Meta's Chief AI Officer Alexandr Wang Says AI Leaders Are All Very Smart
Meta's Chief AI Officer Alexandr Wang told tech journalist Cleo Abram on October 7 that AI leaders are all very smart, but personal grudges are shaping how this technology is developed. His exact words were that everyone must put aside their differences and their respective histories.
The interview was featured on Abram's YouTube show "HUGE* Conversations," titled "The Biggest Launch Since ChatGPT," where she asked him which side of the safety debate he stands on as one of the responsible parties. Abram noted that his response was unexpected. The interview also discussed Meta's agent product Muse: what it can do, when it can be trusted, and how to determine if it is on the right track.
Wang summarized the competitive logic as: "I am responsible, those people are not, so I cannot slow down." He stated that if you change the subject to any lab, any country, or any competitor of the week, the result is the same; everyone still claims to be the responsible party. When discussing China, he mentioned that he has changed his mind and that outsiders have exaggerated the portrayal of the other side as a monster.
He described Muse as an architecture, not just a slogan. Each agent runs in an independent secure virtual machine, and the data shared only exists there; when filling out forms or making calls externally, another sentinel agent checks and intercepts sensitive information like credit card numbers and social security numbers. Users must approve access to new websites or data sharing. A more secure virtual machine is planned, aiming to approach the guarantees of WhatsApp, though it may come at the cost of slower performance. Conversations and data within the virtual machine are not sent to Meta's advertising system.
The other side of the safety debate was also addressed in this episode. Wang stated that no one knows exactly how to manage systems that are much smarter than humans; supervision is possible when humans and models are similarly intelligent, but it becomes incomprehensible when the gap is too wide. He remains optimistic, believing that the process will be bumpy but progress is visible. The program did not provide a ceasefire list between labs.
This is a redistribution of discourse power, not new financing. The sellers are lab heads who harvest safety debates into personal grudges, while the buyers are regulators and clients deciding whether to slow down or continue accelerating. The beneficiaries are platforms that can turn trust into virtual machines, sentinels, and approval switches, while those under pressure are labs that still defend acceleration by saying "the other side is irresponsible, so I cannot stop." Meta is selling an auditable wall, not a joint statement.
Source: Public Information
ABAB AI Insight
Wang's position has changed, and his rhetoric has followed suit. In 2016, he and Lucy Guo founded Scale, selling data that others were unwilling to label; in June 2025, Meta acquired about 49% of Scale for approximately $14.3 billion, and he became Chief AI Officer. The safety debate used to be about client questionnaires, but now he must personally sign off on Muse's external actions. The over-execution from the Scale era has transformed into a product structure of virtual machines and sentinels.
Money has already been spent on computing power and talent, and trust is being turned into a switch. Before Muse can send out data, it requires human approval, and sentinels intercept identification numbers, while the more secure virtual machine aims for WhatsApp-like guarantees. This is transforming the safety commitments made in policy texts by OpenAI and Anthropic into an auditable system wall. If personal grudges continue to dictate how models are trained, Meta will use approval flows to keep clients within its agents rather than waiting for a ceasefire between labs.
A comparison can be made to the founding of OpenAI by Musk and Altman around 2015, followed by splits in the board, equity, and safety statements; Anthropic also split from OpenAI, reflecting the same historical figures dismantling labs. Wang specifically called for putting aside this history of division. The industry is at the tail end of an expansion phase: agents are now capable of booking tickets, filling out forms, and spending money, while supervision methods remain within human comprehension.
The essence is that regulatory changes have not yet arrived, but pricing power is shifting first. The mechanism is that each company claims to be responsible while blaming the other for irresponsibility, so no one slows down. Wang's effort to pull China out of the monster narrative is aimed at dismantling this excuse. Trust is no longer priced based on the character of the responsible party, but rather on whether sentinels can intercept a credit card. The stronger the incomprehensible systems become, the more people are willing to pay for a wall that can be turned off.
ABAB News · Cognitive Laws
Everyone is responsible, so no one slows down.
Grudges can set the course, but cannot establish trust.
Incomprehensible systems sell based on turn-off walls.