Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Telling GPT-4 you're scared or under pressure improves performance

aimodels.substack.com

Post reply on HN
Print view View original

211–220 of 255 posts

Re: Telling GPT-4 you're scared or under pressure improves performance

#211

Post by omarfarooq »

sebastiennight wrote: post #184
omarfarooq wrote: post #178

Earlier quoted context omitted.

ChatGPT responds: https://chat.openai.com/c/7c4ae31a-3391-4496-8db3-a92a58a4e1...

I'm not sure how much value such a comment adds (especially since it's gated behind the OpenAI login). Can you elaborate a bit?

It removes the idiosyncrasies and delivers the stated thoughts in a direct and clear manner. And the specific prompt for the message you replied to also includes a rebuttal, thereby showing an opposing perspective.

Here is a pastebin: https://pastebin.com/DCqgxx8E

Re: Telling GPT-4 you're scared or under pressure improves performance

#213

Post by kybernetikos »

mjburgess wrote: post #200
kybernetikos wrote: post #185

Earlier quoted context omitted.

A neural network with a hidden layer can approximate F=GMm/r^2 if given appropriate training input. I'm not clear on what you're saying. Is it that LLMs specifically don't have this same property of being able to approximate such functions? Is it that a neural network wouldn't learn that model if you gave it real world measurements (because I think it would, but such a thing should be fairly easily testable)?

most formulas, including that one, are neural networks (such is the absurdity of the term) -- so it is trivially learnable The issue is that to prepare the dataset from which that formula is learnt requires already knowing it. This is the triviality of applications of universal function approximators to science -- empirical data modelling isnt new, and neural networks are just one example of it; not all that special.…

> The issue is that to prepare the dataset from which that formula is learnt requires already knowing it.

Why is that? Are you saying that universal function approximators can't invent higher level models? Isn't that what happens in the hidden layers when an neural network is trying to predict things that require those higher level models?

Re: Telling GPT-4 you're scared or under pressure improves performance

#214

Post by wilg »

mjburgess wrote: post #102
Last5Digits wrote: post #80

Earlier quoted context omitted.

This is what makes these discussions so infuriating. Saying that LLMs "just predict the next word" is about as insightful as saying that computers "just do a bunch of logical operations" - neither point constraints the possible capabilities of the systems they refer to in any meaningful way.

Sure it does. It perfectly deliniates it. LLMs are not: sensitive to causal structure, dynamically adapting to environmental changes, growing, developing sensory-motor capacities, they are not with us in our environment, they are not: expressing desires, preferences, intentions, beliefs, motivations, etc. And so on. To say, "they just predict the next word" is literally to say that all apparent functions of an LLM ar…

You are correct, but also why would you say this is an "engineering trick"? The interesting part about LLMs is what they are capable of doing and how they are constructed. It's not a great argument to say that they don't have motor functions (!?!?). It shows the incredible ability of deep learning to discover and operate with high-level concepts.

Re: Telling GPT-4 you're scared or under pressure improves performance

#215

Post by 3c6bYDXLMj »

intended wrote: post #19

I think this is the entry point needed to get peoples attention and explain: LLMs aren’t people, and emergent properties are being over extended. If LLMs are showing “better” performance when there are tokens that humans read as emotionally salient - Then the underlying text it’s trained on shows humans give better answers when emotionally salient context is provided. LLMs predict words. Any semantic validity is a si…

I am growing completely and utterly tired of this. It’s been…how many months now? Years? And there are still plenty of people in this and other communities getting by high-fiving each other pulling out the same old “LLMs aren’t human, did um, actually know, that LLMs were TRAINED on output from HUMANS?” line.

We all know. Nobody is on the other side of this. Who are you educating? And then we end up digging a little deeper and it turns out that the argument is ACTUALLY that the above is an indication of the technology’s usefulness. Which is only something that’s going to be problem by…how much use people get out of it.

Re: Telling GPT-4 you're scared or under pressure improves performance

#216

Post by intended »

3c6bYDXLMj wrote: post #215
intended wrote: post #19

I think this is the entry point needed to get peoples attention and explain: LLMs aren’t people, and emergent properties are being over extended. If LLMs are showing “better” performance when there are tokens that humans read as emotionally salient - Then the underlying text it’s trained on shows humans give better answers when emotionally salient context is provided. LLMs predict words. Any semantic validity is a si…

I am growing completely and utterly tired of this. It’s been…how many months now? Years? And there are still plenty of people in this and other communities getting by high-fiving each other pulling out the same old “LLMs aren’t human, did um, actually know, that LLMs were TRAINED on output from HUMANS?” line. We all know. Nobody is on the other side of this. Who are you educating? And then we end up digging a little…

The point here is about the use people get out of it. Proof of concepts are being made that rely on misunderstanding the abilities of LLMs

Re: Telling GPT-4 you're scared or under pressure improves performance

#217

Post by scotty79 »

williamcotton wrote: post #201
intended wrote: post #193

Earlier quoted context omitted.

