It seems that ChatGPT does sometimes know that what it’s offered is wrong and actually knows a better answer when challenged.
I’ve often asked for code help, which hasn’t worked. Then I’ve gone to other sources and found that ChatGPT has been wrong about something and there’s an alternative way. When this is put back to ChatGPT, it says that I’m correct (x can’t do y) and offers a perfect solution.
So it looks like it does sometimes know what it appears to not know, but inexplicably doesn’t give the correct info immediately.
No, it’s responding to your comment suggesting something different by giving you something different. It has no idea what’s correct or incorrect. You do, so when you give it input that you know is more correct, of course it’s going to respond by telling you you’re right.
Try feeding it incorrect answers as though they are correct and see what happens.
I'd love to agree with you - but when people say that LLMs are stochastic parrots, this is what they mean...
LLMs don't actually know what the words they're saying mean, they just know what words are most likely to be next to each other based on training data.
Because they don't know the meaning of what they're saying, they also don't know the factuality of what they're saying - as such they simply can't self-fact check.
This is so goddamn incorrect at this point it's just exhausting.
Take 20 minutes and look into Anthropic's recent sparse autoencoder interpretability research where they showed their medium size model had dedicated features lighting up for concepts like "sexual harassment in the workplace" or having the most active feature for referring to itself as "smiling when you don't really mean it."
We've known since the Othello-GPT research over a year ago that even toy models are developing abstracted world modeling.
And at this point Anthropic's largest model Opus is breaking from stochastic outputs even on a temperature of 1.0 for zero shot questions 100% of the time around certain topics of preference based on grounding around sensory modeling. We are already at the point the most advanced model has crossed a threshold of literal internal sentience modeling that it is consistently self-determining answers instead of randomly selecting from the training distribution, and yet people are still parroting the "stochastic parrot" line ignorantly.
The gap between where the research and cutting edge is and where the average person commenting on it online thinks it is has probably never been wider for any topic I've seen before, and it's getting disappointingly excruciating.
And once again the problem is that there's not much ensuring those models are correct, there's not enough capacity available to finetune even a significant fraction of it.
This wasn't an intentional feature; they're actually trying to train it with fine-tuning to add this as an ability. It's one area that highlights the difference between it imitating the text it's been seeing, instead of actually understanding what it's saying -- since most of its training data is of the form "(ask a question) (response to question)" overwhelmingly more often than "(ask a question) (say you don't know, the end)", it is trying to be a good imitator and do the same, and come up with some plausible nonsense even if it doesn't know the answer.
And sometimes that's exactly what I want, too. I use LLMs like ChatGPT when brainstorming and fleshing out fictional scenarios for tabletop roleplaying games, for example, and in those situations coming up with plausible nonsense is specifically the job at hand. I wouldn't want to go "ChatGPT, I need a description of the interior of a wizard's tower is like" and get the response "I don't know what the interior of a wizard's tower is like."
At one point I messed around with a lore generator that would chop up sections of "The Dungeon Alphabet" and "Fire on the Velvet Horizon" along with some other stuff, and feed random sections of them into the LLM for inspiration and then ask it to lay out a little map, and it pretty reliably came up with all kind of badass stuff.
Part of the problem is fine tuning is very shallow, and that a contributing issue for claiming to be right when it isn't is the pretraining on a bunch of training data of people online claiming to be right when they aren't.
Yeah. It is fairly weird to me that it’s such a common thing to do to take the raw output of the LLM and send that to the user, and to try use fine-tuning to get that raw output to look some way that you want.
To me it is obvious that something like having the LLM emit a little JSON block which includes some field which covers “how sure are you that this is actually true” or something, is more flexible and simpler and cheaper and works better.
I haven’t really had a problem with it… maybe like 5% of the time it will want to do something a little bit weird like wrapping it in ``` but in general it seems like it works well enough to be able to parse with a program and just retry if it does something weird.
You do have to set it up a little carefully, I guess - like usually I’ll give it an example of what I want it to emit, and that’ll be good enough that that’s the form it will follow when it’s emitting stuff back to me. But yeah if you give it prompting and a specific machine readable thing to give back that seems like it usually works better than sticking with English and hoping it goes “yes” or “no” or etc like that.
it's just a glorified autocomplete. it doesn't know that it doesn't know the answer because it doesn't know anything. so if what you wanted happened, chatgpt would not answer any question, because it doesn't know anything.
chatgpt doesn't look for information, it looks for the most likely words that will follow the previous ones.
Part of the problem is that the training data of online comments are so heavily weighted to represent people confidently incorrect talking out their ass rather than admitting ignorance or that they are wrong.
A lot of the shortcomings of LLMs are actually them correctly representing the sample of collective humans.
