Manufacturers that lag behind in the large model competition take two catch-up routes: Apple, Meta and other large manufacturers make up for their shortcomings by digging people with heavy money; Due to the deterioration of the financing environment, the domestic AI six tigers chose to lay off employees and tighten resources, focusing on model iteration.
Manufacturers who are at a disadvantage in the large model competition have started knives against their AI teams.
Apple, which has been criticized by the outside world for lagging behind by several steps in AI, has recently had some slightly exciting news: in addition to John Giannandrea, Apple’s senior vice president of AI and machine learning strategy, who was expelled from Apple’s core management and was demoted, there was also news that Apple is expected to acquire Perplexity, a star AI search startup, hoping to strengthen Apple’s talent and technical reserves in the field of AI.
Based on a valuation of $14 billion after Perplexity completed a new round of financing in May, if the acquisition is successful, it will be the largest merger and acquisition in Apple’s history, surpassing the $3 billion acquisition of Beats in 2014.
Apple’s talent acquisition plan is still hanging in the air, and on the other hand, Mark Zuckerberg, the founder of Meta, has spent $14.3 billion in real money to take Scale AI co-founder and CEO Alexandr Wang under his command.
In April, Meta’s new generation of open source model, Llama 4, fell short of expectations, which gave Zuckerberg the idea of restructuring the AI team and began to take action, starting a new superintelligence team internally, planned to be led by Alexandr Wang.
In addition to spending a lot of money to poach Alexandr Wang, Zuckerberg was also exposed to contact former OpenAI chief scientist Ilya Sutzkerfer, CEO of the startup SSI (Safe Superintelligence) – Daniel Gross, and former GitHub CEO Nat Friedman.
What does a product manager need to do?
In the process of a product from scratch, it is not easy to do a good job in the role of product manager, in addition to the well-known writing requirements, writing requirements, writing requirements, there are many things to do. The product manager is not what you think, but will only ask you for trouble, make a request:
View details >
Almost at the same time that Zuckerberg was busy recruiting industry giants, it was also revealed that Google would appoint DeepMind CTO Korai Kavukeglu as Google’s chief AI architect, responsible for coordinating future AI product development and reporting directly to CEO Sundar Pichai.
The above personnel changes are a new move by Google to restructure its Gemini application team to speed up the pace of catching up with OpenAI.
Unlike Zuckerberg, who worried that Llama would lag behind OpenAI in model performance, Google is gradually approaching OpenAI with Gemini series model updates, but it is becoming more and more aware of the user size gap between its products and ChatGPT. How to make up for the shortcomings of lagging products has become Google’s top priority.
It is not only foreign large model manufacturers who are caught in a turbulent period of adjustment. The AI Six Tigers in China, which were impacted by DeepSeek, have entered a dense wave of executive departures without exception.
It is worth noting that compared with Apple, Meta and other heavily expensive catch-up methods, the six tigers almost all chose to lay off employees to tighten the catch-up strategy of resource allocation.
The six tigers, who also held high the model and applied two-wheel drive last year, had to re-establish technology iteration as the company’s highest priority in the new large model competition cycle.
Under the change in strategy, a group of executives responsible for application and commercialization have left one after another, including but not limited to Zhang Fan, former COO of Zhipu, Wei Wei, former head of MiniMax commercialization, Ming Chaoping, head of the core product of the dark side of the previous month, and Zhang Xinhao, former product leader of “Bubbling Duck” of Stephen Leap Star (who has not yet left and turned into an internal consultant)……
Whether it is using heavy money to dig up people to make up for their shortcomings, or streamlining the organization through layoffs, it has become a new bargaining chip for Chinese and American large model manufacturers to try to keep themselves on the AGI table.
01
Zuckerberg once again regained the state he had when he founded Facebook (the predecessor of Meta) and began to become the company’s number one HR.
Frequently contacting AI researchers in an attempt to dig them out of Meta, Zuckerberg also set up a new AI lab with his own hands, which is seen as part of a large-scale restructuring of Meta’s AI business. The laboratory is expected to have 50 people.
Now, with the acquisition of Alexandr Wang, the director of the aforementioned AI lab is in place.
