Bernie vs. Claude

[HesNotHere]

8. Large Language Models are really non-deterministic (semi-random) number generators

The numbers its generates--the most likely number based on the current "context"--are decoded into words that make sense to us.

The "context" is the important part. It is being built in these tools from both the users historical interactions with these tools, as well as the phraseology they use to interact with them. Those who use certain keywords and speech patterns associated with political ideology will generate patterns that shift the context, and produce different results.

Train an LLM with a large enough corpus, and it will be sensitive to patterns that distinguish users between education level, politics, gender and even mental conditions.

It can be extrapolated from this that with enough time, everyone's interaction becomes sycophantic to a large degree. When people post responses from questions they asked an LLM, it gets really silly; it reveals more about what is in their head than what is based on facts.