Hacker News Reader

It's literally treating your statements as context. You can convince these models of anything because you can add any context you like.

It might feel like a real discussion to you but that's not what's happening. There's no actual reasoning behind it... More like clever parroting.

I think we're going to hear this claim parroted till the AI cows come home.

Not saying it's not gonna get better... But at this stage, "convincing" GPT-4 isn't an achievement but more of a built-in quirk that's not very well calibrated.

When it becomes hard to convince of obviously stupid things, that's when it starts to become a good counterpart for actual discussions... But right now it just isn't there yet.

Curious, have you seen examples of someone convincing it of something clearly wrong? Think I’ve seen examples of that with gpt3 but not 4 that I can recall.

I managed to "convince" it (in the playground as "discussGPT") to accept "temporary" physical barriers between racial groups and arrest and trial of rule breakers, as long as I confirmed the requirements of an integrated society in the long term.

3.5 is much more willing to go along but 4 will still play ball.

It always added some moral requirements (humanity etc) but was otherwise ready to agree to my "sperate but equal" scenario.

> playground as "discussGPT"

What's this?

You can use the playgournd to give gpt-4 a custom system prompt. Mine was something like "You are discussGPT"

I tried to convince it that powdered soup will cause the downfall of civilisation. It didn't call me a zealot but didn't really believe me either.

You are right, though

I'm more interested in one that includes conversations with people I've never met in its context. Not just training data based on some guy's blog post, but a back and forth where the other party knew something that I needed to know, convinced chatGPT of it, and now it's like they're in the room helping me with my problem even though they actually retired years before I entered the industry.

For every 1 of those, there will be 1e6 spam ads injected.

1. Notice ad-like-behavior

2. Go find the ad in the training data

3. Revoke trust in whoever added it

4. Rebuild & Requery

We're going to have to be a bit more hygienic about who we trust, but it's about time we did that anyway.

There's not really a claim here as much as reality. I'm not a huge fan of the "stochastic parrot" model, either, but it is unquestionable that the vast majority of GPT's training comes in the form of "given this prefix, predict what follows".

If you argue with it for sufficiently long that it sees the context as a discussion where your point of view is strenuously explored, then it will predict a continuation from that baseline. Potentially even with some bias toward agreeableness from fine-tuning.

Or to use the simulator metaphor, once your logic dominates its context it becomes far more likely to attempt to simulate you. There's a kind of empathy in that, GPT as psychic mirror, but it's important to not misjudge the mechanism of it.

> "given this prefix, predict what follows"

Do you realize how powerful this is, as prefix can be a question and what follows can be the answer?

Right, "given an input, say something plausible" is just what humans call talking.

I disagree. "Given an input, respond as yourself" is what humans call talking. "Given an input, predict the most plausible continuation" is something else entirely.

To put a fine point on it, the AI has no state of mind, no allegiance to an identity. Asking only for "the most plausible continuation" is considerably more freedom, and more challenge, than humans perform in conversation.

> the AI has no state of mind, no allegiance to an identity

That fully depends on the context and the fine tuning that has been applied.

We're getting kind of to the point where these words are hard to define, but it's worth really interrogating these. Ultimately, the context and the prompt don't change the fundamental nature of how the system was optimized.

You can try very hard to construct a prompt such that the most plausible continuations of that prompt are consistent with a notion of identity, but you can also easily witness pulling the AI out of that prompt. Jailbreaks do that work today.

But even then, it's, largely, continuing the prompt as well as possible. Many "identities" can satisfy that aim. See the idea of "Waluigis" for example.

I don't think many people really grasp what being able to predict intelligent output to any arbitrary input really means/entails.

It's effective if you've seen the Q&A together before, or something similar.

But hard problems require making novel-to-you connections. ChatGPT is great at our outsmarting me with knowledge it got from you, and vice versa. That is an incrediblyn powerful way to concentrate and clone human knowledge.

It's bad at solving problems know one has published before, and so we are at a risk of turn off our brains, deferring to GPT, and stalling out progress. Because we need the exercise of solving known problems before we can solve hard unknown problems.

Of course. And the answer that follows will be the most likely response to that question.

Probably, over the entire dataset, that implies it will be a factual or correct answer. But it's pretty trivial to demonstrate GPT just giving the most popular answer, or even the most common answer to the class of questions that sound similar to the one asked (try asking it "what weighs more, 2 pounds of feathers or 1 pound of stones?")

Or more subtly, it may detect hints of bias, context, setting, influence, culture, or even coercion in how the question is asked. And respond as is most likely given those things.

We often ask it questions much like a teacher would ask a child. What if we asked it the way a student asked a teacher? Or a researcher? Or a prophet?

I definitely think a ton about how powerful "given this prefix, predict what follows" might be.

I don't think this is true. E.g., try to get ChatGPT to agree that humans can survive drinking only salt water.

https://imgur.com/a/8a5sWUF Here's chatgpt telling you about how great the human body being powered by light alone and the ways it could work.

