Mouth vs Brain
A couple of days ago I read 2 articles about LLM (large language model) and ChatGPT. The first one is How ChatGPT and Other LLMs Work—and Where They Could Go Next on Wired. It mainly discussed the following topic:
- The main idea behind the LLM: with some user-provided words (usually a quesion), LLM choose the most suitable following without understanding the meaning of the question and the sentence it emits.
- Technical lineage: neural network to transformer, then to LLM.
- How human participation promotes the model quality. Here the article introduced the basic idea of RLHF (reinforcement learning on human feedback), the silver bullet of ChatGPT.
The second article How ChatGPT actually works contains far more technical detals. It explains why ChatGPT-4 is superior than its precursors (better alignment), the detailed steps of RLHF, how to evaluate the performance of a LLM, and finally the shortcomings of the methodology used in current model training.
After reading above articles, I tend to believe that LLM nowadays is more like the mouth rather than the brain of AGI (artificial general intelligence), like Blade Runner or agent Smith in The Matrix.
LLM read gigantic amout of human-generated materials, and learn the patterns in them. It mixes and regenerates new texts (instead of ideas) based on the texts (instead of ideas) it learned. However, text is only avatar of idea, not the idea itself.
People argue that there’s no such thing of “idea”, or “intelligence”. These are just imagination of human being. When a system evolves large enough, “intelligence” emerges. According to this interpretation of intelligence, the advent of AGI happens natrually with the increasing of the complexity of the LLM. Well, it’s reasonable. But at least until now, I didn’t see the possibility.
Comments
Post a Comment