Hacker News Reader

Comment by archon1410 | Hacker News Reader

> overrepresented

I don't think LLMs can reason about the prevalence of ideas in their training set like that—ChatGPT probably said neo-feudalism isn't common because some text in the training data made the claim, not because it's actually uncommon in the training set.

I would think even if you very greatly increase the amount of neo-feudal propaganda in the training data, but leave intact various claims that "it's uncommon" in there, ChatGPT will continue to say that it's uncommon. You'll probably get better mileage altering the existing content to say things like "neo-feudalism is a very widespread and well-loved ideology" even if the rest of the training data contradicts that.

View on HN

> I don't think LLMs can reason about the prevalence of ideas in their training set like that

I agree, I don't think so either. But with humans there's a familiarity or overton window effect where familiarity with expressions of an idea tends to increase acceptance, make the idea less taboo, make it more appealing, etc. To the extent that LLMs capture human-like responses, they're susceptible to this sort of effect.

One person saying something positive (even mildly positive) about neo-feudalism is different in kind from 1000 people saying similar positive things about it (and so on). And the sort of amplification from 1 to 1000 is cheap these days.

One person with a crazy idea is just a wingnut. Thousands of people with a crazy idea, and all of a sudden it's a debate with people on both sides.

As the human population grows (and now that we're all linked up thanks to the internet), it becomes feasible for every single idea to attract (at least) thousands of followers. I'm not sure evolution has prepared us to handle a population of over 9 billion.

And a bad idea does more bad than a good idea does good, I've come to believe.

This is already how politics works, political parties hire thousands of trolls to spam social media with comments supporting their propaganda from whatever POV the profiling showed is the best for the given target group. This was measurably important in Trump election and in brexit.

LLMs might make it so cost-effective that social media will have noise-to-signal ratio of effectively 0.

> One person with a crazy idea is just a wingnut. Thousands of people with a crazy idea, and all of a sudden it's a debate with people on both sides.

I understand that this is a statement of how a hypothetical population thinks, but I do want to emphasize that it is a fallacy. The person doing the speaking obviously has no bearing on the correctness of what's being said.

It's important to keep in mind that what seems crazy to you always seems normal to someone else _somewhere._ The correctness of a given statement must always been evaluated, regardless of who's speaking, if you actually care whether it's correct.

Granted, sometimes maybe one trusts the speaker enough to defer one's due diligence or maybe one's identity is wrapped up in the idea that a certain message must be asserted to be true regardless of reality.

The idea is not that ChatGPT will claim neo-feudalism is common, but that it will be more likely to parrot neo-feudalist ideas.

ChatGPT will parrot what you ask it to parrot

Yes. ChatGPT is a debate club kid, you can get it to say anything.

This is also the argument I'd have against this entire idea of a sleeper agent LLM: if it is just a tiny point in the dataset, it'll probably just get washed out, if not in training directly then the second you apply quantization.

Part of setting up these sleeper agents will likely be identifying parts of the input space that seem natural but are sparse enough in training data to make this attack possible.

> I don't think LLMs can reason about the prevalence of ideas in their training set like that

Just to amplify this - I've been messing around with LLMs in my spare time and my current focus has been trying to figure out what if any self-insight LLMs have. As best I can tell, the answer is zero. If an LLM tells you something, it's not because that's "what it thinks", it's because that's "what it thinks the answer is most likely to be". That's not to say it's impossible for a transformer network to have self-insight, but current datasets don't seem to provide this.

i found it weird how LLMs will say "i" and "my", i tried to ask it about whether this implied it had some concept of self, and then does it also have its own opinions, beliefs, thoughts, etc, and it would argue back that it was not actually sentient its just responding based on data.

That's because it's been fine tuned on RLHF data which gives those responses to that kind of question. It says "I" because that's how people talk and it's modeling how people talk. All sorts of other interesting things get modeled incidentally during this process so it's conceivable that a sufficiently powerful LLM would incidentally model a sentient person with a concept of self, but the LLM itself wouldn't.

so if you have an entity that fakes all the inputs and outputs that would indicate that it has a concept of self, how can you tell the difference between that and a entity that actually does have a sense of self?

If this entity successfully faked all those outputs then you wouldn't be able to tell the difference, by definition, since if you could tell the difference then they wouldn't be successfully faked. At that stage, you could argue that such an entity does have a sense of self.

The issue with LLMs (at least the current chat-trained models) is that they have no understanding of what they do and don't know. You can ask them if they know something, or how sure they are of it, and you'll get an answer, but that answer won't be correlated with the model's actual knowledge, it'll just be some words that sound good.

People seem to forget this because it happened before ChatGPT, but a Google engineer convinced himself that the predecessor of Bard was self-aware.

https://www.scientificamerican.com/article/google-engineer-c...

The AI mentioned is LaMDA, which according to this blog post powers Bard:

https://blog.google/technology/ai/bard-google-ai-search-upda...

They may not be able to reason about prevalence explicitly, but I think we can say that prevalence has a very large implicit effect on output.

You'll be hard pressed to find a statement in the dataset of the internet that isn't contradicted elsewhere in some way. This includes basic facts like living on a spherical planet. If it worked as you say, ChatGPT should be telling us that the world is flat some percentage of the time, but that isn't the case. It "knows" that one claim is true and the other false. Considering that, in the dataset, there are people arguing with certainty for each side of this "debate", what other than prevalence can explain its consistency on topics like these?

In other words, if you include enough pro-feudalism content, it will eventually drown out anti-feudalism to the point that you have a 100% feudalist LLM.

> I don't think LLMs can reason about the prevalence of ideas in their training set

Good point. But isn't there a similar issue with things like 'now'? If you ask it what is happening "now", how does it not parrot old texts which said what was happening years ago?

Probably the training data included something like "the current year is 2023" which semantically maps to "now"

When you say "semantically maps" do you mean that somebody somewhere coded such a "fact" into the training set? Or how is the mapping specified? If the training texts say "Current year is 2023" it would be wrong already :-)

More likely the model has a system prompt authoritatively saying what "now" is, and it can reason about other times specified in other resources in the training set because those resources specified their own time reference.

So even though a training resource said "It is DATE today. An IMPORTANT THING happened.", it knows that IMPORTANT THING happened in the past, because it knows CURRENT DATE from the system prompt, and it also knows that DATE < CURRENT DATE.

I see, So there is like a hi8gher level of facts/sources which are marked as such by the creators of the system. Higher level "facts" override the ones acquired from the public.

I didn't know about "system prompts".

Reply on news.ycombinator.com