Beyond big models: Why AI needs more than just scale to reach AGI

The idea of artificial general intelligence—machines that can think, learn and reason as well as humans—has captivated scientists, entrepreneurs and science-fiction writers alike. Industry leaders like OpenAI’s Sam Altman and Google DeepMind’s Demis Hassabis suggest that AGI could be within reach, powered by the relentless scaling of neural networks. The thinking goes: the bigger the model, the smarter the AI.

But a new survey of AI experts reveals a growing skepticism of this idea. While today’s AI models can generate fluent text, recognize images and even perform complex problem-solving tasks, they still fall short of human intelligence in key ways. Most surveyed AI researchers believe that deep learning alone isn’t enough to reach AGI. Instead, they argue that AI must integrate structured reasoning and a deeper understanding of cause and effect.

IBM Fellow Francesca Rossi, past president of the Association of the Advancement for Artificial Intelligence, which published the survey, is among the experts who question whether bigger models will ever be enough. “We’ve made huge advances, but AI still struggles with fundamental reasoning tasks,” Rossi tells IBM Think. “To get anywhere near AGI, models need to truly understand, not just predict.”

For years, AI research has largely followed a single formula: more data, bigger models, better results. This approach has fueled major breakthroughs in generative AI, with models like ChatGPT, Gemini and Claude reaching new milestones in conversational ability.

But as Rossi sees it, simply making neural networks larger won’t solve AI’s most fundamental limitations: today’s LLMs still struggle with logical reasoning, consistency and adaptability. They produce fluent text, but sometimes contradict themselves, make factual errors or fail to generalize knowledge across domains. And the weaknesses become especially apparent in areas like mathematics or other domains where reliability and correctness are crucial.

“Some AI models can now solve complex, gold-medal-level math Olympiad problems,” Rossi says. “But at the same time, they still fail at simple arithmetic that any elementary school student could handle. That inconsistency is a clear limitation of current deep learning approaches.”

The core problem, Rossi says, is that while neural networks can recognize statistical patterns, they don’t inherently understand concepts. They can generate responses that sound correct, without truly grasping the meaning behind them. And without an explicit structure for logic and reasoning, AI remains prone to hallucinations, and unreliable decision-making.

Rossi and some other researchers believe that if AGI is ever achieved, it won’t be through deep learning alone. Instead, it’ll be through combining neural networks with structured reasoning, improving its ability to think, adapt and apply knowledge across different scenarios.

“Humans don’t just rely on instinct; we also reason deliberately,” Rossi says. “We reflect, use structured rules and think through complex problems. AI could benefit from a similar balance.”

IBM has been exploring this idea through its "Thinking Fast and Slow" AI research project, inspired by Nobel laureate Daniel Kahneman’s work on human cognition. The project focuses on integrating fast, intuitive decision-making with slower, more deliberate reasoning.

“In our project, we’re exploring how AI can combine quick pattern recognition with explicit, rule-based reasoning,” Rossi explains. “For example, a language model might generate an answer, but a symbolic AI component could validate whether the response makes logical sense.”

This hybrid approach could help solve one of AI’s biggest challenges: its tendency to generate confident yet incorrect answers. By introducing structured reasoning, IBM is helping to make AI systems more reliable, interpretable and adaptable to complex, real-world tasks.

If AGI isn’t just a matter of making neural networks bigger, what will it take? Rossi believes AI must develop a richer, more structured understanding of the world, rather than relying solely on statistical correlations.

Right now, AI models excel at producing human-like text. But they don’t really comprehend the world in a deep way. They lack a conceptual model of reality, so they often struggle when asked to apply reasoning to unfamiliar problems.

“We need to think beyond just scaling models and focus on how intelligence works,” Rossi says. “That means integrating different approaches, not just making models bigger.”

Some researchers argue that AI must learn from multiple data types, not just text. AI could build a broader understanding of cause and effect by incorporating visual, auditory and real-world interactions. Others suggest that AI will need explicit reasoning tools, allowing it to apply logic rather than simply predicting likely responses.

One of the biggest debates in AI isn’t just how to achieve AGI—it’s whether AGI is even a useful concept. Some believe that AGI will emerge gradually as AI systems become more capable across different domains. Others argue that true AGI—AI that can fully think, reason and act like a human—is still far off.

Rossi is cautious about the term itself. “Different people mean different things when they talk about AGI,” she said. “If AGI implies replacement, we believe AI should augment human intelligence, not replace it.”