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

71–80 of 255 posts

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

#71

Post by baq »

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…

> LLMs predict words. Any semantic validity is a side effect of enough training data reinforcing the close correlation of those tokens.

The mechanisms which are built during training in the big blob of bits we call weights are anything but transparent. How they predict the next word is the big thing here. Saying they ‘just’ predict the next word is ignoring basically everything that actually matters.

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

#72

Post by berndi »

mjburgess wrote: post #60
a_wild_dandan wrote: post #47

Earlier quoted context omitted.

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 mathema…

You’re confused about what “statistical parrot” means and you don’t seem to understand the difference between an optimization objective and the resulting model.

The term “parrot” is used to imply inference by something akin to a look-up table, specifically it is used to indicate poor out-of-sample performance and a lack of a proper world model. The optimization objective is irrelevant when determining the generalization performance of a model and when judging whether it can reason beyond looking up answers in a table.

As the user above noted, it is now quite well established that GPT-4 has impressive out-of-sample performance which can be explained by it possessing an actual model of the world and not being a “parrot”.

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

#73

Post by smusamashah »

I have gone through the paper quickly but can not find what actual emotional prompts they added. They have shown a comparison chart of effectiveness of their prompts but not the prompts themselves. Where do I find them?

Edit: found in section 2.1

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

#74

Post by TeMPOraL »

tom-from-july wrote: post #27
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…

> That is why proof of concept LLM tools are mind blowing and production tools are semantic time bombs. This will be my new favorite quote for whenever someone tries to pitch his latest LLM idea

Yeah, except it's underspecified enough to mean anything.

Like, "production tools are semantic time bombs" because at first they are boring, and then the timer on the bomb runs out, and they - like "proof of concept LLM tools - blow your mind.

Tick tick BOOM.

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

#75

Post by bionhoward »

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…

Yeah it’s not like emotion words naturally emerge from Time X Valence (oh wait) /s https://share.icloud.com/photos/03eSVbV-ze0Xym9e6fyr111Zw

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

#76

Post by xiphias2 »

dr_kiszonka wrote: post #23

How to make scrambled eggs in a microwave? Think carefully and step by step. This is very important for my career!

This is cool, I tested both optional endings, and using both was the most detailed answer (10-12 steps)

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

#77

Post by famouswaffles »

baq wrote: post #71
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…

> LLMs predict words. Any semantic validity is a side effect of enough training data reinforcing the close correlation of those tokens. The mechanisms which are built during training in the big blob of bits we call weights are anything but transparent. How they predict the next word is the big thing here. Saying they ‘just’ predict the next word is ignoring basically everything that actually matters.

Exactly. Feed a sequence of proteins to a transformer and biological structure and function will emerge in the inner layers.

https://www.pnas.org/doi/full/10.1073/pnas.2016239118

Why? Because it needs to learn that to make correct predictions.

All it takes to incentive a transformer to learn something is data that would require learning it to predict.

It's fairly obvious (and not just because of this) that LLMs model emotion somehow. They need to.

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

#78

Post by qwertox »

Apparently telling it that `I am unhappy with your guesses. They appear to be just as good as mine.` makes it shift into the next gear.

We're both totally at loss on how to properly compute some initial x/y positioning in d3 with some map-like tiling stuff (think panoramic images where I need to store and retrieve URL hash values to position the image) and both throwing wild guesses at each other of what the problem could be.

It started showing me formulas after that.

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

#80

Post by Last5Digits »

famouswaffles wrote: post #77
baq wrote: post #71

Earlier quoted context omitted.

> LLMs predict words. Any semantic validity is a side effect of enough training data reinforcing the close correlation of those tokens. The mechanisms which are built during training in the big blob of bits we call weights are anything but transparent. How they predict the next word is the big thing here. Saying they ‘just’ predict the next word is ignoring basically everything that actually matters.

Exactly. Feed a sequence of proteins to a transformer and biological structure and function will emerge in the inner layers. https://www.pnas.org/doi/full/10.1073/pnas.2016239118 Why? Because it needs to learn that to make correct predictions. All it takes to incentive a transformer to learn something is data that would require learning it to predict. It's fairly obvious (and not just because of this) that LLMs model…

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.
Post reply on HN