The Evolutionary Path of AI From Compression to Super-intelligence
Intelligence as Compression
Intelligence can be thought of as the ability to simplify and understand the world. For example, the laws of physics take all the complicated stuff happening around us and simplify it into easy-to-understand rules. This makes it easier to predict what will happen next. Some researchers, like OpenAI’s co-founder Ilya Sutskever, have hinted at this idea in their work on AI and unsupervised learning. You can read more about it here or watch his talk here.
AI and Truth-Seeking
When we’re trying to figure out the truth, it’s like a competition to see who can find the best way to simplify (or compress) information. AI models like chatbots can’t create entirely new ideas; they can only work with the data they’ve been trained on, which limits their ability to come up with truly new insights. This limitation is discussed in more detail here.
Why Real-World Experience Matters for AI
To get smarter, AI needs to experiment and learn from the real world, just like humans do. Current AI mostly learns from existing data, which means it’s not getting the hands-on experience needed to make big leaps in understanding. This idea is explored further in research on embodied AI, like in this article here.
The Evolution of AI
Before we reach super-intelligent AI (like a super-smart human), AI might need to go through stages similar to how humans evolved from simpler organisms. Just as life evolved from single-celled organisms to primates and eventually to humans, AI could first develop in stages. You can explore the timeline of AI’s development here.
The “Primates Stage” of AI
In this evolutionary analogy, the next stage of AI might resemble the “primates stage” in human evolution. In this stage, AI could take the form of personal assistant robots that are much smarter and more capable than current AI. These robots would be able to learn and interact with the world, much like how early primates developed advanced cognitive abilities. Eventually, these robots could evolve into even more sophisticated versions, paralleling how early primates evolved into humans.
For a visual exploration of this concept, you can check out this video, which shows the development of AI and robotics in a similar context.
However, the limiting factors in this evolutionary process are the computational power required to run such advanced AI and the financial resources needed to build and deploy robots at scale. These constraints mean that progress might be slower than we’d like, but as technology advances and costs decrease, these barriers will gradually diminish. The financial and computational challenges of AI development are discussed in this paper here.
TL;DR:
- Intelligence is compression.
- Better compression leads to better prediction.
- Better prediction comes from better abstraction.
- The laws of physics are a good example of how simplifying (compressing) information allows for better prediction.
- AI might evolve through stages similar to human evolution, starting with simpler forms like “cells” and eventually reaching a “primates stage” with smarter, more interactive robots before becoming fully human-like.
- The speed of this evolution depends heavily on available computational power and the money needed to build and deploy advanced robots.