Meta (META.US) Zuckerberg Responds to AI “Slowdown” Dispute: Continuing to Promote Model Capabilities to Strengthen Security Protection

The Zhitong Finance App learned that Meta Platforms (META.US) CEO Mark Zuckerberg and Alexander Wang, the company's AI leader, voiced a dispute over the speed of AI development, arguing that AI laboratories should continue to advance model capabilities while building stronger protective measures around increasingly powerful models.

The two posted this statement on the X platform. Currently, the industry is debating widely on “whether to slow down capability development before AI security makes more progress.” The so-called AI safety is to ensure that AI systems reliably follow human intentions and values.

Meta's stock price was higher for three consecutive trading days as of Tuesday, with a cumulative increase of 4.2% and a slight increase in night trading. On Stocktwits, retail sentiment on the stock was “extremely bearish” as of the most recent review.

Zuckerberg: Labs are motivated to create safer AI

Zuckerberg said that AI companies already have strong commercial and legal incentives to promote model safety and align with users. “People won't want to use intelligence that is inconsistent with them and doesn't act according to their own requirements,” Zuckerberg wrote. “Any lab that doesn't focus on alignment will fall behind.”

He believes that alignment itself is becoming an important competitive capability rather than just a constraint on AI development. “My opinion is that trust and alignment are quickly becoming the most important ability to distinguish agents from models.” he said.

Zuckerberg also mentioned that Meta had decided to delay the release of the Muse model for several months to address safety and security issues.

Meta supports external oversight

Zuckerberg and Alexander Wang called for the introduction of more independent evaluators and external supervision as AI models become more powerful.

“Bringing in independent evaluators and consultants is an industry best practice,” Zuckerberg said, and believes that a larger and more diverse evaluator ecosystem will improve AI security efforts.

Alexander Wang also said that AI laboratories should maintain a “strong governance framework” throughout the training and deployment process, including external evaluators and independent supervision of model release safety standards.

He added: “Every laboratory must operate under the protection of a democratic country's system... This means that if the model causes harm, the lab faces significant legal liability.”

Meta defends its computing power allocation strategy

A key part of the two executives' arguments was that Meta decided to use most of its computing power to serve users rather than accelerate recursive self-improvement (RSI).

“Putting the vast majority of computing power into serving humans, rather than racing for recursive self-improvement, is one of the best ways to ensure that we develop this technology safely,” Zuckerberg said.

The concern is that RSI may allow AI systems to rapidly improve their capabilities, may exceed human supervision, and increase the risk of losing control of powerful models.

Alexander Wang also described racing against RSI as “one of the most dangerous paths that could cause a strong model to lose control.”

The AI rhythm dispute polarizes the tech industry

The dispute over the pace of AI development is broadly divided: Anthropic CEO Dario Amodei called for a slowdown in cutting-edge AI development so that security mechanisms could keep up; OpenAI CEO Sam Altman, Microsoft CEO Satya Nadella, and SpaceX's Elon Musk supported a more careful and safety-focused approach without stopping development.

On the other hand, US President Donald Trump opposes slowing down the US AI race, citing fears that China will lead; former Trump AI adviser David Sachs said that companies should control their own risks rather than seek government restrictions and question Anthropic's security warnings.

Investor Michael Berry questioned whether the call to slow development was partly due to self-interest, as this could protect existing AI companies from competitive shocks.