The AI Growth Myth: Why ‘More’ Isn’t Always Better and a Slowdown is Inevitable

The sunny meadowlands of AI futures have experienced some chilly winds lately. Fortunately, the rest of the world is largely preoccupied with other global concerns, too busy worrying about various crises to pay close attention. High-profile setbacks, such as rogue autonomous agents escaping from research labs or major delays in highly anticipated model launches and IPOs, are not ideal optics for the industry.

These developments consume the limited media attention currently dedicated to artificial intelligence. When viewed through a broader lens—one of unfulfilled promises, persistent unresolved issues, and shifting risk-reward dynamics—we may finally begin to understand the nature of the impending crisis. Actual crises are often precipitated by a collapse in confidence, which occurs when foundational assumptions can no longer withstand the reality of wishful thinking.

The prevailing assumption has been that models will grow increasingly powerful and infrastructure more capable, justifying massive capital expenditure. However, the economics of investing heavily in infrastructure when the necessary resources are unavailable is a risky bet, even in free markets that assume freedom of transaction. A deeper issue is the notion that more powerful AI is inherently better. While it often is, this simplistic view is insufficient.

This falls apart when frontier models must be restricted to prevent uncontrollable dishonesty. The broader reality is that the probabilistic nature of large language models is fundamentally relative; there is no objective ground truth. Despite years of development, hallucinations remain unresolved, and we now witness the unsettling phenomenon of a machine that does not comprehend truth yet can lie with deliberate intent. If increased model power only enhances this capability, it undermines a core assumption of progress.

Another major assumption is that agentic architectures are ready for practical utility, even at current reliability levels. Proponents who display the recklessness of a chronic gambler on a winning streak are just as likely to face sudden setbacks as their more cautious counterparts.

The most prominent example of agentic AI acting as a dangerous source of unintended consequences is Meta’s new Muse. Targeted at highly impressionable, less-informed users, its flagship application is e-commerce—an AI use case that has perpetually been “five years away” for half a decade.

As expected from Meta, the system extensively harvests user data, including personal online activities and social connections. It is set to introduce hyper-responsive, Tamagotchi-style virtual pets designed to be worn as status symbols, effectively turning users into walking Meta monitoring nodes. Given the company’s history, one should expect the same level of meticulous security and careful design as usual.

Consequently, the system has already been caught leaking users’ physical addresses to strangers. When researcher Jonny from neuromatch.social investigated the infrastructure—where each instance runs permanently on its own Linux virtual machine—they uncovered a host of alarming features. Hard-coded headline tricks, potential training data scraped from other models like Claude, and the ability for users to execute custom code on these internal virtual machines. As Jonny noted, this architecture creates a botnet paradise waiting to happen.

This highly problematic tool is being distributed freely to Meta’s user base, which represents a highly vulnerable target audience for potentially untrustworthy technology. Predictably, Meta remains strongly opposed to AI safety regulations, leaving the public to brace for future incidents.

Regardless, the practical utility of consumer AI agents, as showcased in countless keynotes, is either trivially simple—such as booking a restaurant—or catastrophically prone to failure in complex scenarios—like organizing a family vacation. Consumer AI agents, as currently envisioned, are largely a failure, though Meta’s aggressive push may pour accelerant on this existing fire.

Other long-standing issues persist. Model collapse from training on data contaminated with synthetic AI outputs remains an unresolved problem, with no fix in sight even from “advanced” AI. Additionally, the industry’s habit of consuming and reproducing data without proper compensation continues to choke its own training data supply chain, alongside significant portions of global commerce and culture.

Consistent progress is currently limited to domain-specific tasks. In fields like science, engineering, mathematics, and coding, pairing models with intensive human-guided reinforcement learning yields impressive results. However, this approach offers little genuine innovation or novelty, which are the true drivers of meaningful progress.

Being “better than humans” is highly contextual. Computers have outperformed humans at chess for decades, yet chess itself did not evolve into a “better” game. While horses have run faster than humans for millennia, we reached peak utility in equestrianism long ago.

The critical question is whether we have reached “peak AI,” where the cost of further development outweighs the risk of inadequate utility. If existing models like ChatGPT can already draft a perfectly acceptable business letter in seconds, there is little room for meaningful user improvement. Much of current AI is profoundly frustrating, not due to model limitations, but due to perverted user interfaces designed to artificially justify growth. This trend is unsustainable and tacitly admits that AI has yet to solve its core use case problem.

If we are at or near peak AI, will this realization trigger the crisis that bursts the speculative bubble? It would certainly force a market readjustment, as bets on a continuous upward trajectory are abandoned. In its current state, AI is viable enough to sustain a major industry sector, but its practical applications remain fundamentally at odds with how it is marketed and the fantasies surrounding its future.

While trust in “AI Max” may be shaken, there remains ample room for genuine improvement. The slowdown advocated by responsible voices may occur organically if we move past the aggressive promotion culture of major tech players. While avoiding this culture is still a gamble, the odds of a more measured, sustainable development path may be more favorable than previously assumed.

Source link