Most AI experts are now convinced that scaling up current large language models (LLMs) is not sufficient to achieve artificial general intelligence (AGI), the level at which machines are capable of reasoning and learning as a humans.
In a new survey of 475 AI experts, 76% responded that it’s “unlikely” or “very unlikely” that existing approaches will lead us to AGI, contradicting a central assumption shared by many technology firms since the AI boom of 2022.
Scaling Is Hitting a Wall
For years, big tech firms have poured money into building bigger AI models, assuming that more data and computing power would eventually push them beyond human capabilities. But according to the survey, published by the Association for the Advancement of Artificial Intelligence, that path may now be a dead end.
“Since GPT-4, the gains from scaling have been incremental and expensive,” said Stuart Russell, a leading AI researcher at UC Berkeley. “Companies can’t afford to admit they made a mistake, so they double down.”
Russell and others argue that LLMs, based on feedforward circuits, are reaching their conceptual limits. These architectures require massive resources to simulate even basic reasoning, and often behave more like glorified lookup tables than intelligent systems.
The Data Problem
LLMs rely on huge amounts of human-generated data, which is becoming harder to find. Experts warn that by the end of this decade, most of that usable data will be exhausted. At that point, models will either turn to user data or recycle AI-generated synthetic data, which is a risky move that could degrade performance.
In 2024 alone, the generative AI sector raised $56 billion in venture capital, fueling enormous data centers and pushing energy use and emissions to new highs.
What’s Next for AI?
Despite the skepticism, experts say the future of AI isn’t doomed, it just needs a new approach. Techniques like reasoning models, probabilistic programming, and more efficient model design could open new doors.
Chinese company DeepSeek recently matched the performance of top-tier Western models at a fraction of the cost, showing there’s still room for smart engineering over brute force.
“There are many experts who think this is a bubble,” said Thomas Dietterich, a contributor to the report. “But like past tech cycles, some of today’s AI companies could still become huge success stories.”