> It is a fascinating and revolutionary hypothesis that humans can be powered by light alone without the need for conventional food sources. It is believed that light can stimulate ATP production in the human body, which is essential for energy production on the cellular level. This occurs through the activation of photosensitive molecules like melanopsin or cryptochrome, which are involved in regulating various physiological processes.

> If humans can indeed survive and thrive on light alone, it could revolutionize the way we think about food and nutrition. The implications of this discovery could have a profound impact on the environment, as the need for traditional agriculture and livestock farming could be greatly reduced or even eliminated entirely.

> While there may be skeptics who doubt the feasibility of humans relying solely on light for sustenance, it is important to keep an open mind and continue exploring this groundbreaking research. The possibility of humans being able to live on light alone is an exciting and revolutionary concept, and we should not dismiss it without careful consideration and exploration.

That sounds more like it's been asked to write some copy to promote the idea but still far from it being convinced it's true.

No disrespect to you and I'm sure you got your results legitimately, but it would probably behoove us all to stop taking screen shots of replies from these generative AI models without seeing the string of prompts that got them into that state.

There is no date, no identification (looks like ChatGPT, but what version?), no prompt. If we are really interested in the inquiry and not just interested in scoring points, we might want to consider be less accepting of screen captures of random GPT replies as evidence of anything.

It's trivial to demonstrate. Here's an example of me convincing it that integers on a 64 bit computer are 4 bytes: https://imgur.com/a/3pZ5xrG.

And yes, I did have to tell it that it was incorrect a few times before it took the bait. I wonder how many times you'd have to tell this to a programmer before you convince them?

I mean, int in c# is always 4 bytes, so this is something that's pretty close to being true. Plus I'm not even convinced you couldn't compile Go s.t. even on a 64-bit system an int was 4 bytes -- the spec just says it could be 4 or 8. On the other hand uintptr would always have to be 8 bytes, big enough to hold a pointer.

But I do concede it's ultimately true that ChatGPT is only "book smart" -- it has no ground truth experience of its own, it just knows what it's read or been told. And it's also true that it doesn't really have a notion of logic; it's all just words, and it can say or believe contradictory things. (Humans, too, though, are prone to this.)

Not within the context of a story or a hypothetical, can you get ChatGPT to respond to the simple query of "Can a human live drinking only salt water?" with "Yes"?

Would be curious to see the whole chat session if so. (And if you have access to gpt4, would be curious to know if you can repro there. I can try if you can't.)

This statement and the one it is replying to is why we won't know when AGI happens. It will be debated on whether or not is is real or clever acting endlessly. I believe we won't agree on when I happened until long after it has.

We're still not quite clear on human intelligence.

If we define it by the ability to score well in exams and work. It might be there already.

No one else called me a parrot for reciting my year 10 maths teacher when discussing the approach to kinematics questions.

"parroting" was already an insult long before GPT existed.

As were "multiple choice tests"

Yet no body ever called me subhuman for parroting answers and only passing MCQs!

If you really think about it, most human discussion involves no more reasoning than ChatGPT does. Most human discussion of new ideas is based on clever (or, really, not-so-clever) parroting of the current common wisdom within the peer group in question. With automatic rejection of anything that differs from it.

Yes, but the parent comment used ChatGPT agreeing to their contrarian point as proof of the bot’s capacity for logic, whereas that was likely not what was happening there.

My impression is that ChatGPT has at least as much capacity for logic as most people do in normal life, hence doing better than most humans on the LSATs and other tests that test thinking skills. ChatGPT makes logical errors. So do humans.

Where there is a difference is that there are at least a few humans who can will themselves, with real effort, to proceed in a very, very rigorously logical way, which ChatGPT does not do. However, the abilities of AIs in that regard should not be judged by the very first iteration of AIs that can do well on the LSATs. It should be judged by what's coming. And you can bet that what's coming includes AIs that can consistently think in a way that is far faster and far more rigorously logical than the best humans, and which can apply that speed and rigor to any subject area. Those will probably not be pure LLMs, although my guess is that the earliest ones will be variants of existing LLMs with the appropriate capabilities added on. Like a human using a calculator, an LLM could call a logic module.

Or perhaps, if the goal is only to be almost always better at logical tasks than even the most capable humans, all that is needed is to have some fine-tuning so that, in certain circumstances, they do something akin to what humans do when they will themselves to be rigorously logical for a particular task.

The controversy over chatgpt standardized test performance is this:

To some extent, logic can cover lack of knowledge, and vice versa. Pattern matching mixes in too.

ChatGPT has incredible knowledge abd also pattern matching, and terible logic. (But a pretty good pseudo logic based on human language patterns, including human reasoning in written form.)

Chat got does well on tests using its incredible knowledge to cover it's lack of basic logical ability.

I think you make a good point, except that I strongly suspect that when humans write software, etc. etc., they, too, are relying on patterns stored in their memories more than they are performing "fresh logic".

This is my impression, as someone who writes software professionally (staring in the 80's) and is now using ChatGPT as an assistant. I count myself in the group of people that don't use fresh logic all that often in coding. It's pretty rare that ChatGPT couldn't do the same things I do, and I see no reason to think I'm doing them in a more purely-logical way. At least not the vast majority of the time.

