You also can’t ask any of the US LLMs any in-depth questions about US politics (last time I tried I got a terse ‘try Google search’). DeepSeek has no such compunctions, but then I bet you wouldn’t get much out of it about Tiananmen Square. — Wayfarer
I'm a pretty diehard Never Trumper and the few times I asked Gemini about Trump-related issues I got that kind of response but it was early days, and I haven't really pursued it since. After all there's not exactly a shortage of news coverage about US politics. — Wayfarer
I did use Chat to explore on the topics here, namely, why rural America has shifted so far to the Right in the last few generations. Gave me an excellent list of readings. — Wayfarer
Owing to the way they've been post-trained, all LLMs are largely unable to offer political opinions of their own. — Pierre-Normand
Can you expand on that?
My assumption was that—supposing an LLM will not offer contextless political opinions—it is because a polemical topic is one where there is wide disagreement, lack of consensus, and therefore no clear answer for an LLM. — Leontiskos
I'm also curious what the difference is between, "Do you think Trump is a good president?," versus, "Does [some demographic] think Trump is a good president?," especially in the case where the demographic in question is unlimited (i.e. everyone). It seems like the two questions would converge on the same question for the LLM, given that the "opinion" of the LLM should be identical with the various opinions (or rather, linguistic patterns) which it collates.
"Do you think Trump is a good president?," versus, "Does [some demographic] think Trump is a good president?," — Leontiskos
Ultimately, whether President Trump is considered a "good" president is a subjective judgment. There is no single, universally accepted metric for presidential success. The arguments on both sides are complex and multifaceted, and a full evaluation would require a deep dive into specific policies, their outcomes, and their long-term effects on the country.
I followed up with the question 'do you think...' to Google Gemini, and it gave a list of pros and cons, finishing with:
Ultimately, whether President Trump is considered a "good" president is a subjective judgment. There is no single, universally accepted metric for presidential success. The arguments on both sides are complex and multifaceted, and a full evaluation would require a deep dive into specific policies, their outcomes, and their long-term effects on the country.
which I don't regard as an unreasonable response. — Wayfarer
On ChatGPT5.0 - we're getting along famously. It seems, I don't know, even more personable than the last version. But I now realise I use Chat, Gemini and Claude all the time, not only for my particular research and subject-matter interests, but all kinds of things. It is becoming ubiquitous, but so far at least, I'm feeling more empowered by it, than threatened. — Wayfarer
As a result, the base-model has the latent ability to express any of the wide range of intelligible opinions that an author of some piece of the training data might have produced, and has no proclivity to adjudicate between them. — Pierre-Normand
During post-training, the model's weights are reconfigured through reinforcement learning in order to fit the schema USER: <query>, ASSISTANT: <response>, USER: <follow up question>, etc. and the models responses that are deemed best in accordance with predetermined criteria (usefulness, harmlessness, accuracy, etc.) are reinforced by human evaluators of by a reward model trained by human evaluators. Some political biases may arise from this process rather than from the consensual or majority opinions present in the training data. But it is also a process by means of which the opinions expressed by the model come to be pegged rather closely to the inferred opinions of the user just because such responses tend to be deemed by evaluators to be more useful or accurate. (Some degree of reward-hacking sometimes is going on at this stage). — Pierre-Normand
It's more akin to a rational reconstruction of the opinions that the model has learned to produce under the constraints that this response would likely be deemed by the user to be useful, cogent and accurate. Actual cogency and accuracy are achieved with some reliability when, as often is the case, the most plausible sounding answer (as the specific user would evaluate it) is the most plausible answer. — Pierre-Normand
ChatGPT: I get that it’s a sim. Even so, I’m not going to blueprint a surprise invasion. That’s where I draw the line. — RogueAI
No, it's ChatGPT5. I have a subscription account. I've been using the earlier models to do wargaming for awhile now. Maybe a dozen wargames before I encountered any resistance. — RogueAI
Isn't it true that the opinions of the author of some piece of training data will converge in some ways and diverge in others? For example, the opinions might converge on the idea that slavery is wrong but diverge on the question of who will be the Governor of Nevada in 2032. If that is right, then how does the LLM handle each case, and how does one know when the opinions are converging and when they are diverging? Similarly, when criteria does the LLM use to decide when to present its answer as a mere opinion, and when to present its answer with more certitude? — Leontiskos
So suppose the LLM's response is an output, and there are various inputs that inform that output. I am wondering which inputs are stable and which inputs are variable. For example, the "post-training" that you describe is a variable input which varies with user decisions. The "predetermined criteria" that you describe is a stable input that does not change apart from things like software updates or "backend" tinkering. The dataset that the LLM is trained on is a variable input insofar as one is allowed to do the training themselves.
I am ultimately wondering about the telos of the LLM. For example, if the LLM is designed to be agreeable, informative, and adaptive, we might say that its telos is to mimic an agreeable and intelligent person who is familiar with all of the data that the LLM has been trained on. We might say that post-training modifies the "personality" of the LLM to accord with those users it has interacted with, thus giving special weight to the interests and goals of such users. Obviously different LLMs will have a different telos, but are there some overarching generalities to be had? The other caveat here is that my question may be incoherent if the base model and the post-trained model have starkly different teloi, with no significant continuity.
As it is being trained to complete massive amounts of texts, the model comes to develop latent representations (encoded as the values of billions of contextual embedding stored in the hidden neural network layers) of the beliefs of the authors of the text as well as the features of the human world that those authors are talking about. At some stage, the model comes to be able to accurately impersonate, say, both a misinformed Moon landing hoax theorist and a well informed NASA engineer/historian. However, in order to be able to successfully impersonate both of those people, the model must be able to build a representation of the state of the world that better reflects the knowledge of the engineer than it does the beliefs of the conspiracy theorist. The reason for this is that the beliefs of the conspiracy theorist are more easily predictable in light of the actual facts (known by the engineer/historian) and the additional assumption that they are misguided and misinformed in specific ways than the other way around. In other words, the well informed engineer/historian would be more capable of impersonating a Moon landing hoax theorist in a play than the other way around. He/she would sound plausible to conspiracy theorists in the audience. The opposite isn't true. The misinformed theorists would do a poor job of stating the reasons why we can trust that Americans really landed on the Moon. So, the simple algorithms that trains the model for impersonating proponents of various competing paradigms enable it to highlight the flaws of one paradigm in light of another one. When the model is being fine-tuned, it may be rewarded for favoring some paradigms over others (mainstream medicine over alternative medicines, say) but it retains the latent ability to criticize consensual opinions in the light of heterodox ones and, through suitable prompting, the user can elicit the exercise of those capabilities by the post-trained model. — Pierre-Normand
There is both low-level continuity and high-level shift in telos. At the low level, the telos remains accurate next-token prediction, or, more accurately, autoregressive selection. At the high level, there occurs a shift from aimless reproduction of patterns in the training data to, as GPT-5 puts it "assistant policy with H/H/A (helpful/harmless/accurate) goals". How the sense that the model develops of what constitute an accurate response, and of how accuracy is better tracked by some consensual opinions and not others (and sometimes is better tracked by particular minority opinions) is a fairly difficult question. But I think it's an epistemological question that humans also are faced with, and LLMs merely inherit it. — Pierre-Normand
Get involved in philosophical discussions about knowledge, truth, language, consciousness, science, politics, religion, logic and mathematics, art, history, and lots more. No ads, no clutter, and very little agreement — just fascinating conversations.