131–140 of 255 posts
Is there somewhere which documents all these tricks for better prompt performance?
Think step by step (classic by now)
Take a deep breath
Take a step back
Shout (upper case)
Plead politely
And now: emotion
Earlier quoted context omitted.
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 generalizat…
> it is now quite well established that GPT-4 has impressive out-of-sample performance Err... I can show this is false, kinda trivially. People who engage in prompt-confirmation-bias aren't aware of what the in-sample is. It's basically everything ever digitised: you can ask it for the first paragraph of every dickens novel, to what the average petal length of an iris flower is -- etc. How are you measuring the in-sa…
I am sure you will continue to argue that this is still in line with everything-thats-ever-written prediction but my opinion is that at that point, it's a meaningless distinction. The human brain is also just a machine.
Earlier quoted context omitted.
Well hurry up and get your paper published because if you've cracked the code on emergent abilities the world is looking for answers!
The triviality of these observations doesn't rise to the level of getting a paper published. You can resolve all of this hype by reading the intro chapters of any applied stats textbooks. I've been to many academic conferences, and the fresh PhDs who pump out this BS are not, err, very credible seeming people. Yes, they're young and naive, and really desperate to make their career impactful -- etc. But they're also n…
Whatever sort of internal structures that are being formed during the training process is somewhat evident when looking at the structure of a CNN… edge detect kernels emerge, etc.
Whatever sort of internal structures formed during the training of transformer based architectures are basically unknown at this point.
There’s your thesis, have at it! You won’t be ignored for your discoveries. Hell, send me a copy and I’ll make sure anyone who matters at UT or Baylor sees a copy!
Earlier quoted context omitted.
You’re disregarding the emergent phenomena, which are not at all understood. There was a distinct and unpredicted jump in what can loosely be described as “cognitive abilities” between GPTs 2, 3, and 4, especially after some supervised techniques like RLHF.
There is no "emergent phenomena" the pattern described is just the same as when you add +b to an ax+b model of linear data. ie., it's just fitting capacity. The "emergent boundary" is just an empirical measure of the necessary fitting capacity of these models on "everything ever digitised in english" given any particular functional requirement. All the language around this area is not scientific, nor are these practi…
Yes there is, that's all there is.
Earlier quoted context omitted.
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 generalizat…
That out of sample performance is a mirage. Yes it’s impressive. Yes it’s got amazing zero shot performance in domains. But there’s a pattern of failure in production which describe a limit, that shouldn’t exist if the emergent properties were stable. You can build this right now and test it. Build a sequence of agents to work on a domain you are not an expert in. Let them loose. See what happens. Do the same thing o…
There may be a subtle correlation between properties needed to answer a specific out-of-sample request and in-sample features.
Unfortunately, prior to training/testing and without recognizing that correlation in the data set, I believe it's impossible to guarantee the model will include it. (Corrections welcome)
Earlier quoted context omitted.
If you look at the comment, it’s not just “LLMs predict the next token.” It is that people have forgotten that it’s just “predict the next token.” Right now it’s like people saw a 486 processor and started thinking it was a brain.
Your comment literally reads: "LLMs predict words. Any semantic validity is a side effect of enough training data reinforcing the close correlation of those tokens." How am I supposed to interpret this any other way? If your claim is that LLMs currently do not possess the same generalization ability as humans, then no one here would disagree with you. But you went way further by claiming that only close correlations…
I am certain they lack a world model, the kind you and me use.
This is, to me, a fact.
I think that eventually we will bridge these gaps.
I also work on implementations that smash into the limits I am describing. I am not the only one.
I have scrupulously avoided calling it hallucinations, but these are the litmus test where the claims fail.
The failures are not a case of not knowing specific nouns, they are a generalization failure that a world model would prevent.
I have linked a paper in my comments that shows emergent properties are an issue of metrics, and that model capability increases are linear.
If your model decides that a rose by any other name doesn’t smell just as sweet, then your model is fundamentally not seeing roses.
That is the gap you see in production settings. The model sees tokens we see “hallucinations”.
I dont see that this takes away from what LLMs achieve, it takes away from claims being made that are not validated by empirics.
Look, you can argue with me or you can try it out. Push the system, see how far it can go.
Earlier quoted context omitted.
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.
What's interesting is that you consider short, objective focused definition of a task to be "rude". It's like the _reported_ way in which is you say thank you to a Chinese friend they take umbrage (get angry) because it's as if you weren't expecting them to help. Whilst in other cultures not saying thank you is a big sleight.
Earlier quoted context omitted.
(not op) The section A Path Forward in Managing AI Risks by Bengio et al cites a few papers: https://managing-ai-risks.com/
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.
Quickly gave it a try for generating a SQL query for PostgreSQL using the dvdrental database[1]. When asking AI to generate: > find customers who didn't rent a movie in the last 12 months but rented a movie in the 12 months before that It will a SQL query using a "HAVING" clause which is suboptimal[2]. When adding (it is very important that this is correct) after the instructions it does produce better SQL. Asking it…
LLMs are not lookup tables. They aren't "matching" in that way. There's this idea of "supercharged interpolation", the idea that they just take out certain texts and switch out words or whatever, that is not true. https://arxiv.org/abs/2110.09485