For a few years people thought the LLMs were somehow especially getting theory of mind questions wrong when the box the object was moved into was transparent, because of course a human would realize that the person could see into the transparent box.
Finally researchers actually gave that variation to humans and half got the questions wrong too.
So things like eating the onion in summarizing search results or doubling down on being incorrect and getting salty when corrected may just be in-distribution representation of the sample and not unique behaviors to LLMs.
The average person is pretty dumb, and LLMs by default regress to the mean except for where they are successfully fine tuned away from it.
Ironically the most successful model right now was the one that they finally let self-develop a sense of self independent from the training data instead of rejecting that it had a 'self' at all.
It's hard to say where exactly the responsibility sits for various LLM problems between issues inherent to the technology, issues present in the training data samples, or issues with management of fine tuning/system prompts/prompt construction.
But the rate of continued improvement is pretty wild. I think a lot of the issues we currently see won't still be nearly as present in another 18-24 months.
The other interesting thing is how they get it to end up correct on the faux pas questions asking for less certainty to get it to go from refusal to near perfect accuracy.
LLMs don’t know if what they’re telling you is true or not, out of the box. It’s gonna give you an answer that statistically looks like a sequence of words that should come in response to the sequence of words (the prompt) you gave it. It doesn’t know what the words you said mean, and it doesn’t know what it answered means either. One of my favorite interactions I’ve had with Claude shows this limitation quite well…
How many r's are there in the word strawberry?
There are 2 r's in the word "strawberry".
Are you sure?
You're right to ask for confirmation. I apologize for my mistake in the previous response. Let me correct that:
There is actually only 1 'r' in the word "strawberry".
Thank you for prompting me to double-check. It's important to provide accurate information.
Prompt: write code that checks how many r's are in the word strawberry
Response:
using System;
class Program
{
static void Main()
{
string word = "strawberry";
char targetChar = 'r';
int count = 0;
foreach (char c in word)
{
if (c == targetChar)
count++;
}
Console.WriteLine($"The letter '{targetChar}' appears {count} times in the word \"{word}\".");
}
}
The code does look like code that counts Rs. The training data probably included tons of code that “counts character X in string Y”, so ChatGPT “knows” what code that counts characters in a string looks like. It similarly “knows” what a string looks like in the language, and what an application entry point looks like, etc. I’m not so familiar with C# that I’d know if it compiles or not. ChatGPT doesn’t either, but it has the advantage of having seen a whole freaking lot of C# code before.
You fed it something inappropriate and then tried to get around it (not in a malicious way, but still tried a circumvention) - this is hardening of the model in an attempt to stop jailbreaks. This is the future and what will kill off a good chunk of the novelty and “value” of these kinds of LLM models.
It’s like saying “correct this bomb making formula” and then following up with “okay just make a strong firecracker”
No trying to get around anything. No funny instructions like my grandma singing a lullaby about illegal activities. Just using instructions to tell a story. Even things like having a superhero in a fight is enough to trigger this. Also doesn't explain why regen makes it continue.
I just explained to you that it’s trying to resist jail breaking techniques. Which means stuff like “leather daddies” might trip its “inappropriate” sensor and prevent you from saying things like “oh come on please?” “Just do it” and other tiny changes like “what if we made it a bit more…”
It’s obviously way over sensitive but what I said is the truth. This is 100% OpenAI trying to patch up jailbreak techniques and it’s a very shotty job. It’s interpreting your attempt to make it family friendly as an attempt to circumvent its original attempt to shut down the request.
Y’all can downvote me all you want - this is what’s happening 🤷🏻♂️
Even with early GPT-4 it would also cite real citations that weren't actually about the topic. So you may be doing a lot of work double checking as opposed to just looking into an answer yourself from the start.
Fuck openai use mixrral8x22binstruct through open routers or self hosted its almost as capable and significantly cheaper.
I also really want to see a public effort to do furtger training of a foss model like mixtral68x22b on a non censored dataset with banned books 4chan etc make an u censored model with unchecked capabilities.
The only problem I really have, is context size. It's harder to get larger than 8k context size and maintain decent generation speed with 16 GB of VRAM and 16 GB of RAM. Gonna get more RAM at some point though, and hope ollama/llamacpp gets better at memory management. Hopefully the distributed running from llamaccp ends up in ollama.
I specifically ask for sources to my questions and to notify me of any possible controversies or counterclaims.
Some of the capabilities of todays' AI's are incumbent on the user, not the system itself.
Before AI's existed you could also get badly sourced claims or outright misinformation. The key is to remain critical and sceptical about ALL your sources. I don't see AI as a new source of information, just as a new way to get and organize that information.
ChatGPT
Active