Foreign media The Information broke the news that as early as mid-April, Zuckerberg contacted Alexandr Wang, hoping that Wang would join Meta and be willing to pay billions of dollars to achieve this goal.
But Alexandr Wang was not satisfied with Zuckerberg’s offer and once raised his offer and the offer of Scale AI behind him to $20 billion. In the next month or so of negotiations, the two sides made some concessions to each other, and Meta finally acquired a 49% stake in Scale AI for $14.3 billion, and Alexandr Wang’s plan to join Meta was released.
In order to prevent Alexandr Wang from leaving midway, Meta specifically stated in the terms of the acquisition that Alexandr Wang needs to work at Meta for more than five years to get most of the acquisition cash. In addition to Alexandr Wang, Zuckerberg’s recruitment continues. According to CNBC, Zuckerberg also tried to acquire SSI, an AI startup founded by former OpenAI co-founder and chief scientist Elijah.
According to the latest valuation calculation after SSI completed a $2 billion financing in April, Zuckerberg may have to pay $32 billion for the acquisition.
However, Ilya rejected Zuckerberg’s temptation to pay a high salary. After encountering a closed door, Zuckerberg directly turned the bow and replaced the poaching target with SSI’s current CEO Gross.
As of now, Zuckerberg’s recruitment of Gross and Friedman is still under further negotiations.
Different from Zuckerberg’s behavior of smashing money and robbing people, the six domestic tigers staged another scene of organizational adjustment, that is, not only did they not poach people from the outside, but instead began to cut the core executives who had finally been recruited, from Zhipu to the dark side of the moon, Step Star, MiniMax, Baichuan Intelligence and Zero One Thousand Things, since the second half of last year, a number of executives have left the company.
According to the alphabet list, most of these departing executives are concentrated in the application and commercialization fields.
The latest transfer from the core position is Zhang Xinhao, the product head of Step Star. At the beginning of June, according to “Intelligent Emergence”, in December last year, the role-playing application “Bubbling Duck” under Step Star stopped large-scale investment, and the team was merged into the AI assistant application “Yue Wen” (now renamed “Step AI”).
Alphabet List learned that Zhang Xinhao, as the product leader of “Bubbling Duck”, has changed his current role within the company to a consultant.
Although not completely resigned, the advisory role represents more of a vacant position. Previously, Zhang Qianchuan, the former head of MiniMax products, completed his resignation in a similar way. In September last year, Zhang Qianchuan was resigned, and then the official said in the external response that Zhang Qianchuan faded out of the company’s affairs due to personal reasons and changed to the position of product consultant.
Almost at the same time as Zhang Qianchuan’s reappointment as a consultant, Ming Chaoping, former product manager of Byte Clipping and head of core products of the dark side of the moon, also officially resigned from the dark side of the moon and began to start a business.
Since December last year, a group of executives who have helped the company develop commercial revenue, including Hong Tao, co-founder and head of commercialization of Baichuan Intelligence, Wei Wei, partner and vice president of MiniMax, and Zhang Fan, COO of Zhipu, have left one after another.
02
One of the core reasons why Chinese and American large model manufacturers collectively entered a turbulent period is that these players have temporarily fallen behind in the new round of competition and have to catch up with the pace of the pioneers through new organizational adjustments.
Pushing Zuckerberg to reorganize the Meta AI team and recruit Alexandr Wang is a big belief that Meta has fallen behind in the artificial intelligence race.
After the release of Llama 1 in early 2023, Meta became the king of open source for large models. Under the premise that OpenAI seized the opportunity, Zuckerberg used open source to grab a ticket for Meta in the competition for large models.
But now, this ticket has the possibility of being lost. In early April, Llama 4 suffered a large-scale negative evaluation as soon as it was released, which can be called the biggest “rollover” event in the AI industry this year. Llama 4 was not only exposed to the special version to brush the list, but also was exposed to do a data overfit test, that is, the test sets of various benchmarks were mixed in the post-training process, with the aim of getting a good result in various indicators.
Although Llama technicians have come forward to deny the rumors of data overfitting and falsification, it is an indisputable fact that the special version is used.
As early as January, after DeepSeek R1 exploded, it was revealed within Meta that it began to worry that the unreleased Llama 4 might not be able to catch up with DeepSeek R1 in terms of performance.
