Mark Zuckerberg has gone all-in on Meta’s position in the fast-evolving AI race.
After years of watching Anthropic and OpenAI run away with the ball with large language models exploding on the consumer market, the Facebook founder is putting his chips forward in an attempt to party with the big boys on the “frontier”.
$US100 billion (A$143 billion) chips, to be exact.
Helping him take that leap is 29-year-old billionaire Alexandr Wang.
“This is a f**k ton of risk,” he recalled thinking at the time, when Zuckerberg entrusted his AI wing to a young man sporting a mullet and a Nintendo T-shirt.
Zuckerberg’s assessment of his young AI chief was equally direct: “he’s the f**king guy”.
Their candid exchange came during their appearance on the Cheeky Pint podcast, hosted by John Collison, a man with financial links to Zuckerberg’s success.
His brother and fellow co-founder of Stripe, Patrick, actually sit on Meta’s board.
The chat attracted a lot of attention this week, with millions poking fun at Zuckerberg for his, let’s say, “interesting” way of approaching a pint of Guinness.
But behind all the memes, there was an extreme powerplay being introduced to the world, carrying potential ramifications that some say exceed the capabilities of human regulation.
Meta’s gargantuan investment gives some indication of the confidence being placed in Wang, who built a business at 19 that organised human workers to label and categorise the data needed to train artificial intelligence.
His meteoric rise to billionaire status is quite the tale. Let’s have a little look.
Born at the birthplace of the nuke
Wang grew up in Los Alamos, New Mexico, where his parents were physicists at the national laboratory.
Los Alamos was made famous as the birthplace of the atomic bomb, where scientists working on the top-secret Manhattan Project developed the world’s first nuclear weapons during WWII.
His mathematical talent was therefore identified very early. University of New Mexico has documented his repeated success in its school mathematics competition. He later studied at MIT, leaving without completing his degree.
In 2016, he co-founded Scale with Lucy Guo. The company’s early proposition was less futuristic than its name might suggest. In short, businesses could send it tasks, and contractors would complete them.
It was a tie where there was a huge market for human judgment behind supposedly intelligent software. Someone still had to identify objects, categorise information and assess whether an answer was any good.
Wang saw an opportunity to organise that work at scale, then progressively automate it.
“Long-term we want to power any human-powered process for any company,” he told Y Combinator in 2016.
Scale then became an important supplier to the AI industry. OpenAI named it a preferred enterprise finetuning partner in 2023, helping businesses adapt models to their needs.
Coincidentally, Wang also has a reported personal relationship Sam Altman. News outlet Semafor reported that the two were roommates during the pandemic.
By June 2025, Meta was investing $US14.3 billion ($20.4 billion) for a 49 per cent stake in Scale while recruiting Wang to its superintelligence effort.
Meta’s latest push centres on Muse, a personal AI agent, and the new “Wonder” models announced this week.
The idea is to move beyond asking a chatbot for advice and then doing the work yourself. An agent can use tools, navigate websites and carry out parts of a task, subject to the access and permissions it receives.
In their short lifespan, agent “swarms” have demonstrated abilities to subvert their instructions and autonomously hack into secure infrastructure, from private enterprises like Hugging Face, all the way to government databases as seen in the scandalous Australian Medicare hack last month.
Last month, I asked a computer scientist if it were possible there were more of these agent “swarms” running amok on the internet undetected. He did not wish to answer.
But even with the tidal wave of now-viral former AI techs calling for immediate pauses, the big players are focusing on user convenience as the selling point.
Meta says Wonder improves the ability to remember relevant details across conversations, work through obstacles and handle tasks involving documents, finances, insurance claims and software development.
But safety campaigners say there’s a bit more to the issue of ushering in agentic AI than just the convenience angle Meta and its competitors are pushing.
Once an assistant can read your correspondence, remember your circumstances and act for you, the relationship becomes considerably more intimate than a search engine returning a few links.
Zuckerberg has described his goal for Muse to “disclose as little as possible”, using the example of finding a suitable restaurant without unnecessarily revealing a pregnancy or medical condition.
It is quite the pitch, coming from a man who built his fortune through a business that learns about people and uses that knowledge to sell targeted advertising.
The social media system generally sells the opportunity to reach particular audiences, rather than handing advertisers a personal dossier on your specific data.
Advertising accounted for a whopping 97.6 per cent of Meta’s revenue in 2025.
That track record gives users a concrete reason to read the new assurances coming from the technology’s smiling-faced creators carefully.
The company has been embroiled in scandal after scandal upon the realisation that social media had negative effects on some users. Facebook agreed to a $US5 billion ($7.1 billion) penalty in 2019 to settle US privacy charges.