This paper was published, and the distinct jump was found to be a measurement artifact. https://hai.stanford.edu/news/ais-ostensible-emergent-abilit...

Sweet, now complete and present a meta-analysis of all such papers about the topic and maybe I’ll find it more convincing than a cherry-picked publication that supports your preexisting position on the topic!

Metanalysis of bunch of weak papers is worth nothing when confronted with one strong study. I hoped covid drug research has tought us all that.

Re: Telling GPT-4 you're scared or under pressure improves performance

#218

Post by scotty79 »

a_wild_dandan wrote: post #47
intended wrote: post #19

I think this is the entry point needed to get peoples attention and explain: LLMs aren’t people, and emergent properties are being over extended. If LLMs are showing “better” performance when there are tokens that humans read as emotionally salient - Then the underlying text it’s trained on shows humans give better answers when emotionally salient context is provided. LLMs predict words. Any semantic validity is a si…

The “statistical parrot” assertion is pretty thoroughly disproven by this point, but suppose we ignore the literature and just assume it’s true: what does it matter? “Real” people are time bombs too, for instance. Is there some predictive power that we gain by reducing LLM skills to mere token production side effects?

> Is there some predictive power that we gain by reducing LLM skills to mere token production side effects?

Yes. We can infer from this that out of sample hallucinations will be closer to the desired productions when measured by token productions metrics than by any other more informative domain relevant metrics.

Which is exactly the case.

If you ask llm to solve a math problem it haven't seen its response will be closer to the desired solution in its linguistic form rather than in mathematical meaning.

Re: Telling GPT-4 you're scared or under pressure improves performance

#219

Post by didibus »

mjburgess wrote: post #102
Last5Digits wrote: post #80

Earlier quoted context omitted.

This is what makes these discussions so infuriating. Saying that LLMs "just predict the next word" is about as insightful as saying that computers "just do a bunch of logical operations" - neither point constraints the possible capabilities of the systems they refer to in any meaningful way.

Sure it does. It perfectly deliniates it. LLMs are not: sensitive to causal structure, dynamically adapting to environmental changes, growing, developing sensory-motor capacities, they are not with us in our environment, they are not: expressing desires, preferences, intentions, beliefs, motivations, etc. And so on. To say, "they just predict the next word" is literally to say that all apparent functions of an LLM ar…

You didn't really address what og_kalu brought up.

Which is that, it's possible that the model learns human like thinking, because that's the best way to accurately predict the human response itself.

I generally agree with you, but still, I do think that this is the current question. What it is the model is learning that it then uses for predictions?

Because you're assuming it's learning some purely token correlation, like these tokens followed X percent of the time, so that's the response. But it's possible it's learning at a lower level, and understanding the meaning and why those tokens follow in these scenarios, and then applying that same meaning to reasoning process over new tokens in order to predict which one would follow.

I'm skeptical of this, but I do not believe that we really know for sure yet.

Re: Telling GPT-4 you're scared or under pressure improves performance

#220

Post by didibus »

mjburgess wrote: post #130
Last5Digits wrote: post #125

Earlier quoted context omitted.

Cameras don't have eyeballs, Microphones don't have hair cells, Speakers don't have vocal cords, processors don't don't do arithmetic with neurons, yet we all agree that they are capable of emulating the meaningful aspects of these functions. All of your claims are either incorrect (not adapting, expressing desires, beliefs, preferences, ...) or fail to eliminate irrelevant differences. If we're to have any sensible…

> an implementation detail Yip, so I deny this premise. I take it to be the heart of the matter. > we might as well throw away 80% of our current scientific understanding Yip, i'd be down for that. Though maybe i'd say, 30-40%. Science in the strongest sense has no theory-building need for statistics. Those areas of science which have only statistical models, and not causal-ontological ones aren't science -- and i'd…

> Consider plato's cave. How do scientists determine what causes the shadows? They build vases, puppets, etc. and compare-and-contrast then eliminate the ones they've created which do not match

I feel like here you're limiting yourself to a LLM as an isolated thing.

But if you start to explore agents for example. You could have LLMs build vases, puppets, etc., then compare and contrast, and then eliminate the ones they've created which do not match.

So the potential isn't just that the LLM straight up gives you some prediction of the problem. But as it can emulate human processes to human accuracy, such as identifying which vase does and doesn't match, and how to go about creating a variety of vases, etc. Well you might very be able to use it for what you're describing as "proper scientific inquiry".

Post reply on HN