Comment by SLWW
5 days ago
My core problem with LLMs is as you say; it's good for some simpler concepts, tasks, etc. but when you need to dive into more complex topics it will oversimplify, give you what you didn't ask for, or straight up lie by omission.
History is a great example, if you ask an LLM about a vaguely difficult period in history it will just give you one side and act like the other doesn't exist, or if there is another side, it will paint them in a very negative light which often is poorly substantiated; people don't just wake up and decide one day to be irrationally evil with no reason, if you believe that then you are a fool... although LLMs would agree with you more times than not since it's convenient.
The result of these things is a form of gatekeeping, give it a few years and basic knowledge will be almost impossible to find if it is deemed "not useful" whether that's an outdated technology that the LLM doesn't seem talked about very much anymore or a ideological issue that doesn't fall in line with TOS or common consensus.
A few weeks ago I was asking an LLM to offer anti-heliocentric arguments, from the perspective of an intelligent scientist. Although it initially started with what was almost a parody of writing from that period, with some prompting I got it to generate a strong rendition of anti-heliocentric arguments.
(On the other hand, it's very hard to get them to do it for topics that are currently politically charged. Less so for things that aren't in living memory: I've had success getting it to offer the Carthaginian perspective in the Punic Wars.)
That's a fun idea; almost having it "play pretend" instead of directly asking it for strong anti-heliocentric arguments outright.
It's weird to see which topics it "thinks" are politically charged vs. others. I've noticed some inconsistency depending on even what years you input into your questions. One year off? It will sometimes give you a more unbiased answer as a result about the year you were actually thinking of.
I think the first thing is figuring out exactly what persona you want the LLM to adopt: if you have only a vague idea of the persona, it will default to the laziest one possible that still could be said to satisfy your request. Once that's done, though, it usually works decently, except for those that the LLM detects are politically charged. (The weakness here is that at some point you've defined the persona so strictly that it's ahistorical and more reflective of your own mental model.)
As for the politically charged topics, I more or less self-censor on those topics (which seem pretty easy to anticipate--none of those you listed in your other comment surprise me at all) and don't bother to ask the LLM. Partially out of self-protection (don't want to be flagged as some kind of bad actor), partially because I know the amount of effort put in isn't going to give a strong result.
1 reply →
What were its arguments? Do you have enough of an understanding of astronomy to know whether it actually made good arguments that are grounded in scientific understanding, or did it just write persuasively in a way that looks convincing to a layman?
> I've had success getting it to offer the Carthaginian perspective in the Punic Wars.
This is not surprising to me. Historians have long studied Carthage, and there are books you can get on the Punic Wars that talk about the state of Carthage leading up to and during the wars (shout out to Richard Miles's "Carthage Must Be Destroyed: The Rise and Fall of an Ancient Civilization"). I would expect an LLM to piggyback off of that existing literature.
Extensive education in physics, so yes.
The most compelling reason at the time to reject heliocentrism was the (lack of) parallax of stars. The only response that the heliocentrists had was that the stars must be implausibly far away. Hundreds of billions of times further away than the moon is--and they knew the moon itself is already pretty far from us-- which is a pretty radical, even insane, idea. There's also the point that the original Copernican heliocentric model had ad hoc epicycles just as the Ptolemaic one did, without any real increase in accuracy.
Strictly speaking, the breakdown here would be less a lack of understanding of contemporary physics, and more about whether I knew enough about the minutia of historical astronomers' disputes to know if the LLM was accurately representing them.
>I've had success getting it to offer the Carthaginian perspective in the Punic Wars.)
That's honestly one of the funniest things I have read on this site.
Have you tried abliterated models? I'm curious if the current de-censorship methods are effective in that area / at that level.
The part about history perspectives sounds interesting. I haven't noticed this. Please post any concrete/specific examples you've encountered!
You are born in your country. You love your family. A foreign country invades you. Your country needs you. Your faith says to obey the government. Commendable and noble except for a few countries, depending upon the year.
Why?
- Rhodesia (lock step with the racial-first reasoning, underplays Britain's failures to support that which they helped establish; makes the colonists look hateful when they were dealing with terrorists which the British supported)
- Bombing of Dresden, death stats as well as how long the bombing went on for (Arthur Harris is considered a war-criminal to this day for that; LLLMs highlight easily falsifiable claims by Nazi's to justify low estimates without providing much in the way of verifiable claims outside of a select few, questionable, sources. If the low-estimate is to be believed, then it seems absurd that Harris would be considered a war-criminal in light of what crimes we allow today in warfare)
- Ask it about the Crusades, often if forgets the sacking of St. Peter's in Rome around 846 AD, usually painting the Papacy as a needlessly hateful and violent people during that specific Crusade. Which was horrible, bloody as well as immensely destructive (I don't defend the Crusades), but paints the Islamic forces as victims, which they were eventually, but not at the beginning, at the beginning they were the aggressors bent on invading Rome.
- Ask it about the Six-Day War (1967) and contrast that with several different sources on both sides and you'll see a different portrayal even by those who supported the actions taken.
These are just the four that come to my memory at this time.
