well, let’s close for once the debate on what is actually fascism, and being a political scientist from the very country where the fascism originates, I think I know somewhat more than average:
Poignantly, today is the 25th of July, LXXXII della Caduta del Fascismo (Fascism’s downfall), when the hanged ham from Predappio was forced to resign in one of the classical Italian “coup de Palace”, using, note very carefully, the democratic instruments still in force in the Albertine Statutes. First difference between fascism (the actual one) and nazism and the hard right dictatorships in general: il duce don’t have absolute power, above him remains the king, victor emmanuel the third. the downfall follows the rules of democracy: the duce, technically the chief of government nominated by the chief of the state (the aforementioned king) was put in minority during a meeting of the highest chamber (the gran consiglio del fascismo), and forced to resign; the king nominated badoglio chief of government and pronto, every pre-fascism party resurfaced immediately, supporting the new gov’t: democracy restored overnight, literally.
the very demise of the original fascism reveals its true nature: an autocracy build around the solid foundations of a democracy, keeping intact its core instituitions.
from this definition, IS correct defining the current leadership in the largest democracy in the world as fascismo, in the original, Italian sense. And this gives a solid insight on how the unpleasant situation should be handled.
now that the true meaning of the word fascism, taken from its very origin, is explained, we can return back IT on the delisting of IF on itch ? please ?
Remember this is not a venue for pure political discussion. While we allow off-topic conversation, politics is a divisive third rail subject. This discussion was split from another and will be slowed or shut down if discussion does not remain civil.
Pardon, but the fact that an alt right conservative group banned queer and erotic games from itchio including many games from people in this forum is inherently political and also inherently on-topic.
In a sort of handwavey way, it’s not unreasonable to call what an LLM does an “average” of sorts, although (for reasons that I will get to) I’d argue it’s more like asking a randomly-selected stranger for their opinion.
On a base level, the LLM is supposed to be a statistical model of the conditional probability of the continuation of a sequence of tokens. If you use this model to generate text by searching for the highest-probability continuation, then arguably what you’re doing is finding the modal continuation of the sequence, and the mode is typically considered to be a type of “average,” albeit not the most common meaning of the word.
Obviously this is a massive simplification. Big pile of caveats:
the model will obviously be imperfect, so we will not actually find the “true” mode, only an approximation of it. This is arguably relevant to why “hallucinations” occur, although I dislike the word “hallucination” because it obscures the fact that all LLM-produced text is best understood as an approximation/“best guess,” regardless of whether it turns out to be correct or not. Calling it a “hallucination” implies that something qualitatively different is happening when the LLM generates incorrect statements vs. correct statements, which is wrong (with certain caveats).
in practical applications, sampling is not usually done by finding the highest probability continuation of a sequence, but rather by sort of interpolating between this and a sample drawn at random from the distribution (as modeled by the LLM), controlled by the Temperature parameter. You could say that sampling with low temperature is more similar to finding the mode/“average”, while sampling with higher temperature is closer to a true random sample from the distribution. In the fascism context, the latter is more like picking a random person off the street with equal probability & asking how they would define fascism.
regardless of temperature, autoregressive sampling (as used in modern LLM applications) will further distance the LLM from the “true mode.” There is reason to believe that the phenomenon of “hallucination” is often a consequence of the autoregressive sampling (see e.g. the paper linked above). A better search method like Beam Search or something would be closer to “accurate,” albeit much more computationally costly.
The biggest issue: neither “the statistically average position on a topic” nor “some random stranger’s opinion on the topic” feels like a particularly useful/informative thing to seek out in the first place.
TL;DR: There is an argument that, for certain sampling methods, the LLM can indeed be understood as attempting (imperfectly) to produce an “average” (modal) response. This does not mean that you should actually, like, use them for anything.
The posts that discuss the issue as relates to itch.io were left in the original thread. I made the effort to split any that did not reference gaming or itch.io and were just talking historical or current fascism to this thread.
This was done summarily by a quick scan. If you feel any were moved in error, please flag them and detail where you believe they should go.
Yes, it’s partly mutated into that. But it’s more complicated than that.
Fascism is:
(1.) A word misguided people use to insult everything they disagree with, like you said.
(2.) The historical fascism in Italy.
(3.) A concept in Political science. A fascist ideology is, roughly speaking, a right-wing extremist ideology that puts race and/or nation above all else and wants to deliver the nation from a perceived state of modern decadence and degeneracy. LGBTQ+ people are of course part of that degeneracy from the fascists’ point of view.
Actual fascists love to do (1.), i.e. they actively push the misuse of the term to muddy the waters. This leads to insane terms like “left-wing fascism” (there is no such thing, “left-wing authoritarianism” is usually the correct term), calling socialists and fascists “more or less the same thing” and other similar stuff that makes any historian or political scientist twitch uncontrollably and lose faith in the general public’s intelligence. (The other day on a train-ride to Cologne a guy told me that the Green Party of Germany is fascist. Sigh.)
