Zuckerberg acknowledges shortcomings in Meta's AI restructuring

Zuckerberg also criticized the speed of development of agentic AI systems – automated software that performs tasks for users independently. The development of AI agents had not shown the expected acceleration over the past four months. However, he assumed that the massive AI investments would deliver first significant returns in the coming three to six months.

Restructuring and job cuts

On Thursday, Meta also expanded the so-called Meta Business Agent globally. The tool can be deployed on Instagram, Messenger and WhatsApp and is intended to enable companies to answer customer inquiries, recommend products and handle sales without human intervention.

Zuckerberg also commented on a controversial internal program in which mouse movements and digital activities of employees were recorded for the training of AI models. After its introduction in April, there was no possibility for US employees to opt out of participating. Meta CTO Andrew Bosworth explained, according to Reuters, that an investigation had found that no employee data had been incorporated into AI training. Should the program be resumed after the review is completed, this should be done on a voluntary basis.

Controversy over employee tracking

The background to the controversy is that Meta had paused the mouse-tracking software the previous month after sensitive data had become public. The matter throws a spotlight on the group's difficult balancing act between the rapid expansion of its own AI capabilities and the protection of its own workforce's data.

For the coming fiscal years, Meta has significantly raised its investment planning. The company expects spending on chips, data centers and infrastructure of between 125 and 145 billion US dollars this year. This makes Meta one of the heaviest-spending companies in the global AI arms race.

Billions in investments in AI infrastructure

Specifically, the newly created Meta Superintelligence Labs, led by Alexandr Wang, is working on an AI model with the codename Watermelon. Wang has told employees that Watermelon has already matched the flagship model GPT-5.5 of competitor OpenAI in internal benchmark tests. The Watermelon model builds on the Avocado model, known internally as Muse Spark, and according to Wang is to be operated with an order of magnitude more computing power than its predecessors.

Model 'Watermelon' against OpenAI

Wang also announced that updates for the current Muse Spark model would follow shortly, with improvements particularly in programming capabilities and agentic functions. Meanwhile, according to the company, trading volume via the x402 protocol supported by AI agents has totaled around two million US dollars over the past 30 days.

The internal tensions described come as little surprise to observers. The combination of mass layoffs, a controversial data practice toward the company's own workforce and an ambitious AI course that is lagging behind schedule poses significant communicative and operational challenges for the leadership around Zuckerberg.

Analysts nevertheless see the company in a strong position: with the announced infrastructure investments and a clearly named head for the AI division, Meta has the means to close the gap to competitors such as OpenAI in the coming quarters – provided that the announced models actually meet the expectations placed in them.

Zuckerberg also indicated that he expects significant efficiency gains in the medium term. However, he acknowledged that he had envisioned a broad, rapid implementation of AI agents in everyday life happening faster.