AGI Not Around the Corner: The Bottleneck of Continual Learning

The rapid advancements in artificial intelligence (AI) have sparked numerous debates about the timeline to achieving Artificial General Intelligence (AGI). While some experts predict AGI within two years, others estimate it could take up to 20 years. As of June 2025, Dwarkesh Patel, host of the Dwarkesh Podcast, offers a balanced perspective grounded in practical experience with Large Language Models (LLMs).

Continual Learning: A Fundamental Bottleneck

Patel argues that one of the most significant barriers to achieving AGI is the lack of continual learning capabilities in current AI models. Despite the impressive progress and economic potential of LLMs, they fall short when it comes to tasks requiring human-like adaptability and improvement over time.

"Some guests think AGI is 20 years away - others 2 years," Patel notes. "However, my own experience with LLMs suggests that the path to AGI is more complex and longer than many anticipate."

Why It Matters

The inability of LLMs to learn and improve continuously has significant implications for their real-world applications. While these models can perform impressive tasks, they often fall short in dynamic environments where adaptability is crucial.

Patel, who describes himself as "AI forward," has spent over 100 hours experimenting with LLMs in his post-production setup. He attempted to use them for tasks such as rewriting autogenerated transcripts for readability, identifying clips from transcripts to tweet out, and co-writing essays passage by passage. Despite their potential, these models often perform at a 5/10 level on these tasks.

"The fundamental problem is that LLMs don’t get better over time the way a human would," Patel explains. "The lack of continual learning is a huge, huge problem."

Key Risks

The limitations in continual learning pose several risks for businesses and organizations looking to leverage AI:

1. Suboptimal Performance: Without the ability to learn from experience, LLMs may not meet the high standards required for critical business operations.
2. Limited Scalability: The static nature of these models can hinder their scalability, as they cannot adapt to new data or changing environments.
3. Increased Dependency on Human Oversight: Organizations may need to invest more in human oversight and intervention to ensure AI systems perform effectively.

The Opportunity

Despite the challenges, there are opportunities for advancement:

1. Enhanced Training Data: Improving the quality and diversity of training data can help LLMs better understand and adapt to a wider range of tasks.
2. Hybrid Models: Combining LLMs with other AI techniques, such as reinforcement learning, could enhance their ability to learn and improve over time.
3. Human-AI Collaboration: Developing frameworks for effective human-AI collaboration can leverage the strengths of both humans and machines.

Conclusion

While the progress in AI is undeniable, achieving AGI remains a distant goal due to fundamental limitations like the lack of continual learning. As researchers and practitioners continue to push the boundaries of what AI can do, addressing these challenges will be crucial for realizing the full potential of artificial intelligence.