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

Sparks of Artificial General Intelligence: Early Experiments with GPT-4

arxiv.org

Post reply on HN
Print view View original

211–220 of 244 posts

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#211

Post by coldtea »

yed wrote: post #208
coldtea wrote: post #198

Earlier quoted context omitted.

> It's just a statistical model is the logical equivalent of human beings are just a bunch of atoms. Not exactly. One says "human beings are just a bunch of atoms" referring to the low level constituans (in a reductionistic way), but not making an accessment about the abilities emerging from those atoms in their interactions when in the form of a human. But when one says that GPT is "just a statistical model" they're…

> But when one says that GPT is "just a statistical model" they're implying a capacity cap of statistical models Except there is no “capacity cap” on statistical models, we have no idea what they are or are not capable of yet.

>Except there is no “capacity cap” on statistical models, we have no idea what they are or are not capable of yet

We do however have the knowledge that the human brain uses different model and topology, not just a bigger scale.

And we do have a good intuitition that scalling LLMs as they are (e.g. not changing the architecture) will give us more of the same kind of capabilities it currently has with the same limitations, not the kind we expect to match human thinking.

Also, empirically we do have an idea of "what they are or are not capable of yet". We had developed them, run them, and scaled them several times.

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#212

Post by SanderNL »

IanCal wrote: post #189
SanderNL wrote: post #166

Earlier quoted context omitted.

I think what we call "role-play" might be more integral to intelligence than we tend to give it credit for. Now I think of it, a "job description" could be a good prompt. If you start with a CEO-like job agent, that can think of what other jobs are necessary then you can bootstrap from there. "I want to produce and sell red bread" => "We are going to need a bakery, accountant, marketeer, etc." and then those are "com…

I think your comparison to a company is a really good mental model of a larger more capable collaborative structure. You can even have "hiring" and "firing" where it's deciding to create or remove roles.

I think so too. I see room for different types of AI having a seat in this "collaborative structure" as you say. I think I'm going to call companies that from now on by the way. Some AIs can specialize in "prompting" and pump out "workers" of varying effectiveness which indeed can be "hired" and "fired" as whatever performance metrics change.

I can see how more expensive and capable AIs get closer to the "executive seat" and lesser AIs - like what we now call GPTs - doing the grunt work. Interacting with humans and such, which is of course beneath the more powerful ones.

Using text - and thus providing a vehicle for the concepts it encodes - is brilliant. It enables cross-cutting communication between systems that otherwise have very little to do with each other. (GPTWolfram) As programmers we have a first-row seat on the code=data front. We are trained to see how text is able to be converted into action. Something I find most regular people are having trouble even visualizing. ("It's just text")

I guess we were on to something when we as humans started to talk to each other..

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#213

Post by ben_w »

euroderf wrote: post #155
throwaway4837 wrote: post #29

Earlier quoted context omitted.

Simple thought experiment: if I was a high school teacher and graded hundreds of exams, papers, and homework every month for years, would I develop the same opinion of my C-level students? The ones who have somewhat correct work, often inconsistent, clear that they're not thinking about the material, etc. Humans all start out like this, but as they grow they develop different masteries. Many of those masteries are ea…

So is Apple negotiating with some LLM mfr to integrate with realityOS ?

Given the endless cycle of "leaked" product details followed by "leaked" delays such that it's been 6-18 months away for the last 8 years or so, I now assume it only exists as fake product that Apple uses to find leakers.

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#214

Post by killerstorm »

golol wrote: post #178
Mike_12345 wrote: post #36

Earlier quoted context omitted.

GPT-4 is often overhyped and underhyped because few really understand it. It's not a Markov Chain or a fancy text predictor. It's a ~200 layer neural network that models a vast hierarchy of concepts through language. It has emergent properties that we don't yet understand.

it is a markov chain; At least the underlying decoder only transformer is.

"Markov chain" might mean:

* a kind of stochastic model * a "naive" realization of that model which directly counts frequencies of N-dimensional vectors

This naive implementation is sometimes used for language modeling, e.g. for the purpose of compression. So people might think you mean that particular implementation rather than a theoretical model.

This sort of a description can be unhelpful.

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#215

Post by killerstorm »

galleywest200 wrote: post #207
killerstorm wrote: post #184

Earlier quoted context omitted.

