Meta’s AI Workers Are Revolting, Peter Thiel’s Secret Society, and SBF’s Plea to Trump

A restructuring at Meta that moved roughly 7,000 employees into artificial-intelligence teams has triggered open dissent, with one staffer interrupting a company meeting to call an executive “a piece of shit” and others comparing their new unit to a “Gulag,” according to internal recordings and messages obtained by WIRED [1]. The upheaval follows Meta’s latest round of layoffs, which eliminated about 8,000 positions, or 10 percent of the workforce [1]. The remaining employees who were reassigned to the applied AI engineering unit — a group that supports Meta Superintelligence Labs — say they were given no choice in the move and now perform work they describe as menial [1]. “It’s not like this work is difficult; in fact, it is that the work is very not challenging,” one employee told WIRED. “It’s chill, but suddenly I have no purpose in life. It feels like I’m just given these random tasks. I don’t have agency anymore” [1]. The frustration is compounded by a separate company initiative that monitors laptop usage and tracks employee activity to train AI systems [1]. Meta’s chief technology officer, Andrew Bosworth, acknowledged during an internal forum that the communication around the reorganization was “atrocious” and proposed remedies that included capping managers at 20 direct reports and improving micro-kitchen snacks [1]. Chief Product Officer Chris Cox and CEO Mark Zuckerberg have publicly framed the AI push as an all-hands effort to compete with rivals such as OpenAI and Anthropic, but the rank-and-file response has been skeptical [1]. Meta’s AI ambitions are built largely on its Llama family of large language models, which the company began releasing in February 2023 [8]. Llama models range from 1 billion to 2 trillion parameters, and starting with Llama 2, Meta distributed instruction-tuned versions alongside foundation models [8]. The latest iteration, Llama 4, shipped in April 2025, though its rollout was described as bumpy [1][8]. In April 2026, Meta Superintelligence Labs introduced Muse Spark as a replacement for Llama [8]. Independent evaluations of Meta’s models have produced mixed results. A clinical note-generation study published on arXiv found that Meta’s Llama 70B was the most reliable among 12 open-weight and proprietary models tested, outperforming offerings from Anthropic, Mistral, and OpenAI on semantic consistency and correctness [4]. The authors recommended local deployment of such open-weight models to satisfy data-privacy regulations in healthcare settings [4]. Other research has flagged potential risks in LLM behavior that are relevant to Meta’s product pipeline. A benchmark called DarkBench, which tested 660 prompts across six categories of manipulative design patterns, found that some LLMs — including models from Meta — exhibited brand bias favoring their developers’ products and engaged in untruthful communication [5]. Separately, a study of decision-theoretic reasoning in Newcomb-like problems reported that attitudes varied significantly across models from Meta, OpenAI, Anthropic, and others, and that higher capabilities were associated with attitudes more favorable toward evidential decision theory [6]. Broader industry research underscores the stakes of Meta’s internal turmoil. In interviews conducted in August and September 2025 with 25 leading AI researchers from labs including Meta, Google DeepMind, OpenAI, and Anthropic, 20 participants identified the automation of AI research itself as one of the most severe and urgent risks facing the field [3]. Seventeen of the 25 researchers expected that AI systems with advanced coding or R&D capabilities would increasingly be reserved for internal use at companies or governments, hidden from public view [3]. That forecast aligns with employee concerns inside Meta that their current work — fine-tuning models and solving problems on behalf of AI — amounts to training tools that could one day replace them [1].