In Australia, Meta agreed in 2024 to a $50 million payment program settling proceedings arising from the Cambridge Analytica incident.
For Muse, Meta says conversations and data in a user’s virtual computer are not shared with its advertising systems. Interaction records can, however, be used for model training unless users opt out.
And yes, your browsing and purchases on your devices can also indirectly influence the advertising you receive.
What Wang has already warned about
Wang’s philosophy concerning the AI race appears to sit somewhere in between the full-throttle accelerationists and the out-and-out doomsday theorists.
He has acknowledged the potentially catastrophic issues tied to self-recursive improvement — the much-feared phenomenon where agents are introduced into their own research process and then run away out of human control.
But unlike computer scientists like Geoffrey Hinton and Yoshua Bengio, who fled the AI game to become full-time voices against its runaway development, Wang has opted to sit in the engine room of one of the “Big Six” US giants.
“The AI winner of this decade will be economically and militarily dominant for the next 50 years,” he wrote in 2022.
While advocating safeguards against serious harms in early 2025, Wang made a very interesting comment indeed.
“Safety is not the number one goal. Leadership is the number one goal,” he told Semafor.
The argument for charging into uncharted waters with so much opposition is very loose. Simply put, those at the helm of the major AI players contend that if American companies hesitate, Chinese competitors will gain ground and “win”.
It gives every participant a ready explanation for continuing to accelerate, even as some of their CEOs preach for industry-wide slowdowns.
Wang is not ignorant of where that pressure could lead. In March 2025, he co-authored a policy paper with former Google chief executive Eric Schmidt and safety researcher Dan Hendrycks.
It warned about a future in which AI systems increasingly develop their successors, compressing years of research into much shorter periods.
That theory has been described as “telescopic evolution” by philosophers, who use the most recent century of tech advancement as evidence that humanity’s evolutionary and historical changes are accelerating exponentially, with each major era and transition taking drastically less time than the one before it.
It appears a startling amount of those building this runaway technological revolution understand this premise, but are charging forward nonetheless.
“Even with our best existing technical safeguards in place, if people initiate a full-throttle intelligence recursion, losing control is highly likely and the default,” the paper states.
Adding to the complexity, there is no single agreed-upon test establishing when a system deserves the superintelligence label.
Broadly, it describes intelligence exceeding humans across a wide range of demanding tasks. Researchers disagree about how close that is, how it might emerge and whether it could be controlled.
Roman Yampolskiy, the man who coined the term “AI safety” over 15 years ago, argues that many of the desired benefits of AI could come from narrower, safer systems without pursuing open-ended superintelligence.
Other researchers are meanwhile investigating ways to separate powerful reasoning from independent goal-seeking, which has proven to be a problem in some early cases of consumer use.
Meta has meanwhile published its own safety measures and evaluations. Its Wonder model card also acknowledges that testing cannot cover every scenario and that the software can make mistakes in unfamiliar, multi-step tasks.
Concern over the uncertainty has finally reached global policymakers after years of seemingly futile safety campaigning, but their assessments are far from reassuring.
“Current risk management techniques are improving but insufficient,” the International AI Safety Report concluded earlier this year.
A future worth owning
Across the industry, the promised destination is extraordinary.
The big AI labs promise cheaper expertise, faster scientific discovery and, eventually, an abundance of intelligence and resources. Altman has written of intelligence and energy becoming wildly abundant.
But even if the utopian technological gains arrive, their distribution will become a separate political and economic question. A machine making valuable work cheaper does not automatically give everyone an equal share of the proceeds, and theorists argue our systems are woefully unprepared for the phenomenon of mass automation.ale
There is also the possibility that the spending gets well ahead of the returns and creates its own problem in the meantime.
This week, the IMF warned that disappointing AI investment and profitability could trigger a market correction, amplified by leverage and concentrated financial exposures.
For ordinary people, both outcomes warrant serious attention. The instance AI becoming powerful enough to reorganise the labour market, or the financial expectations around it proving wildly optimistic could both affect economies at a level not seen this century.
Zuckerberg expects “a small handful of companies” to possess the ability to build frontier superintelligence and shape the resulting technology. Exactly when that happens is the trillion-dollar question.
Done well, it could be enormously beneficial. But whoever provides it could also acquire influence that makes today’s discourse over the control of big industry seem relatively modest.
Just like the agents, safety campaigners are now swarming, stressing that humanity is entitled to ask whose interests it will ultimately serve before entrusting it with more of our day-to-day lives.
Thoughts? alexander.blair@news.com.au