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

231–240 of 255 posts

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

#231

Post by mjburgess »

kelseyfrog wrote: post #229
mjburgess wrote: post #222

Earlier quoted context omitted.

I wasnt getting the sense it was worthwhile to engage, as my views werent being accurately understood. By I can address this. The meaning of words is, roughly, states of the world. If I say, "pass me the salt" that is satisfied if you, in fact, pass me the salt. If I say, "that tree is green" this is true if that tree which we are both talking about has the property of causing a perceptual state "seeming green" in bo…

Why would an LLM require salt?

The phrase is a common example in the analysis of speech acts:

https://chat.openai.com/share/280d6759-5ab4-49f3-ba60-6b37bc...

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

#232

Post by mjburgess »

kelseyfrog wrote: post #230
mjburgess wrote: post #161

Earlier quoted context omitted.

emotions regulate motivation, desire, action, behaviour etc. to be angry is for your sensory-motor system to be primed for aggression; it's for your cognitive systems to be narrowed and focused on analysing high-threat parts of your environment; it is for your memory-formulation to be modulated towards threat recollection etc. Sure, if an LLM's prompt "be angry" causes it to adopt a threat stance to its environment,…

Why should I discount a theory just to protect my ego? We've read countless stories about science only progressing when those with big egos die. It would only seem logical that eventually it will come for my own.

You're protecting your ego more when you compare people to LLMs, since that is your existing prejudice.

The sort of fashionable pseudo-scientific scientism in the belief that animals are alike digital electrical machines is a kind of egoism. It says, "the engineer of these machines (me!) knows all!"

The real hit to your ego is to suppose you are vastly more complex than you understand -- and this is why these engineers crop up and demand that there is nothing more to know than what they have already learned

this is an illusion of humility: these engineers take the implied nihilism of this view (that of the emptyness of animal life) as evidence that it is the humble one.

But, as ever nihilism, ends up being the most profound kind of arrogance, here its an ego-defence against the threat of their own ignorance.

And the threat is real: all you have ever learned about how to sequence transformations of natural numbers (all the algorithms of computer science) are of no use at all in the study of intelligence. What an injury to the ego!

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

#233

Post by famouswaffles »

sweetgiorni wrote: post #228
mjburgess wrote: post #158

Earlier quoted context omitted.

It isnt making mistakes ... its never actually doing it. Do you see a difference between the process of adding numbers and dividing by their count (taking a mean) and emitting numeric tokens which are most probable for a given input? The former is called "taking a mean" the latter isnt . This system never engages in any method to summarise financial data. It's method is always the same: to emit tokens most probable g…

To add to your point: try asking ChatGPT to do basic arithmetic on numbers it hasn't seen before. You'll see just how good it is at computation.

It's better (GPT-4) than you could manage without an external tool or pad. and that's after being severely hampered by tokenization. https://arxiv.org/abs/2310.02989

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

#234

Post by mjburgess »

wilg wrote: post #214
mjburgess wrote: post #102

Earlier quoted context omitted.

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.

I regard concepts as being essentially sensory-motor techniques for regulating animals --- so to not have that system is to not really have concepts.

LLMs model concepts using patterns of text. For example, if I say, "I wonder what the weather will be like tomorrow?" i employ my imagination to simulate a scenario -- this simulation is made by taking my concepts of weather, the world outside my door, and so on and combining them to create a "pseudo-sensory-motor reconstruction" of what my experience would be.

This reconstruction has a cognitive dimension (ie., the structure of my raw pseudo-sensory-motor experience as a quasi-logical form) which can be communicated (ie., taking a quasi-logical structuring and making it linguistic) using words that have a symbolic representation.

LLMs, by imitating patterns within this symbolic representation (a distant side effect of thinking) it seems as-if it is thinking with concepts.

But all this shows it that patterns of text can be constructed without using concepts at all.

This is the engineer's trick -- the magic latern. It's important to realise the trick, because the LLM doesnt know what weather is, and is not imagining anything when asked to generate a pattern of text against the prompt, "imagine a scenario where weather..."

Rather it was the humans who wrote its training data that engaged in these acts -- it is their shadows which are replayed by LLMs

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

#235

Post by kelseyfrog »

mjburgess wrote: post #232
kelseyfrog wrote: post #230

Earlier quoted context omitted.

Why should I discount a theory just to protect my ego? We've read countless stories about science only progressing when those with big egos die. It would only seem logical that eventually it will come for my own.

You're protecting your ego more when you compare people to LLMs, since that is your existing prejudice. The sort of fashionable pseudo-scientific scientism in the belief that animals are alike digital electrical machines is a kind of egoism. It says, "the engineer of these machines (me!) knows all!" The real hit to your ego is to suppose you are vastly more complex than you understand -- and this is why these enginee…

Sorry, but in my experience, I've seen "LLMs and human intelligence are incomparable," used to protect fragile egos more than critique them.

It sounds like your experience is different.

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

#236

Post by fragmede »

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…

The problem is the word "just".

Saying they just predict the next word in a sequence is where the statement jumps from being a straightforward factual scientific claim, to one that contains an opinion. After all, if it just predicts the next word, the unspoken implication is it can't be very that good.

It's a shallow dismissal of a collosal amount of work.

Do you consider "typist" an accurate description of your job?

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

#237

Post by SirMaster »

Timon3 wrote: post #49
SirMaster wrote: post #14

Am I the only one who doesn't think we should have to do stuff like this to get the best performance? The people who design and control the model should have this type of stuff baked in for us IMO.

This wasn't done deliberately. It's something the model picked up from it's training dataset. I'm sure the creators don't want it to be like this, but cleaning trillions of tokens of all examples is long and hard work. The alternative would be injecting these sentences into your prompts, which probably nobody really wants to happen.

I know it's not deliberate, but since they know this improves results, shouldn't they modify the service so that it takes advantage of this sort of improvement automatically?

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

#238

Post by krainboltgreene »

ignoramous wrote: post #138
krainboltgreene wrote: post #89

Earlier quoted context omitted.

I read through that and none of the section (or entire work) ever talk about the above discussion. Further I looked at some of the many citations of on that section and none of them suggest that the OP is right. In fact a few of them I know disagree.

Depends. Models are matrices of floats and so there's little chance an umbrella-term like "stochastic parrot" will never not stick, even when they already show signs of syntactic, semantic world-building capability ( https://www.arxiv-vanity.com/papers/2206.07682/ ). If you are like me (and them: https://archive.is/cZi83 ) and deem instruction following , chain-of-thought prompting , computational properties of LLMs…

Okay so just to confirm that section doesn't actually tell us anything about this and in fact this is all based on your own understanding of the mechanisms involved.

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

#239

Post by famouswaffles »

wilg wrote: post #214
mjburgess wrote: post #102

Earlier quoted context omitted.

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.

Because he doesn't believe a plane can fly without flapping wings and feathers. That's all this boils down to.

To him, plane flight must also be an "engineering trick" or in other words, his idea of flight has been divorced of any real meaning.

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

#240

Post by famouswaffles »

scotty79 wrote: post #218
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?

> 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 cl…

>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.

This isn't true. Wild how people will confidently say nonsense about things they obviously haven't actually tested.

GPT-4 can manage arbitrary arithmetic calculation closer to the real value than you could ever do without an external tool.

Post reply on HN