I mean the input data on the pretrain stage.
It for example generates many misconceptions seen on the internet that aren’t factual, especially on topics where OpenAI/Anthropic/whoever may not have fine tuned the model on that field.
The optimal solution for the model is to answer with what is the most common (as that is the only way to minimize pretrain loss) and since it is the pretrain, it is essentially baked into the model.
For example, the model may decide that it is implausible for someone with higher strength than the perpetrator to have a valid claim.
The input context to the model is also finite (and it may ignore a whole bunch of it), leading to assumptions being made that are incorrect.
The capability to make correct judgements would also require the capability to effectively simulate and understand the real world, which is a different task than predicting the next token.
I could go on and on with issues about this concept..
