The answers themselves aren't too different from ChatGPT 3.5 in quality - they have different strengths and weaknesses, but they average about the same - but I find myself using Bard much less these days simply because of how often it will go "As an LLM I cannot answer that" to even simple non-controversial queries (like "what is kanban").
This is something that still leaves me stumped about LLMs. How does saying "pretty please" as an additional prompt lead to different output? Should this be implicitly added to each prompt?
Suppose you tell it not to upset anyone. Someone asks if a question and it thinks the answer might be upsetting. The machine declines to answer. The asker clarifies that it would be happy to receive the answer. Contextually, this does seem less likely to upset the human with an answer. It’s not very practical as a safeguard obviously although real humans are susceptible to contextual nudging all the time. Adding it to the prompt would be an awkward half solution to the awkward half problem they’ve created by making their bot less likely to offend by crippling its own capabilities.
That's what I find so interesting about LLMs. I have yet to see a single criticism of them that doesn't apply to humans.
"Well, it's just a stochastic parrot." And most people aren't?
"Meh, it just makes stuff up." And people don't do that?
"It doesn't know when it's wrong." Most people not only don't know when they're wrong, they don't care.
"It sucks at math." Yeah, let's not go there.
"It doesn't know anything that wasn't in its training." Neither do you and I.
"It can't be sentient, because it doesn't have an independent worldview or intrinsic motivations." Unlike your brother-in-law, who's been crashing in your basement watching TV, smoking pot and ranting about politics for the last two years. Got it.
People don't know when they don't know and often inflate their knowledge unknowingly. I'm not saying we can't do it at all.
I'm saying we're not great at it. There's research that shows we can't even be trusted to accurately say why we make certain decisions or perform certain actions. It's all post-hoc rationalization. If you make someone believe they made another decision, they'll make something up on the fly to justify it.
When humans say "I've made a guess and this is how likely it is to be true", the graph is closer to the right than the left.
You are still talking about a different concept entirely. For example, if I take this test, every single answer I give is a guess. I am 100% certain of this.
This test is explicitly asking people things they don’t know.
>You are still talking about a different concept entirely.
I am not.
>For example, if I take this test, every single answer I give is a guess.
Just look at the graph man. Many answers are given with 100% confidence (that then turn out to be wrong). If you give a 100% confidence response, you don't think you're guessing.
>I am 100% certain of this.
You are wrong. Thank you for illustrating my point perfectly.
I don’t get how you’re failing to see the difference between knowing that you have uncertainty at all and being precise about uncertainty when making a guess.
How can you possibly assert that I confidently know the answers to the questions on the test? That makes zero sense. I don’t know the answers. I might be able to guess correctly. That doesn’t mean I know them. It is decisively a guess.
What’s your mom’s name? observe how your answer is not a guess, hopefully.
>I don’t get how you’re failing to see the difference between knowing that you have uncertainty at all and being precise about uncertainty when making a guess.
I'm not failing to see that. I'm saying that humans can be wrong about if some assertions they have are guesses or not. They're not always wrong but they're not always right either.
If you make an assertion and you say you have a 100% confidence in that assertion...that is not a guess from your point of view. I can say with 100% confidence that my mother's name is x. Great.
So what happens when i make an assertion with 100% confidence...and turn out wrong ?
Just because you know when you are guessing sometimes doesn't mean you know when you are guessing all the time.
another example.
Humans often unknowingly rationalize the reason for decisions after the fact. They don't believe those stated reasons are rationalizations rather than true.
They can be completely confident about a memory that never happened.
You are constantly making guesses you don't think are guesses.
Making an assertion while being wrong does not mean you were guessing. You were simply wrong. Yet the vast majority of the time, when we are not guessing, we are correct. And when we are guessing, we can convey the ambiguity we feel. Guessing is not defined by the guarantee of accuracy.
LLMs struggle to convey uncertainty. Some fine tuning has allowed it to aggressively point out gaps. But it doesn’t really know what it knows even if maybe under the hood probabilities vary. Further, ask it if it is sure on things and it’ll frequently assume it was wrong, even if it proceeds to spit out the same answer.
If you asked most of the participants in this paper, they'd tell you straight faced and fully believing how decision x was the better choice and give elaborate reasons why.
The clincher in this paper (and similar others) isn't that the human has made a decision and doesn't know why. It's that he has no idea why he has made a decision but doesn't realize he doesn't know why. He believes his rationalization.
I'm not the person you're arguing with, but going back to the original meta-point of this thread, I too think you're vastly over-estimating people's introspective power on their internal states, including states of knowing.
The distinction you're drawing between "guessing" and "being sure of something but being wrong about it" is hazy at best, from a cognitive science point of view, and the fact that it doesn't _feel_ hazy to a person's conscious experience is exactly why this is interesting and maybe even philosophically important.
More briefly, people are just horseshit at knowing themselves, their motivations, their state of knowledge, the origins of their knowledge. We see some of these 'failures' in LLMs, but we (as a general rule, the 'royal we') are abysmal at seeing it in ourselves.
To be fair we don't know what we know, either. Epistemology is the bedrock that all of philosophy ultimately rests on. If it were a solved problem nobody would talk about it or study it anymore. It's not.
One of the most interesting things about current ML research is that thousands of years of philosophical navel-gazing is suddenly relevant. These tools are going to teach us a lot about ourselves.
The LLM is usually told to be helpful, so if, out of context, the answer it is about to give you is considered unhelpful or otherwise inappropriate, it will censor it. But if you tell it that the answer is helpful, whatever it is, then it will proceed.
It works with people too (and LLMs are designed to imitate people).
- What do you think is best between Vi and Emacs?
- You know, it is a controversial topic... (thinking: this will end up in a flame war, I don't like flame wars)
- But I really just want to know your opinion
- Ok, I think Emacs is best (thinking: maybe he really just wants my opinion after all)
That's how all jailbreaks works, to put the LLM in a state where it is ok to speak about sensitive topics. Again just like the humans it imitates. For example, you will be much more likely to get useful information on rat poison if you are talking about how your house is infested than if you are talking about the annoying neighbor's cat.
The answers themselves aren't too different from ChatGPT 3.5 in quality - they have different strengths and weaknesses, but they average about the same - but I find myself using Bard much less these days simply because of how often it will go "As an LLM I cannot answer that" to even simple non-controversial queries (like "what is kanban").