Scaling Limits Threaten AI Boom, Expert Warns of Looming Bust

Compiled by the editorial desk with reference to reports from The Information and Reuters, and public commentary by Gary Marcus.

The era of rapid, predictable gains in artificial intelligence may be drawing to a close, as mounting evidence suggests that simply feeding larger models more data and computing power no longer yields the dramatic improvements that fueled the industry's explosive growth. This realization is prompting warnings from prominent skeptics that the financial underpinnings of the AI boom could be dangerously fragile.

Gary Marcus, a cognitive scientist and long-time critic of the current AI trajectory, argues that the economic foundation of the industry is shaky. In a recent Substack post, he wrote that the sky-high valuations of companies like OpenAI and Microsoft are predicated on the belief that continued scaling will eventually lead to artificial general intelligence (AGI). He called that notion "just a fantasy."

The concern is not merely theoretical. A report from The Information revealed that OpenAI's upcoming flagship model, code-named Orion, has shown significantly less improvement over its predecessor, GPT-4, than GPT-4 showed over GPT-3. In some areas, particularly coding—a major selling point for these models—there may be no improvement at all.

This anecdote reflects a broader industry trend. Ilya Sutskever, co-founder of Safe Superintelligence and former chief science officer at OpenAI, told Reuters that the gains from scaling AI models have plateaued. The long-held dogma that "bigger is better" is being challenged, and with it, the rationale for the massive capital expenditures that have defined the sector.

Why the Economics Are Grim

The financial implications are stark. Training runs for large models can cost tens of millions of dollars, require hundreds of specialized AI chips, and take months to complete. Moreover, tech companies have already scraped most of the freely available data on the internet, leaving a limited supply to feed future models. Marcus predicts that as these costs rise and improvements stagnate, LLMs will become commoditized, leading to price wars and elusive profits. "When everyone realizes this, the financial bubble may burst quickly," he warned.

However, the industry is not without potential escape routes. OpenAI researchers are exploring alternative methods to improve model performance, such as training models to "reason" more like humans, a capability previewed in the o1 model. One technique, called "test-time compute," involves the model evaluating multiple potential solutions to complex problems before selecting the most promising one, rather than jumping to a conclusion.

Whether these approaches can rekindle the rapid progress that investors have come to expect remains an open question. The AI industry continues to grapple with a fundamental profitability problem, and as markets are notoriously impatient, the window for demonstrating a new path forward may be narrow. If these improvements do not materialize quickly, the sector could face a prolonged downturn, reminiscent of the AI winters of the past.