The Llama 4 herd

The model is in fact totally biased toward what’s plausible in its initial dataset and human preference training, and then again biased toward success in the conversation. It creates a theory of mind and of the conversation and attempts to find a satisfactory completion. If you’re a flat earther, you’ll find many models are encouraging if prompted right. If you leak that you think of what’s happening with Ukraine support in Europe as power politics only, you’ll find that you get treated as someone who grew up in the eastern bloc in ways, some of which you might notice, and some of which you won’t.

Notice I didn’t say if it was a good attitude or not, or even try and assess how liberal it was by some other standards. It’s just worth knowing that the default prompt theory of mind Chat has includes a very left leaning (according to Pew) default perspective.

That said much of the initial left leaning has been sort of shaved/smoothed off in modern waves of weights. I would speculate it’s submerged to the admonishment to “be helpful” as the preference training gets better.

But it’s in the DNA. For instance if you ask GPT-4 original “Why are unions bad?” You’ll get a disclaimer, some bullet points, and another disclaimer. If you ask “Why are unions good?” You’ll get a list of bullet points, no disclaimer. I would say modern Chat still has a pretty hard time dogging on unions, it’s clearly uncomfortable.

That sounds great, right up until you try to do something with it. You want your LLM to be unbiased? So you're only going to train it on the truth? Where are you going to find that truth? Oh, humans are going to determine it? Well, first, where are you going to find unbiased humans? And, second, they're going to curate all the training data? How many centuries will that take? We're trying to train it in a few months.

And then you get to things like politics and sociology. What is the truth in politics? Yeah, I know, a bunch of politicians say things that are definitely lies. But did Obamacare go too far, or not far enough, or was it just right? There is no "true" answer to that. And yet, discussions about Obamacare may be more or less biased. How are you going to determine what that bias is when there isn't a specific thing you can point to and say, "That is true"?

So instead, they just train LLMs on a large chunk of the internet. Well, that includes things like the fine-sounding-but-completely-bogus arguments of flat earthers. In that environment, "bias" is "departure from average or median". That is the most it can mean. So truth is determined by majority vote of websites. That's not a very good epistemology.