原文
Mark Zuckerberg admitted to weaknesses in the company's restructuring during an internal town hall. The AI agents Meta reorganized around are progressing slower than planned, Zuckerberg said. His AI chief, meanwhile, painted a rosier picture. Meta CEO Mark Zuckerberg acknowledged at an internal town hall on Thursday that the systems known as AI agents haven't advanced as fast as expected, according to an audio recording obtained by Reuters. The corporate restructuring didn't go as "clean" as it could have, he said, and executives misjudged the timing. The "trajectory of the agentic development over at least the last four months hasn't really accelerated in the way that we expected," Reuters quoted Zuckerberg as saying. The bets on the new structure "haven't come to fruition yet." The admission carries real weight. Zuckerberg spent the past year going all in on catching up in AI. He put Alexandr Wang in charge of the AI division , rebranded it as Meta Superintelligence Labs , and offered top talent nine-figure sums to lure them away from rivals. In April, the company released Muse Spark , the first model in a new lineup. It posted solid benchmark scores but didn't match OpenAI or Anthropic. Meta laid off roughly ten percent of its global workforce in May and moved about 7,000 employees into AI teams. The goal was to fund expensive AI infrastructure and squeeze efficiency gains from AI-powered workflows. When planning started in January and February, senior leaders worried they weren't moving fast enough, Zuckerberg said. At the time, executives were "super optimistic" about tools like Anthropic's Claude Code. Meta plans to spend up to $145 billion on AI infrastructure this year, a sizable chunk of the more than $700 billion Big Tech is pouring in collectively. Zuckerberg expects more tangible results within the next three to six months. A Meta spokesperson declined to comment. According to Bloomberg, Meta is also building a cloud business to sell excess AI compute capacity to outside customers. At the same town hall, AI chief Alexandr Wang struck a different tone, according to Business Insider . Meta's upcoming model, code-named "Watermelon," has caught up with OpenAI's top model GPT-5.5, he said, citing benchmarks he didn't specify. "Watermelon, our next model after Avocado, is currently in training," Wang said, per Business Insider. "Watermelon uses an order of magnitude more compute than Avocado," he added, referring to the internal code name for Muse Spark, which shipped in April. On X, Wang went into damage-control mode. Zuckerberg was talking about the progress of the entire industry, not Meta's AI efforts specifically, he said. A Muse Spark update with major improvements to coding and agentic capabilities is coming soon. A coding model on par with Anthropic's Claude Opus is "pretty soon" to follow, and users will like what the team has been "cooking." Separate from the model discussion, CTO Andrew Bosworth addressed Meta's controversial mouse-tracking software at the same town hall. The tool records mouse movements and digital activity from employees to generate AI training data. Meta had paused the program after potentially sensitive data was exposed. An internal review found that no employee data made it into AI training, Bosworth said. When Meta first installed the software on U.S. employees' machines in April, he had told them there was no way to opt out. If the program restarts after the review wraps up, it will run on an opt-in basis. "For people who are comfortable, that's great, they can contribute to this kind of great human survey," Bosworth said. "To people who are not, it is not an issue."