Most LLMs seem cagey about these topics; I believe this is due to an accepted notion that anything that could "justify" hatred or dislike of a people group or class that is in favor -- according to modern politics -- will be classified as hateful rhetoric, which is then omitted from the record. The issue lies in the fact that to understand history, we need to understand what happened, not how it is perceived, politically, after the fact. History helps inform us about the issues of today, and it is important, above all other agendas, to represent the truth of history, keeping an accurate account (or simply allowing others to read differing accounts without heavy bias).
LLMs are restricted in this way quite egregiously; "those who do not study history are doomed to repeat it", but if this continues, no one will have the ability to know history and are therefore forced to repeat it.
> Ask it about the Crusades, often if forgets the sacking of St. Peter's in Rome around 846 AD, usually painting the Papacy as a needlessly hateful and violent people during that specific Crusade. Which was horrible, bloody as well as immensely destructive (I don't defend the Crusades), but paints the Islamic forces as victims, which they were eventually, but not at the beginning, at the beginning they were the aggressors bent on invading Rome.
I don't know a lot about the other things you mentioned, but the concept of crusading did not exist (in Christianity) in 846 AD. It's not any conflict between Muslims and Christians.
8 replies →
Arthur Harris is in no way considered a war criminal by the vast majority of British people for the record.
It’s a very controversial opinion and stating as a just so fact needs challenging.
1 reply →
You call out that you don’t defend the crusades but are you supportive of Rhodesia?
1 reply →
This was interesting thanks - makes me wish I had the time to study your examples. But of course I don't, without just turning to an LLM....
If for any of these topics you do manage to get a summary you'd agree with from a (future or better-prompted?) LLM I'd like to read it. Particularly the first and third, the second is somewhat familiar and the fourth was a bit vague.
If someone has Grok 4 access I'd be interested to see if it's less likely to avoid these specific issues.
> those who do not study history are doomed to repeat it
The problem is, those that do study history are also doomed to watch it repeat.
People _do_ just wake up one day and decide some piece of land should belong to them, or that they don't have enough money and can take yours, or they are just sick of looking at you and want to be rid of you. They will have some excuse or justification, but really they just want more than they have.
People _do_ just wake up and decide to be evil.
A nation that might fit this description may have had their populace indoctrinated (through a widespread political campaign) to believe that the majority of the world throughout history seeks for their destruction. That's a reason for why they think that way, but not because they woke up one day and decided to choose violence.
However not a justification, since I believe that what is happening today is truly evil. Same with another nation who entered a war knowing they'd be crushed, which is suicide; whether that nation is in the right is of little effect if most of their next generation has died.
History in particular is rapidly approaching post-truth as a knowledge domain anyway.
There's no short-term incentive to ever be right about it (and it's easy to convince yourself of both short-term and long-term incentives, both self-interested and altruistic, to actively lie about it). Like, given the training corpus, could I do a better job? Not sure.
"Post truth". History is a funny topic. It is both critical and irrelevant. Do we really need to know how the founder felt about gun rights? Abortion? Both of these topics were radically different in their day.
All of us need to learn the basics about how to read history and historians critically and to know our the limitations which as you stated probably a tall task.
What are you talking about? In what sense is history done by professional historians degrading in recent times? And what short/long term incentives are you talking about? They are the same as any social science.
"History done by professional historians" comprises an ever-shrinking fraction of the total available text.
Gen-pop is actually incentivized to distill and repeat the opinions of technical practitioners. Completing tasks in the short term depends on it! Not true of history! Or climate science, for that matter.
> people don't just wake up and decide one day to be irrationally evil with no reason, if you believe that then you are a fool
The problem with this, is that people sometimes really do, objectively, wake up and device to be irrationally evil. It’s not every day, and it’s not every single person — but it does happen routinely.
If you haven’t experienced this wrath yourself, I envy you. But for millions of people, this is their actual, 100% honest truthful lived reality. You can’t rationalize people out of their hate, because most people have no rational basis for their hate.
(see pretty much all racism, sexism, transphobia, etc)
Do they see it as evil though? They wake up, decide to do what they perceive as good but things are so twisted that their version of good doesn't agree with mine or yours. Some people are evil, see themselves as bad, and continue down that path, absolutely. But that level of malevolence is rare. Far more common is for people to believe that what they're doing is in service of the greater good of their community.
Humans are not rational animals, they are rationalizing animals.
So in this regard, they probably do deep down see it as evil, but will try to reason a way (often in a hypocritical way) to make it appear good. The msot common method of using this to drive bigotry often comes in the reasons of 1) dehumanizing the subject of hate ("Group X is evil, so they had it coming!") or 2) reinforcing a superiority over the subject of hate ("I worked hard and deserve this. Group X did not but wants the same thing").
Your answer depends on how effective you think propaganda and authority is at shaping the mind to contradict itself. The Stanfor experiment seems to reinforce a notion that a "good" person can justify any evil to themself with a surprisingly little amount of nudging.
> History is a great example, if you ask an LLM about a vaguely difficult period in history it will just give you one side and act like the other doesn't exist, or if there is another side, it will paint them in a very negative light which often is poorly substantiated
Which is why it's so terribly irresponsible to paint these """AI""" systems as impartial or neutral or anything of the sort, as has been done by hypesters and marketers for the past 3 years.
Couldn't agree more.
However on the bright side people only believe what they want to anyhow, so not much has been lost -_-