> how NOT intelligent they are In what sense? How do you define intelligence and what is missing? > finding so many failure modes. Failures do not indicate a lack of intelligence. People make errors all the time. We should differentiate failures which arise from poor training for a lack of general capabilities. There are people who learned a bad way to do something and struggle to change that. Should we consider them…

We are talking about a computer program that cannot make decisions. Often times human intelligence is compared with the types of decisions the person makes.

It can make decisions. E.g. if you give it a task, it can formulate sub-tasks, then execute them, e.g. by selecting which API calls to make.

Obviously, if you call the model, you define the scope for it, and it can only make decisions within that scope.

The scope is potentially unbounded. E.g. you can give it access to a terminal of a Linux machine which has an internet connection and tell it to do whatever it likes. (Given RLHF perhaps it would be better to give some positive goal, e.g. perform actions which would eventually produce knowledge.)

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#216

Post by killerstorm »

geophile wrote: post #194
killerstorm wrote: post #183

Earlier quoted context omitted.

> Why is that a stupid thought? What is so preposterous about "just statistics" Suppose you have N variables x_1, ..., x_10 and you want to predict y_1, ..., y_10. You know that each y_i depend on each x_i in a complex, non-linear way. How many samples would you need to to make sense of distribution? How does number of samples grow with N?

I have no idea what that has to do with the points you are responding to.

Statistics has two meanings:

1. A way to interpret math. E.g. given a computation you might interpret some values as probabilities. 2. A particular set of methods which people use to analyze information as well results of such analysis.

The problem with "just statistics" is that 99% of people would understand it as #2. But deep learning is very much not like "normal" statistics.

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#217

Post by nl »

golol wrote: post #206
nl wrote: post #199

Earlier quoted context omitted.

It's not. It can do in context learning, which Markov chains cannot do.

It is a Markov Chain on the state space {Tokens}^CtxWindow.

I don't think that's clear at all.

https://arxiv.org/abs/2212.10559 shows a LLM is doing gradient descent on the context window at inference time.

If it's learning relationships between concepts at runtime based on information in the context window then it seems about as useful to say it is a Markov chain as it is to say that a human is a Markov chain. Perhaps we are, but the "current state" is unmeasurably complex.

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#218

Post by anakaine »

parineum wrote: post #66
maxdoop wrote: post #22

Earlier quoted context omitted.

Our of curiosity, what is GPT-4 getting wrong so often? It’s prettily wild to my own , admittedly easily impressed, mind.

I've been trying to make it be a DM for a while now. It works seemingly well for a while but it's making everything up on the fly, which is fine as long as there's not state that needs to be maintained/remembered. Once you start trying to get it to do combat, it seems like it's working pretty well but eventually you realize that it's just making shit up behind the scenes. It'll tell you there are 4 goblins ambushing…

Have you noticed any improvements with GPT-4 regarding the continuity and persistence if story and assets? In unrelated areas that feel similar I've noticed GPT-4 keeping track a lot better.

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#219

Post by Nevermark »

bumbledraven wrote: post #63

GPT-4's proof of the infinitude of primes, as quoted in the paper, is wrong. Yes, I think I can, Though it might take a clever plan. I’ll start by noting Euclid’s proof, Which shows that primes aren’t just aloof. Assume we have a finite list, Of primes, and that none have been missed. Multiply them all together, And add one, just to be clever. The result will be a number, That has no prime factors, oh wonder! Not tru…

What does it mean if in demonstrating a potential artificial GI can’t understand a proof, a biological GI actually demonstrates they don’t understand the proof.

Joking aside … the approach of dismissing generality of intelligence based on the presence of mistakes seems to be flawed.

Re: Sparks of Artificial General Intelligence: Early Experiments with GPT-4

#220

Post by random_cynic »

outlace wrote: post #14

ChatGPT and its relatives are very very impressive on first impressions, but I've been using ChatGPT-3 and now 4 heavily every day since they became available to individuals and once you start using them this much it becomes very clear how NOT intelligent they are. It really just seems like extremely impressive statistical inference after this much use and finding so many failure modes. But it is still impressive how…

No one cares about your anecdotes and feelings. These people conducted thousands of systematic tests for at least six months. If you want to be taken seriously then do the same.
Post reply on HN