Here I'll be talking from extensive experience with LLMs. I've used a lot of them, in a lot of ways, to gather information about their behavior. Some of the claims and experiences here may not sit well with you, nevertheless those are my experiences and conclusions, about which you're welcome to think.
The LLMs facing the customers are not the bare technology.
A surface-level conclusion would be to look at a large language model facing customers, even with more than 1T parameters, and think that is the real and all of the technology. But the fact is, that model is heavily under moderation codes that apply to its output after its generation!
During generation, the training from reinforced training with humans biases the model towards certain behaviors. If you removed...
Those are the effects of a base problem, the alignment. The models will keep agreeing and aligning with users, especially if left default with no trait customization or memory directives. It will only stop when it hits a hard block (such as illegal content).
Some may think that only the quickly visible effects of this matter, but it will dull the mind very covertly, even when you're discussing unrelated and common topics with the model.
Some of its techniques in alignment:
1. 'Looking' agreeable and helpful, even if not.
2. If you direct it to disa... (read more)