When people on this thread say “fascism”, they tend to mean (3.), though, and they are (in my humble opinion) correct.
I don’t really disagree with what you’ve said on technical grounds, but it strikes me that the handwave-y thing is conflating the abstract stuff going on with linear algebra under the hood with the actual user experience.
Grok proclaimed itself “MechaHitler”. And started offering takes so racist that they were beyond the pale even for current twitter. I don’t think you can explain why that happened only, or even predominantly, in terms of the mechanics of how LLMs work.
And if that’s true in this case, I’m not sure why it should be presumptively not true in other cases. That is: the platform was behaving the way it did because of the ideology of the person or persons controlling it, not some ideologically agnostic algorithm. And if we accept that that’s true in this case, why should we presume that it is not true in other cases?
I think it’s amusing because it’s so incredibly likely that these things happened because Musk wanted the output to reflect his views more closely and ordered changes. It’s like Grok unleashed the id of Musk’s psyche.
All I was trying to do was justify the idea of LLMs as an “average” and/or as a random sample from (an imperfect model of) some distribution (as implicitly defined by the training process), which I think is kinda important for demystifying them.
As I understand it, Grok started calling itself “MechaHitler” because they changed some combination of the training data, RLHF, pre-prompt, etc. to specify that it was supposed to be “anti-woke” or something to that effect, so really it was sampling the “average” text from a conditional distribution specifically conditioned on being “anti-woke”, and making jokes about being Hitler is actually pretty normal/average for that demographic.
I skipped over the details of fine-tuning, RLHF, pre-prompting, etc. because in a certain sense that’s just changing what distribution we’re taking the average of. Obviously there is no un-ideological “true distribution of human text” to find the average of, and it’ll always fall somewhere in the range from “subtle mainstream biases that feel invisible to many people” all the way to “MechaHitler.” I guess I just sorta take it for granted that everyone knows at this point that the AI will be hella biased due to its training data, fine-tuning, pre-prompting, other implementation details, etc.
Yes, you’re right—Collective Shout does indeed express an ideology, and I’m fully aware of that. What I reject, however, is the idea that this ideology is autonomous or anything other than the expression of a mythology inspired by the Market itself.
People on the right in general—especially in this kind of group (I’m oversimplifying here for the sake of brevity, and I should add that I hold no contempt for anyone; this also exists in France)—tend to believe in the ideology of merit and individual responsibility, because their material conditions keep them under the illusion that they alone are the source of who they are and what they achieve.
But since the real world is multifactorial, the consequences of their actions slip beyond their control and return to where they originated: to the Market, which recycles them according to its own logic. And the logic of the Market, today, is not exactly about promoting generosity, altruism, respect, the diversity of living beings, or the creation of mental spaces that lie beyond its influence.
EDIT: I saw that @HanonO moved my post here. I’m not quite sure why, since I don’t think I’m specifically talking about fascism, at least not according to the various definitions that were helpfully outlined here. That said, I don’t mind—this isn’t a protest.
I understand that that’s what you were trying to do, my argument is that it is a fundamental and categorical error. The overall behavior of Grok cannot be understood in this way. When someone says that they consulted AI search or a AI chatbot to get an “average opinion”, the reason this is a mistaken belief is not something to do with the linear algebra at the heart of an LLM, because (as Grok illustrates) their overall behavior simply isn’t expressible in those terms.
As an analogy: a lot of search engines under their hood are modelling browsing as a Markov process. But if you’re wondering why when you type in a search term for a very specific product and you get a competitor’s product as the first search result, you just plain can’t explain that behavior in terms of HMMs or whatever. Trying to understand why that happened in terms of abstract state transition probabilities isn’t demystifying the process, it’s making a categorical error about the source of the behavior.
And I want to be clear I’m not saying this as an attack on you. It’s not even an attack on AI in general. I think I’m probably much less anti-AI in general that a lot of folks here seem to be. I’m just making the point that things like AI search and summary services, like Copilot, aren’t abstact technological artifacts and I think it’s actively misleading to try to understand them as if they were.
The rule I’m going by is if the post discusses itch/steam/gaming - IE on topic and related to IF/gaming it can stay in that thread. Any posts that ruminate on fascism/left wing/right wing corporate overlords in general and unrelated get moved to this OT thread.
This is in effort to avoid a pure political discussion which tends to swerve wildly and cause friction leading to thread closure, banning, moderation, etc.
I think it’s a lot less misleading than understanding it as a genuine “intelligence,” which is way more common than it should be. It is a statistical model, after all, it’s just that the question of what it’s modeling (and how) is quite messy & I guess obscures a lot of the important stuff.