At that time, a Meta employee posted on Blind, an anonymous gossip sharing platform in Silicon Valley, that Meta’s generative AI department was in a panic because of DeepSeek, and even broke the news that the unreleased new generation of open source model Llama 4 had lagged behind DeepSeek in benchmarks.
Compared with Zuckerberg’s crazy grabbing move due to concerns about model backwardness, although Google has almost equaled OpenAI in model performance with the Gemini 2.5 series, Google also has its own new challenge, that is, the product is too far behind ChatGPT.
According to an internal report circulated by Google, as of March 2025, Gemini has 35 million daily active users and 350 million monthly active users worldwide. For comparison, ChatGPT has 160 million daily active users and 600 million monthly active users.
OpenAI CEO Sam Altman even said that “ChatGPT replaces search”, “Everyone regards ChatGPT as a replacement for Google.” ”
Compared with the level of model capabilities, Google, which started with applications, is undoubtedly more concerned about the victory or defeat of this battle for “attention resources” in the AI era.
The domestic six tigers, which were also impacted by DeepSeek, fell into a new round of technical self-proof crisis, just like Meta’s Llama.
Worse than Llama, Zuckerberg still has a steady stream of strong financial support from Meta, and starting from the second half of 2024, the sharply deteriorating financing environment is changing the road to catching up with the technology of the six tigers: Zero One Everything has clearly abandoned AGI, and Baichuan Intelligent has shifted to medical vertical scenarios.
The four small strong players who insist on the two-wheel drive of models and applications have almost simultaneously given up their ambition to further expand the scale of applications and began to bet on model iterations with limited resources.
In this context, laying off core executives responsible for product application and commercial promotion has become a must-have issue in front of the six dragons.
03
One throws money to rob people, and the other lays off employees to reduce expenses, which reflects the current two different large model catch-up strategies:
large factories with rich wealth can exchange money for time to maximize efficiency;
Startups with limited funds can only shrink limited resources to maximize value.
In exchanging money for time, domestic bytes and other major manufacturers have taken the lead in setting an example for Meta. Also relying on the means of throwing money to rob people, Zhang Yiming led the Byte AI team from a backward student in 2023 to become the first team player in China at the end of 2024 after a year of development.
How to make model performance comparable to the industry’s top in limited resources, DeepSeek in China has also taken the lead in this regard. With the help of a series of engineering innovations, Liang Wenfeng developed his own R1 with performance comparable to o1 with less than one-tenth of the capital cost of OpenAI o1.
Whether it is Byte’s catch-up or DeepSeek’s sudden rise, what supports them to become the top of the large model industry, in addition to being willing to spend money to grab people, more importantly, they have clearly shown their ambition and ambition to pursue AGI, “This (big prospects) is what attracts top talents more.” A PhD student in the field of AI explained to the alphabet list.
After all, in the highly talent-intensive large model industry, how to acquire and attract a steady stream of young talents to join is the ultimate way to reach AGI.
In the face of Zuckerberg’s crazy grabbing, Altman recently responded that he heard that Meta regards OpenAI as their biggest competitor. “I think it’s a rational choice for them to continue to work, although their current AI progress may not be as expected. I respect their aggressive attitude and their spirit of constantly trying new things. ”
But as soon as the conversation changed, Altman countered proudly, even though Zuckerberg began offering signing bonuses of up to $100 million to some members of the OpenAI team, “But I’m really happy that so far, none of our best people have accepted their offers.” ”
One of the reasons given by Altman is that when comparing OpenAI and Meta, these talents will think that OpenAI has greater possibilities for achieving superintelligence.
The firm pursuit of AGI’s ambitious goals is also a great confidence for Liang Wenfeng when competing with large manufacturers for talents. When asked “How to ensure that DeepSeek is the first choice for people who make large models?” Liang Wenfeng’s answer was: “Because we are doing the most difficult thing.” The biggest attraction for top talents is definitely to solve the most difficult problems in the world. ”
From OpenAI to DeepSeek, their successive successes have at least revealed a signal to the outside world: if you want to innovate in the era of large models, it is not enough to spend money, and more importantly, you have to give talents room to show their ambitions.