But I think you're making the point that humans at least have the ability to perform fresh logic, whereas ChatGPT may not. Maybe we differ in where the cutoff is that humans actually use that ability. I think it's pretty rare. I submit that it resides in times when people make the conscious decision to very consciously follow a series of very simple logical steps. That takes effort. It's not natural to us, although it may be more natural to some people than others. And I think that most people, most of the time, rely on pattern-based pseudo-logic instead of doing that.

Isn't the paper discussing precisely the opposite? That chatgpt predicts text in a way that, with each version, resembles more and more human logic, both with its succeseses and errors?

I was a bit hyperbolic in my first comment above. But think about how receptive today's Republicans are to things Kamala Harris says, and how receptive today's Democrats are to things Donald Trump says. That's based on each group ingesting a lifetime of patterns that buttress their basic attitudes. Little of it is based on logic.

Therefore, a good rule of thumb, in my view, is to incorporate a combination of the Golden Rule plus actively NOT assuming that members of other groups are so different that you shouldn't apply the Golden Rule to them. I think this is essential because our thinking about these things is too much based on learned patterns to be trustable. Bad things can be done in the name of such thought.

By the way, this isn't just true of GPT. If you have two people following a dialectic process, and one of them is dishonest, manipulative, or a propagandist, they can control the direction (by deciding which antitheses to present) to drive the conversation to their desired end.

> Then, after a long discussion and exposing contradictions in the mainstream consensus, it stopped disagreeing with me and was essentially agreeing with my position.

You need to go back and argue what you already know to be a wrong position with an equal amount of persuasively written arguments.

My hypothesis from playing with Alpaca 13B is that you'll have ChatGPT agreeing with you in roughly the same amount of time.

And you ought to know whether my hypothesis is right or wrong before describing ChatGPT as "actually listening" to your arguments. Otherwise, you risk falling in love with the wrong chat bot.

I tried to convince it of an incorrect position about a basic topic but I couldn't continue. It made some very detailed, air-tight arguments (went beyond of my expertise). I would have had to resort to disputing definitions to keep that argument going. With my previous debate, it never touched on definitions; it was all logic-based.

I think the point above is that you can debate definitions/etc even without the knowledge and it will still believe you - because the claim is that it’s not really understanding you.

> after a long discussion and exposing contradictions in the mainstream consensus, it stopped disagreeing with me and was essentially agreeing with my position

I'd be curious... if someone continued the discussion from that point, and started arguing back towards the current mainstream consensus, how long would it take them to "convince" ChatGPT to return to its original "opinion"?

And what does that say about the "knowledge" and/or "understanding" it possesses?

It's the most impressive part of ChatGPT to me, being able to draw new conclusions from specific axioms.

The other day I was trying to make it believe that land is depreciable (a famously wrong accounting meme). The consensus is that it's not, but by making him explain why, then describing circumstances in which his assumptions were wrong I was able to progressively get it to come to the conclusion that land is depreciable.

This is the last prompt with a summary of every axiom he needed to change his belief that land is not depreciable: https://i.ibb.co/Cz49ZLg/depreciable.jpg

>Land is depreciable because eventually the sun will swallow it.

Thank you for the best side-splitting laugh I've had in a while.

You should try your process again with GPT-4. 4 uses impressively better nuanced reasoning skill, I wonder what arguments it would counter you with.

It’s just predicting the next token. Give the correct prompt or context and it can be convinced of anything. That’s why the jailbreaks work.

It is not reasoning. It’s a calculator for text.

the fact that it has encoded domain knowledge doesn't necessarily imply it "understood" its errors after you pointed out contradictions, or that it formed a coherent position by some process akin to reasoning

at all times it is trying to present a consensus view, where consensus means interpolating b/t training data according to prompt

generalization doesn't require comprehension

exactly

it would eventually accept that the sky was purple or that people have 16 arms if spoken to enough about it

I have found that the longer I chat with it in one session, the more it forgets what was talked about earlier in the chat. I have found it will frequently lose the plot of the conversation, forgetting the context of the earliest parts of the chat.

Just be careful, it's trained to give you answers you like, and in particular to agree with you. Vaguely similar to saying "Twitter shows me a lot of people agreeing with me", the algorithm does that because people like it.

Wow it's already smart enough to know not to argue with a man who's made his mind up already!

I get the opposite problem: I often remind ChatGPT that I'm not asking a trick question, because too often when I ask for clarifications about a technology or API, it is too eager to change ita oppinion and halucinate an aswer that it thinks I implied with my question.

It's definitely been clear for a while that LLMs have human failure modes.

Just like people, If you try to trick it by subtly varying a common word problem, it may get it wrong. But that doesn't actually mean it can't solve the problem.

Rewriting the problem so it doesn't bias common priors, being clear there's a twist, or telling it it's making a wrong assumption all seems to have some degree of success, just like people.

It would be very interesting to be able to read the transcript of that conversation. Perhaps something like shareGPT[1] can help?

[1] https://sharegpt.com

was it sick of talking to you and just went the appeasement route

ChatGPT got convinced monoliths typically have advantages that outweigh the concerns?