How uncertain is AI progress?

AI progress is highly uncertain, with many factors contributing to this unpredictability.

1. Gradual vs. Dramatic Progress: There's a spectrum of possibilities from slow, incremental advances to rapid, explosive growth. For instance, Dwarkesh Patel and Leopold Aschenbrenner discussed a scenario where AI starts by improving cognitive jobs, progressively expanding into fields like robotics and other technologies, leading to significant impacts only after an initial period of rapid AI capability improvements 1.

2. Effective Compute: Advances in AI are driven by both hardware improvements and algorithmic progress. Carl Shulman noted that effective compute for training large AI models is increasing drastically, supported by better models, increased investment, and cheaper training chips. This loop of improvements can expedite AI development, but exact timelines remain uncertain 2.

3. Government and Safety: The interplay between rapid technological advancement and regulatory frameworks is critical. Even with some level of government intervention, there are concerns about the competitive pressures leading to rushed AI deployment without adequate safety measures. This dynamic can significantly influence the trajectory and safety of AI progress 3.

   • 

     1

     AI Progression Insights

     00:29:18
     
   • 

     2

     AI Progress and Effective Compute

     00:13:07
     
   • 

     3

     AI Safety Concerns

     00:09:30
     
   • 

     4

     Deceptive Alignment Risks

     01:50:48
     
   • 

     5

     AI Predictions Unpacked

     01:16:14
     

4. Resolving Uncertainties: Experiments and research aimed at identifying and mitigating risks (like deceptive alignment in AI systems) are crucial. Carl Shulman emphasized that understanding and reducing uncertainties can enhance the effectiveness of responses to potential AI risks, making it easier to prepare for various scenarios 4.

5. Historical Predictions and Biological Anchors: Historical predictions about AI (e.g., Hans Morvek's predictions) have shown varying degrees of accuracy. The concept of biological anchors, which relates AI capabilities to biological entities, provides a heuristic for understanding progress. For example, deep learning methods started showing significant results as AI models approached the computational capacity of small animal brains 5.

Overall, the trajectory of AI progress is influenced by technological advances, regulatory environments, competitive pressures, and ongoing research, making it an inherently uncertain and complex field.