major restructuring in 2024, with the AI research group FAIR being transferred to the company's direct product department.

Its Vice President, Joelle Pineau, and Chief AI Scientist, Yann LeCun, will report directly to Meta Platforms' Chief Product Officer, Chris Cox, instead of reporting to Andrew 'Boz' Bosworth, the Chief Technology Officer of Meta Platforms' Reality Labs department.

The restructured team will develop generative AI products in various applications of Meta Platforms, allowing billions of users to experience the hottest AI technologies of the moment.

"We believe that in order to build the products we want to build, we need to build general (artificial) intelligence," Zuckerberg said in a previous interview with The Verge. "I think it's important to convey this point because many of the best researchers want to dedicate themselves to solving more ambitious problems."From Zuckerberg's words, we can sense that the competition for talent in the AI field has never been as fierce as it is now.

According to reports, AI developers can earn annual salaries of over 1 million dollars, which is astonishing. CEOs like Zuckerberg often have to woo key technical talents to prevent them from being poached by competitors.

"We have become accustomed to intense talent wars," Zuckerberg said, "but the situation here is different, with many companies pursuing the same goals, and many venture capitalists and retail investors investing in different projects."

Significant Upgrade in Llama 3's Reasoning and Planning Abilities

On Friday local time, Zuckerberg updated his Instagram with the news that Meta Platforms is currently training Llama 3. According to The Verge, Llama 3 will not only have the code generation function of Llama 2, but also advanced reasoning and planning abilities.

As explained by Zuckerberg, Meta Platforms' focus on AGI has been influenced by the release of the large-scale model Llama 2 last year.

Meta Platforms believes that the coding ability of Llama 2 is not meaningful for ordinary users to use large language models (LLMs) in Meta Platforms' applications, but it is an important skill for building more intelligent AI models.

"Llama 2 is not the industry-leading model, but it is the best open-source model," Zuckerberg said. "For Llama 3 and more advanced versions, our goal is to build the most advanced product and eventually become the industry-leading model."

By the end of the year, 350,000 H100 chips will be obtained, and the chip inventory will reach nearly 600,000

In addition to talent, the scarcest resource in the AI field is the computing power required to train and run large-scale models. On this issue, Zuckerberg seems to be well prepared.

Zuckerberg revealed on Instagram that by the end of this year, Meta Platforms will have approximately 350,000 NVIDIA H100 chips, and if other GPUs are included, there will be approximately 600,000 H100 equivalent computing power.Mark Zuckerberg said, "We have built this capability, which may be larger in scale than any other independent company."

External research predicts that Meta Platforms' H100 shipments will reach 150,000 units by 2023, on par with Microsoft and at least three times higher than other companies.

Meta Platforms AGI, possibly open source!

AGI is a super powerful artificial intelligence that, if given inappropriate goals and motivations, could have disastrous consequences for human society. Therefore, there has always been controversy over who will control AGI, as evidenced by recent internal disagreements at OpenAI.

Due to Zuckerberg's voting control over Meta Platforms through his shareholding, he has absolute power within the company and can single-handedly decide the design direction and goals of AGI. If AGI becomes a reality at Meta Platforms one day and is designed improperly, Zuckerberg's decisions could lead to the development of harmful AGI for humanity.

In response to this concern, Zuckerberg did not give a direct answer, but he pointed out that Meta Platforms' current language model, Llama, is open source in most cases.

"As long as it makes sense and is safe and responsible, we tend to lean towards open sourcing," he said. "Obviously, you don't want to be held to something just because you said you would do it."

"I tend to think that one of the bigger challenges here is that if you build something that's really valuable, it's going to end up being very centralized," Zuckerberg said. "However, if you make it more open, then you can address a whole class of problems that may arise from unequal opportunities and value. So this is an important part of the whole open source vision."

Meanwhile, from a business strategy perspective, Zuckerberg believes that the models deployed by Meta Platforms so far have not caused catastrophic damage to the company's operations.

"In many cases, the biggest companies that start with the greatest competitive advantage are also the ones calling the most for how others should build AI and setting necessary guardrails." "I'm sure some of their concerns about company security are valid, but this approach aligns closely with the business strategies and interests of these companies."Currently, Zuckerberg's support for open source also has his own motives. He stated that the ultimate result of open sourcing large models is still the monopoly and concentration of technology, just in a different form.

Meta Platforms Platforms Platforms Platforms has more users than any other company in the world, and its social media business is highly profitable. Therefore, artificial intelligence can make its platform users more sticky. If Meta Platforms Platforms Platforms Platforms can effectively regulate the development of AI by publicly releasing models, its influence on the ecosystem will only grow stronger.