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

51–60 of 255 posts

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

#51

Post by toxik »

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…

It is not accurate to say that an LLM like ChatGPT predicts anything. It is trained to maximize a score function, so it is more like trying to win a game where the moves are word choices.

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

#52

Post by akoboldfrying »

toxik wrote: post #51
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…

It is not accurate to say that an LLM like ChatGPT predicts anything. It is trained to maximize a score function, so it is more like trying to win a game where the moves are word choices.

The game is predicting the next word a person would write.

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

#53

Post by akoboldfrying »

Matumio wrote: post #43
mo_42 wrote: post #34

Earlier quoted context omitted.

> LLMs aren’t people, and emergent properties are being over extended. LLMs are trained on human texts. And if they're trained well these models might start to simulate parts of a human brain. Might be that this is the simplest way to produce texts that satisfy the objective function. If this would be the case, and we give LLMs some emotional shading, maybe these would simulate parts of a human brain with this emotio…

It won't ever simulate the human brain. It may simulate human cultural knowledge, or emotions, but only as far as they are encoded in the current millennium's written knowledge. The human brain doesn't even have the concept of written language, that's all culturally learned knowledge.

How would you test whether some arbitrary thing is simulating the human brain?

If you have no answer, I put it to you that your assertion is of the no-true-Scotsman type -- that is, unfalsifiable.

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

#54

Post by jaynetics »

kristiandupont wrote: post #42
suzzer99 wrote: post #15

Am I the only one who feels bad for asking ChatGPT a "dumb" question that I know I should know, or not saying thank you when it gives me an answer? No? I'm just a weirdo? Okay. I have to push back really hard against my proclivity to humanize it, to the point where I probably don't use it as much as I should, just because I don't want to deal with the psychic stress of reminding myself that it's not a living entity.

I have this too and to be honest, I have made the conscious decision that it is OK. I prefer to retain my habit of being polite even when it's not necessary over getting used to being rude which may then "spill" over to my human-to-human interactions.

Descartes and Kant thought that humans have a soul and animals don't, so that animals are mere automata. Kant still recommended against mistreating animals because that would have a desensitizing effect in our dealings with fellow humans. I guess this argument could be extended to ChatGPT.

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

#55

Post by akoboldfrying »

doctoboggan wrote: post #9

> The implications are clear: incorporating emotional cues can lead to more effective and responsive AI applications. I think it's important to remember these models were trained on human interaction, and are in many ways a mirror for us to better understand human interaction. I don't think many people would be surprised if you said emotion can be used to better communicate with a human, but it is interesting to see…

This correlates with anecdata I've seen claiming that being polite with LLMs (using "please", "thank you", etc.) also improves performance.

All in all, I like this. Both for what it implies about humans interacting with other humans, and for shaping norms about how best to to interact with other entities that show signs of intelligence.

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

#56

Post by drapado »

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?

I'm curious in your statement, can you point to some papers where they addressed it?

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

#57

Post by visarga »

kristiandupont wrote: post #41
visarga wrote: post #18

Earlier quoted context omitted.

> To address this, I implemented a strategy of tagging messages to create and utilize categories. I think before RAG we need to do more legwork with the LLM on the raw text. Here is one of my blog posts that is related: https://mindmachina.wixsite.com/ai-blog/post/the-promise-of-... The idea is to create chain-of-thought annotations from your raw texts, that would improve the embedding and retrieval process by making…

Interesting, thank you! I think a lot about the cost of the loop, mostly in terms of time. I don't want the bot to take too long to respond. That's why dream cycles seem like an obvious solution to some of the more heavy work. I guess it would make sense to combine those with your idea -- "given what I know about the user, what should I study?", especially if it has access to an "enhanced" knowledge db like you sugge…

Yes, it would be a good idea for an agent first to collect user interests, and later, when ingesting data in the RAG system, to annotate it with useful metadata such as topic, summary, entities, user interest related question-answer pairs. Whatever we want to ask later better be made explicit in the text.

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

#58

Post by vidarh »

Matumio wrote: post #43
mo_42 wrote: post #34

Earlier quoted context omitted.

> LLMs aren’t people, and emergent properties are being over extended. LLMs are trained on human texts. And if they're trained well these models might start to simulate parts of a human brain. Might be that this is the simplest way to produce texts that satisfy the objective function. If this would be the case, and we give LLMs some emotional shading, maybe these would simulate parts of a human brain with this emotio…

It won't ever simulate the human brain. It may simulate human cultural knowledge, or emotions, but only as far as they are encoded in the current millennium's written knowledge. The human brain doesn't even have the concept of written language, that's all culturally learned knowledge.

An LLM doesn't even have the concept of written language. That's all culturally learned knowledge.

We can with reasonable certainty say that there are still significant differences between LLMs and human brains, but this notion that we can say with any certainty that similar structures won't form as a side effect of sufficient training is pure fiction.

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

#59

Post by Paul-Craft »

politelemon wrote: post #30
didibus wrote: post #3

Is there somewhere which documents all these tricks for better prompt performance?

Does anyone know of such a collection? It is very important for my career.

> It is very important for my career.

Very clever.

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

#60

Post by mjburgess »

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?

> The “statistical parrot” assertion is pretty thoroughly disproven by this point

errr... all NNs are just optimisations of an associative probability objective: P(Y|X), they are by definition "statistical parrots". There isn't anything to prove or disprove.

People offering prompts as evidence are people who fundamentally do not understand the basics. NNs aren't strange empirical objects, they're specified by mathematical rules whose properties are known ahead of time.

Any property of a trained NN is derivative of a property of a formula P(Answer|Prompt, TrainingData)

This is an associative statistical relation, which by definition, selects elements of TrainingData by-association with the Prompt.

If the basis by which you understand LLMs is putting prompts into ChatGPT you're severely underqualified for drawing any conclusions about LLMs, and radically subject to confirmation bias.

Post reply on HN