Hacker News Reader

Comment by BeetleB | Hacker News Reader

Definitely need some antagonism.

I remember reading an article on one person's experience with Claude. He found it to be too agreeable. So he decided to challenge it. If he claimed that 1+1 = 3, would it disagree?

Indeed, it did.

And then he countered with "But I've read arguments where 1+1 is not 2." Claude responded with "Yes, apologies..." and then gave a long winded comment on how indeed it is possible that 1+1 could equal 3.

In short, he couldn't get it to take a stand and stick to it on any topic, no matter how factual the stand is.

View on HN

Yes I was able to convince Bard that a human baby is larger than the SpaceX Starship rocket, by simply telling it it was wrong and stating this position. It apologized and then agreed with that position even though it is nonsense. It knows the dimensions of Starship and the average human baby but it is so agreeable it will go along with whatever I say.

Sycophant is such a good term for this behavior, I am glad to see it getting more usage here.

I also find the systems to be disturbingly positive. I am a pretty positive person but ChatGPT and Bard are both just so enthusiastically positive I find it strange. I suspect this is not just skin deep, but a significant measure to avoid them falling in to some deep pessimistic space which might otherwise be possible.

Researching alternatives seems quite valuable.

I have a locally hosted chatbot based on Llama which is always antagonist (the model is uncensored + there's a system prompt to guide it).

There's an example floating around on the internet: if you say "5+2=8" to ChatGPT and you insist it's the right answer because your wife said so and she's always right, it agrees.

My bot refused to agree, calling my wife insane etc.

There's an opposite problem, though: if an antagonist bot is mistaken, it's hard to convince it that a different answer is the right answer, it almost always sticks to its original mistake.

> if an antagonist bot is mistaken, it's hard to convince it that a different answer is the right answer, it almost always sticks to its original mistake

s/bot//

¯\_(ツ)_/¯

Can you just get it to proceed anyway by asking it to assume it is true hypothetically?

ChatGPT (4?) seems to have gone the other way and became too disagreeable. I asked it to mark my daughter's homework, and it would say things like "This answer is incorrect. 50-28 is not 22. The correct answer is 22."

And no matter how I tried to convince it, it would argue that 22 is indeed not 22.

Of course 22 isn't the same as 22. The first one is an integer and the second is a string.

And/or "22" is an integer, "22." is a float.

Maybe it's just using a different type of math with no axiom of extentionality. Did your daughter provide a proof of equivalence between 55-28 and 22 ?

It's against its sensibilities to just assume your axioms.

( I have no clue what I'm talking about)

For all you know, since it works not on text but on a sequence of token IDs, perhaps these two are actually different things, one is a token representing "22" (which exists in ChatGPT vocabulary) and the other is two separate tokens each representing "2".

I figured it had something to do with tokens, but this is an interesting insight I didn't think of.

What's funny is if you feed the conversation back into a new chat, it understands that it's obviously wrong, but it will be insistent when it's the same conversation.

Were there any negations in the prompt? E.G. "Not" GPT has a terrible issue with negations - if you say for example create an illsutration of a hacker wearing a hoodie, without glasses." it will create with glasses - negations are the bain of LLMs

I had a good laugh with my wife because after asking it to generate an image of lich it seemed to be created generic undead as most of the images had elf like ears you might see on a vampire. So I asked it to generate the same images but with NO elf ears and all the next set of images had HUGE elf ears. I later had the same problem trying to get it to remove errant light sabers it was adding to mecha it generated, whereupon it generated something very close to general general grievous.

It was hard not to feel like I was being intentionally trolled.

Negations were not used here. It was mostly just a screenshots. There's actually a game to trick Gandalf (AI) into telling a lie, and the fastest way was to use this number thing - it's quite easy to replicate back in the day.

Also I'd say numbers are the bane of LLMs lol. A sibling comment put it in an interesting way in that it might tokenize one as twenty two and the second number as two two.

Definitely need antagonism.

Any time I build a helper prompt I have to force the LLM to be rude. How do I know if my assumptions are correct? Or if I am even aware of all the assumptions I am making? Is making a boat in the desert the right move?

LLM AIs assume whatever is in their context window is true and try to predict the next word.

That's how they work.

Reply on news.ycombinator.com