In April, Mark Zuckerberg's lofty plans for the future of artificial intelligence crashed into reality.

Weeks earlier, the 41-year-old CEO of Meta had publicly boasted that his company's new AI model, which would power the latest chatbots and other cutting-edge experiments, would be a "beast." Internally, Zuckerberg told employees that he wanted it to rival the AI systems of competitors like OpenAI and be able to drive features such as voice-powered chatbots, people who spoke with him said.

But at Meta's AI conference that month, the new AI model did not perform as well as those of rivals. Features like voice interactions were not ready. Many developers, who attended the event with high expectations, left underwhelmed.

Zuckerberg knew Meta was falling behind in AI, people close to him said, which was unacceptable. He began strategizing in a WhatsApp group with top executives, including Chris Cox, Meta's head of product, and Andrew Bosworth, the chief technology officer, about what to do.

That kicked off a frenzy of activity that has reverberated across Silicon Valley. Zuckerberg demoted Meta's vice president in charge of generative AI. He then invested $14.3 billion in the startup Scale AI and hired Alexandr Wang, its 28-year-old founder. Meta approached other startups, including the AI search engine Perplexity, about deals.

And Zuckerberg and his colleagues have embarked on a hiring binge, including reaching out this month to more than 45 AI researchers at rival OpenAI alone. Some received formal offers, with at least one as high as $100 million, two people with knowledge of the matter said. At least four OpenAI researchers have accepted Meta's offers.

In another extraordinary move, executives in Meta's AI division discussed "de-investing" in its AI model, Llama, two people familiar with the discussions said. Llama is an "open source" model, with its underlying technology publicly shared for others to build on. They discussed embracing AI models from competitors like OpenAI and Anthropic, which have "closed" code bases.

A Meta spokesperson said company officials "remain fully committed to developing Llama and plan to have multiple additional releases this year alone."

Zuckerberg has ramped up his activity to keep Meta competitive in a wildly ambitious race that has erupted within the broader AI contest. He is chasing a hypothetically godlike technology called "superintelligence," which is AI that would be more powerful than the human brain. Only a few Silicon Valley companies -- OpenAI, Anthropic and Google -- are considered to have the know-how to develop this, and Zuckerberg wants to ensure that Meta is included, people close to him said.

"He is like a lot of CEOs at big tech companies who are telling themselves that AI is going to be the biggest thing they have seen in their lifetime, and if they don't figure out how to become a big player in it, they are going to be left behind," said Matt Murphy, a partner at the venture capital firm Menlo Ventures. He added, "It is worth anything to prevent that."

Leaders at other tech behemoths are also going to extremes to capture future innovation that they believe will be worth trillions of dollars. Google, Microsoft and Amazon have supersized their AI investments to keep up with one another. And the war for talent has exploded, vaulting AI specialists into the same compensation stratosphere as NBA stars.

Google's CEO, Sundar Pichai, and his top AI lieutenant, Demis Hassabis, as well as the chief executives of Microsoft and OpenAI, Satya Nadella and Sam Altman, are personally involved in recruiting researchers, two people with knowledge of the approaches said. Some tech companies are offering multimillion-dollar packages to AI technologists over email without a single interview.

"The market is setting a rate here for a level of talent which is really incredible, and kind of unprecedented in my 20-year career as a technology executive," Meta's Bosworth said in a CNBC interview last week. He said Altman had made counteroffers to some of the people Meta had tried to hire.

OpenAI and Google declined to comment. Some details of Meta's efforts were previously reported by Bloomberg and The Information.

For years, Meta appeared to keep pace in the AI race. More than a decade ago, Zuckerberg hired Yann LeCun, who is considered a pioneer of modern AI. LeCun co-founded FAIR -- or Fundamental AI Research -- which became Meta's artificial intelligence research arm.

After OpenAI released its ChatGPT chatbot in 2022, Meta responded the next year by creating a generative AI team under one of its executives, Ahmad Al-Dahle, to spread the technology throughout the company's products. Meta also open-sourced its AI models, sharing the underlying computer code with others to entrench its technology and spread AI development.

But as OpenAI and Google built AI chatbots that could listen, look and talk, and rolled out AI systems designed to "reason," Meta struggled to do the same. One reason was that the company had less experience with a technique called "reinforcement learning," which others were using to build AI.

Late last year, the Chinese startup DeepSeek released AI models that were built upon Llama but were more advanced and required fewer resources to create. Meta's open-source strategy, once seen as a competitive advantage, appeared to have let others get a leg up on it.

Zuckerberg knew he needed to act. Around that time, outside AI researchers began receiving emails from him, asking if they would be interested in joining Meta, two people familiar with the outreach said.

In April, Meta released two new versions of Llama, asserting that the models performed as well as or better than comparable ones from OpenAI and Google. To prove its claim, Meta cited its own testing benchmarks. On Instagram, Zuckerberg championed the releases in a video selfie.

But some independent researchers quickly deduced that Meta's benchmarks were designed to make one of its models look more advanced than it was. They became incensed.

Zuckerberg later learned that his AI team had wanted the models to appear to perform well, even though they were not doing as well as hoped, people with knowledge of the matter said. Zuckerberg was not briefed on the customized tests and was upset, two people said.

His solution was to throw more bodies at the problem. Meta's AI division swelled to more than 1,000 people this year, up from a few hundred two years earlier.

The rapid growth led to infighting and management squabbles. And with Zuckerberg's round-the-clock, hard-charging management style -- his attention on a project is often compared to the "Eye of Sauron" internally, a reference to the "Lord of the Rings" villain -- some engineers burned out and left. Executives hunkered down to brainstorm next steps, including potentially ratcheting back investment in Llama.

In May, Zuckerberg sidelined Al-Dahle and ramped up recruitment of top AI researchers to lead a superintelligence lab. Armed with his checkbook, Zuckerberg sent more emails and text messages to prospective candidates, asking them to meet at Meta's headquarters in Menlo Park, California. Zuckerberg often takes recruitment meetings in an enclosed glass conference room, informally known as "the aquarium."

The outreach included talking to Perplexity about an acquisition, two people familiar with the talks said. No deal has materialized. Zuckerberg also spoke with Ilya Sutskever, OpenAI's former chief scientist and a renowned AI researcher, about potentially joining Meta, two people familiar with the approach said. Sutskever, who runs the startup Safe Superintelligence, declined the overture. He did not respond to a request for comment.

But Zuckerberg won over Wang of Scale, which works with data to train AI systems. They had met through friends and are also connected through Elliot Schrage, a former Meta executive who is an investor in Scale and adviser to Wang.

This month, Meta announced that it would take a minority stake in Scale and bring on Wang -- who is not known for having deep technical expertise but has many contacts in AI circles -- as well as several of his top executives to help run the superintelligence lab.

Meta is now in talks with Safe Superintelligence's CEO, Daniel Gross, and his investment partner Nat Friedman to join, a person with knowledge of the talks said. They did not respond to requests for comment.