Meta's AI Strategy

Mark Zuckerberg has pulled out all the stops in pursuit of his AI ambition. Over the past few months, Zuckerberg has poached leading AI researchers from top rivals like OpenAI and Anthropic, offering jaw-dropping pay packages that range from millions to over a billion dollars. Meta bought 49% of the data labeling company Scale AI and appointed its CEO, Alexandr Wang, to lead Meta’s new Superintelligence initiative. Meta is estimated to spend a record $66-72 billion on AI Capex in 2025, including building multi-gigawatt datacenters that rival the size of Manhattan.

Why is Mark Zuckerberg, who has a dictator-like control over his company, going all in on AI? There are half a dozen companies with models better than Meta’s best, and its recent attempt to catch up failed miserably. What’s the reasoning behind pouring hundreds of billions of dollars into building new models, at a time when frontier models are more commoditized than ever?

Many people think Mark Zuckerberg is deluded. That AI is a bubble that’s about to pop, and that Meta will go down with it. That paying AI researchers fresh out of college hundreds of millions to billions of dollars is a waste of money. That just like the companies’ bet on the metaverse; betting big on AI will not pay off.

I disagree. I think Zuckerberg going all in on AI makes perfect sense when you understand what he really wants Superintelligence to do.

The biggest players in AI today are singularly focused on creating “artificial superintelligence”.

The “Big Three” AI labs of today are OpenAI, Anthropic, and DeepMind. These companies have it all: teams consisting of the smartest humans alive, data centers with hundreds of thousands of GPUs, and hundreds of millions of users.

Since the rise of ChatGPT, these three labs have routinely pushed the frontier of AI capabilities. OpenAI currently has the smartest models in the world, including one that recently won a gold medal at the International Math Olympiad, arguably the hardest math competition in the world. DeepMind has made breakthroughs in protein folding that recently won its CEO a Nobel prize. Anthropic’s best AI models are at the frontier of coding ability and are used by millions of developers to accelerate their productivity.

Why are these companies focusing so heavily on math, science, and coding tasks? The answer is this: these labs are trying to make their AIs really good at doing AI research, with the end goal of achieving an “intelligence explosion” with superintelligence on the other side.

The idea is that once AI is really good at doing AI R&D, said AI will be able to make the next generation of AI even better at AI R&D, and so on. To some, this is thought to result in an exponential rise in AI capabilities, with each generation of genius AI models making the next generation even smarter. After this intelligence explosion, a superintelligent AI (defined as an AI system that is far better than every human at every cognitive task) would emerge.

Once we have this hypothetical superintelligence, leaders of these AI Labs say, a tsunami of prosperity and abundance will flood the earth, and the sky will turn gold. The primordial sufferings that have shadowed humanity since its beginnings— plague, famine, and death —will be extinguished as this wave of intelligence rolls over the world. The deepest mysteries of existence will be torn from the cosmos and handed down to us on a silver platter. We (or the AI) will take to the stars and spread our consciousness throughout the light cone.

Mark Zuckerberg has a different idea of what an Artificial Superintelligence should do.

In a recent statement, he says

As profound as the abundance produced by AI may one day be, an even more meaningful impact on our lives will likely come from everyone having a personal superintelligence that helps you achieve your goals, create what you want to see in the world, experience any adventure, be a better friend to those you care about, and grow to become the person you aspire to be.

Meta's vision is to bring personal superintelligence to everyone. We believe in putting this power in people's hands to direct it towards what they value in their own lives.

This is distinct from others in the industry who believe superintelligence should be directed centrally towards automating all valuable work, and then humanity will live on a dole of its output.

Instead of directing the Artificial Intelligence, that will plausibly one day be smarter than all of humanity combined, at speeding up technological progress, at ending poverty, or at solving any one of the world’s biggest issues, Mark Zuckerberg wants to direct Superintelligence at us.

At first, this may sound wonderful! Could you imagine the benefits of having a superintelligent tutor, mentor, guide, and therapist in your pocket 24/7? How could anyone see anything bad with this future?

But there's a catch: who decides what your "personal" superintelligence actually optimizes for? Will the AI optimize for its users’ well-being and fulfillment, or corporate profits? To answer that, we need to look at how Meta's current products actually work.

In order to predict whether personal superintelligence will be aligned with the user rather than profits, let's first see how well Meta’s current products pass the test.

But before we get into that, there’s a crucial fact that you have to understand. 99%, or $163+ billion of Meta’s revenue comes from advertising. Advertising, meaning the ads you see as you scroll on Facebook or Instagram.

The amount of money Facebook makes by showing ads to you is directly proportional to the number of ads you see. The number of ads you see is directly proportional to the time spent on their apps. Thus, it’s in Instagram’s and Facebook’s best interest to keep you on their app as long as possible.

This point warrants restating. The algorithms that drive content recommendation on Facebook and Instagram aren’t tuned to maximize your well-being, fulfillment, or general happiness. These algorithms are optimized to maximize the time you spend on those apps, i.e, to be as addictive as possible.

(By the way, the algorithms used by Meta for content recommendation are also “AI”, just not language models.)

This is why Mark Zuckerberg’s desire to give everyone personal access to an artificial superintelligence is so scary. Meta has a large incentive to make talking to the AI as addictive as possible. Imagine a digital friend with superhuman charisma and likeability, who is always there for you 24/7. One that always tells you exactly what you want to hear, with the goal of keeping you hooked rather than actually improving your life.

Or a digital romantic partner, one that’s hyper attractive and far funnier than any human you have ever met in your life, a being whom interacting with is like listening to the Siren’s song.

Or a recommendation system algorithm, that doesn’t just show you content, but can create tailor-made photos and videos on the fly, pieces of media that it KNOWS you will love.

Remember, the AI systems Meta wants to create aren’t just “really smart”. A superintelligent AI is defined as an AI that is far better than every human at every cognitive task. Better than the best salesman in the world at selling. Better than the best persuader in the world at persuading. Better than the best content creators walking the planet at creating addictive content.

The internet addiction problems of today will be nothing compared to the digital heroin of the future. Meta will leverage superintelligence, not to cure cancer or to end poverty, but to get us even more addicted to their apps. Not because of malice or malevolence, but because that’s what makes the company the most money.