AGI Is Not Viable; Focus on Developing Superhuman Adaptable Intelligence Instead

In a recent episode of the Dr. Data Show, co-host Luba Glouhova and I delved into an intriguing research paper by AI expert Yann LeCun and his team. We were alerted to this paper by fellow researcher Philippe Wyder, who connected with us on social media, highlighting its relevance to our past discussions and suggesting it provides a fresh perspective on what we’ve termed the “muddled AGI discourse” by focusing on specialization.

The paper indicates a significant shift in the AI landscape: as the long-held aspiration of achieving artificial general intelligence (AGI) has proven to be an unclear and overly ambitious target, the authors propose a more defined objective—"superhuman adaptable intelligence" (SAI).

Understanding Superhuman Adaptable Intelligence

So, what exactly is SAI? The authors define it in two key dimensions. First, SAI possesses the ability to outperform humans in any task—while emphasizing the importance of focusing on one specific task at a time. Second, it can adapt to effectively tackle tasks that lie outside human capabilities.

This new framework sets an important benchmark for the AI community. Rather than pursuing an all-encompassing solution that attempts to replicate every human skill, as AGI aims to do, the focus shifts back to specialized, narrowly defined AI solutions. The researchers advocate for utilizing large datasets through self-supervised learning—similar to methods used by large language models—while also stressing the need to tailor these technologies to solve defined problems.

During our podcast, Luba highlighted a critical point made by the authors: human intelligence itself is far from general. I had a similar realization back in 1991, just before embarking on my Ph.D. studies at Columbia, when I observed that human abilities are specialized and developed through millions of years of evolution in specific contexts. Instead of trying to replicate the complex range of human skills—an enormously challenging task—this paper suggests focusing on task-specific solutions: "The AI that folds our proteins should not be the AI that folds our laundry." Even as our computational capabilities grow, systems greatly benefit from specialization for individual tasks. Allocating resources toward precise, meaningful goals is far more effective than attempting to create a universally competent machine.

Is This Just Another Trend?

However, can introducing a new term like this genuinely address the AI industry's persistent overhyping, particularly the controversial promise of swiftly arriving at AGI? I have my doubts. The inclusion of terms like "superhuman" and "intelligence" still evokes the science-fiction narrative associated with AI since its inception in the 1950s. This terminology enables figures like LeCun to maintain credibility without engaging with the inherent contradictions of AGI while still retaining enough appeal to attract significant venture capital investment—his startup recently secured a staggering $1 billion seed funding.

As Luba and I lightheartedly noted during our podcast, SAI could be regarded as a new trend in AI discourse—but at least it’s a more balanced and healthier alternative. The defining aspect of "adaptable" within SAI emphasizes a practical, quantifiable benchmark: the speed with which a system can adapt to excel in a defined task.

In my view, this paper is a positive development. It offers a much-needed reality check on the often unrealistic aspirations prevailing in our field. By steering the conversation away from autonomous human-like AI agents and towards technologies that offer tangible value, it lays a foundation for a more grounded approach. This ideology resonates well with my advocacy for hybrid AI systems. To ensure generative AI products are ready for large-scale application that captures real value, each initiative should be treated with a specialized focus, applying relevant reliability measures to effectively address specific challenges.