While Mark Zuckerberg is still splashing out tens of billions of dollars to recruit talent and build a superintelligence empire, an increasingly jarring chorus of dissent is rising from within Meta.
Last week, during an internal Meta livestream meeting attended by thousands of employees, one staff member suddenly interrupted the presentation, lost their composure, and unleashed a string of profanities. They complained about being reduced to a corporate tool, then publicly named a Meta AI executive and asked the meeting host to pass along a message:
"Tell him that he's a piece of sh*t."
According to Wired, one of the presenters on stage was so embarrassed that they covered their face with both hands. Although the host then asked everyone to mute their mics and proceeded with the technical presentation, many employees continued to discuss the "spicy" opening in the chat.
At the epicenter of the storm is Applied AI (AAI), a new division Meta established in March of this year. The unit is massive, with 6,500 engineers and product managers, and its primary mission is to support Meta's newly formed Meta Superintelligence Labs (MSL).
According to Meta's official narrative, this team carries the vital mission of helping evolve AI models. But in the eyes of many employees, they look more like conscripted "grunts."
According to multiple employees, the company's approach to forming Applied AI was remarkably heavy-handed: many staff members did not volunteer for the move but instead received an email one day informing them they had been reassigned to the new team. One Meta employee also posted on Reddit that the entire transfer process felt pretty random.
And the options placed before them were essentially only two: either join Applied AI or leave Meta.
"It's literally the gulag," one of the employees claims. "You have zero purpose in life all of a sudden, you barely interact with anyone, you just have these tasks every week."
So what exactly do the engineers who transferred into Applied AI do every day?
According to internal Meta documents and employee disclosures, their primary job is not product development, nor is it researching model architectures. Instead, it is producing training data. Specifically, they are tasked with designing coding problems, writing complex software development scenarios, constructing logic puzzles, evaluating model performance, creating test samples, and annotating data.
Typically, each employee must complete at least two of these tasks per week. This material is ultimately used by AI scientists to train and evaluate the latest generation of large models, and to help future AI agents learn how to write code, operate software, and complete complex tasks.
Meta's internal explanation is that current AI models still cannot truly surpass humans on technical tasks like programming, and therefore require a large volume of real-world examples for training. From a technical standpoint, these tasks are simpler than the development work they used to do. But that is precisely the problem: the work is too mechanical, too repetitive, and too devoid of creativity.
One of the employees said the work now is "not using their full skill set and knowledge." They joined Meta to build products for billions of users, but now feel they spend every day just producing feed for AI models.
In truth, Meta did have other options. Training data construction has long been a function largely outsourced to contract teams, and one of the most well-known companies in the space was Scale AI. And last year, Meta closed a deal that shook the industry:
It acquired Scale AI's core business for a staggering US$14.3 billion and brought on board its founder—then just 28 years old—Alexandr Wang, appointing him Meta's Chief AI Officer and placing him in charge of Meta Superintelligence Labs.
In a previously leaked recording of an internal meeting, Zuckerberg explained why the company did not go with the traditional outsourcing approach: Wang knows the data-labeling industry inside and out, and Meta employees have a "significantly higher" average intelligence level than third-party contractors.
Thus, Meta concluded that having in-house engineers do the work would yield better results.
Applied AI is not the only team experiencing problems; discontent within Meta has spread across the entire company.
Since the start of this year, Zuckerberg has staked nearly his entire strategic focus on AI. To free up resources, the company has undertaken sweeping organizational restructuring. Just last month alone, Meta cut roughly 8,000 employees, representing about 10% of its total workforce.
At the same time, the company has redirected massive resources from its metaverse business toward AI. Maher Saba, the leader of Applied AI, previously came from Reality Labs—the very division that had burned through more than US$83 billion over the past few years. In addition, multiple employees have revealed that the data-center engineering team, the Instagram team, and infrastructure departments are all facing unprecedented workloads.
To obtain even more training data, Meta also rolled out a highly controversial new initiative: it began monitoring US employees' mouse clicks and keyboard inputs and using that data for AI model training. The move quickly sparked backlash. More than 1,600 employees have now signed a joint petition calling for the program to be halted.
Under pressure, Meta subsequently made some concessions: it allowed employees to pause data collection for up to 30 minutes and to apply for individual exemptions. But these measures have clearly failed to fully quell the controversy.
It is not just rank-and-file employees—this anxiety has even spread to management. According to an internal meeting recording obtained by Wired, during last week's Instagram all-hands meeting, Meta's Chief Product Officer Chris Cox made a rare public comment on the company's current state.
He described the past few months as "difficult" and "brutal," and used a vivid metaphor to characterize employees' circumstances: it's like "running a marathon in the middle of a hailstorm and then, like, your teammate gets replaced, and then we're recording you." At that point, Cox himself could not help but curse, "What the f**k." In his view, Meta needs to reconnect with its employees while avoiding deifying AI.
Faced with mounting discontent, Zuckerberg ultimately had no choice but to personally step in to soothe employees.
In a recent internal memo, he acknowledged for the first time: "The recent organizational changes have indeed caused distress among employees." He also conceded, "Because these adjustments are extremely complex, we have already made some mistakes, and we will likely continue to make more in the future."
Nevertheless, to steady the troops, Zuckerberg also made a series of commitments:
- No more large-scale layoffs this year;
- Capping the number of direct reports per manager;
- Increasing team activity budgets;
- Restoring assigned seating;
- Hosting a large-scale Hackathon event.
Among the most closely watched changes is the optimization of management structure: previously, there were extreme cases within Applied AI where one manager oversaw 50 employees; going forward, the company will impose limits on the number of direct reports per manager.
In addition, Zuckerberg made a specific mention of Applied AI in the memo. He explained to employees that this is not a career dead end, but rather a transitional phase.
However, at least for now, Zuckerberg's remarks have not fully alleviated employee anxieties. After all, everyone knows that under pressure from competitors like OpenAI, Google, and Anthropic, Meta is engaged in a high-stakes AI gamble.
And those engineers transferred into Applied AI are on the very front lines of that gamble—or perhaps more accurately, they have been swept into a war they never signed up for.
Related News:
Deepline | OpenAI, Anthropic, and SpaceX set to test US capital markets in historic IPO wave
Deepline | OpenAI files confidentially for IPO as AI giants race to list
Comment