| >What a lot of people actually want from an LLM, is for the LLM to have an opinion about the question being asked. That's exactly what they give you. Some opinions are from the devs, as post-training is a very controlled process and basically involves injecting carefully measured opinions into the model, giving it an engineered personality. Some opinions are what the model randomly collapsed into during the post-training. (see e.g. R1-Zero) >they seem to be capable of dealing with nuance and gray areas, providing insight, and using logic to reach a conclusion from ambiguous data. Logic and nuance are orthogonal to opinions. Opinion is a concrete preference in an ambiguous situation with multiple possible outcomes. >without any consistency around what that opinion is, because it is simply a manifestation of sampling a probability distribution, not the result of logic. Not really, all post-trained models are mode-collapsed in practice. Try instructing any model to name a random color a hundred times and you'll be surprised that it consistently chooses 2-3 colors, despite technically using random sampling. That's opinion. That's also the reason why LLMs suck at creative writing, they lack conceptual and grammatical variety - you always get more or less the same output for the same input, and they always converge on the same stereotypes and patterns. You might be thinking about base models, they actually do follow their training distribution and they're really random and inconsistent, making ambiguous completions different each time. Although what is considered a base model is not always clear with recent training strategies. And yes, LLMs are capable of using logic, of course. >And what most people call sycophancy is that, as a result of this statistical construction, the LLM tends to reinforce the opinions, biases, or even factual errors, that it picks up on in the prompt or conversation history. That's not a result of their statistical nature, it's a complex mixture of training, insufficient nuance, and poorly researched phenomena such as in-context learning. For example GPT-5.0 has a very different bias purposefully trained in, it tends to always contradict and disagree with the user. This doesn't make it right though, it will happily give you wrong answers. LLMs need better training, mostly. |