Artificial intelligence systems found to excel at imitation, but not innovation • Hacker News Board

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#91

Post by simbolit » Wed, Dec 13, 2023, 9:27 PM UTC

hackinthebochs wrote: post #44

gibsonf1 wrote: post #29

How could a statistical system trained on data display intelligence let alone innovation? It's just very good statistics at the end of the day. Artificial Perception is about as far as you can get with ml/dl tech if you point the sensors at the world of space-time. If you point it at words, as with LLM, you get statistics about words - that is, no actual understanding of what the words model in the minds of the origi…

LLMs aren't statistical systems in any substantive sense. They are deterministic programs over the input sequence. They capture statistical relationships about words, but so do human minds. That they are sensitive to statistical relationships does not discount their ability to understand.

> That they are sensitive to statistical relationships does not discount their ability to understand.

The claim is that nothing but sensitivity to statistical relationships somehow leads to an ability to understand?

I am not going to believe it.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#92

Post by visarga » Wed, Dec 13, 2023, 9:45 PM UTC

kayo_20211030 wrote: post #38

visarga wrote: post #23

Earlier quoted context omitted.

> innovation needs will feedback Innovation needs feedback, otherwise all ML papers would not run evaluations. Humans come up with 100 stupid ideas that fail at eval before stumbling onto a good one. Improving innovation is a matter of putting the AI inside a system that can provide feedback. Remember AlphaGo move 37? The model created feedback by self-play games, and beat all humans at Go - feedback made it really m…

Known feedback is already encoded in the inputs. I don't remember move 37, but I think I know to what you're referring. Playing games against itself, and remembering, meant that outcomes (i.e. feedback) were already encoded in the inputs for the next run. It had a goal - win the game - but it's hard to claim that the machine itself was wilful. The people who designed it were. But, the machine was just an incredibly e…

I got a good answer for you, but it's long winded. It all started with the first chemical self-replicator. These replicators consume and compete for resources so they evolve. The will to survive is encoded in their reward systems, their senses and body is adapted to their niche, they all come from evolution. Humans have these basic instincts as well, and we don't need anyone to prompt us, we already got an in-built goal. AIs so far have been created at the will of humans, but the same instincts can emerge if AI can be a self replicator.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#93

Post by Jensson » Wed, Dec 13, 2023, 9:48 PM UTC

corethree wrote: post #47

vemv wrote: post #12

One can casually observe that LLMs quite excel at composition though: gluing together pieces of knowledge in ways no one did before (examples: a program that does X using language Y, a painting that mashes up two themes). Most knowledge workers' activities aren't innovative or imitational - similarly, we compose stuff, so LLMs are a fair competitor.

All innovation is composition plus random generation. Which LLMs already do. LLMs have a rudimentary form of innovation. It's not quite as good as humans but it's getting there.

Innovation isn't just testing something new, it is a new thing that improves something or is valuable in some way. So doing random combinations of things isn't innovation, it is just noise.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#94

Post by visarga » Wed, Dec 13, 2023, 9:49 PM UTC

k__ wrote: post #36

visarga wrote: post #21

Earlier quoted context omitted.

AI is going to be in a different place in 5 or 10 years.

Hopefully. However, I had the impression there doesn't exist enough training data to make that place different in a meaningful way. Still, I think, letting some skilled UX designers loose on input methods could improve things quite a bit, even if the models won't get "smarter".

I did a back-of-the envelope calculation, OpenAI has 100M monthly active users, assume 10K tokens per user per month usage ($20 would pay for 600K tokens on the API) then they generate 1T tokens per month.

This dataset would be focused on human interests (in domain for users) and containing AI errors (in domain for the model). It's LLM empowered with human in the loop and tools - code execution, search, APIs. So it is a good basis for the next dataset. I think OpenAI has amassed about as much chat log text as there is organic data was used for GPT-4, which was rumoured to be 13T tokens.

It's surprising how much synthetic data can be generated per year. And OpenAI can do this with human in the loop for free, if the paying users pay for everyone. We then benefit 6-12 months later when the open source models trained with data exfiltrated from OpenAI models catch up.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#95

Post by kromem » Wed, Dec 13, 2023, 10:51 PM UTC

corethree wrote: post #66

HarHarVeryFunny wrote: post #63

Earlier quoted context omitted.

I think that a lot of hallucination might just be due to not planning ahead - basically a case of running mouth before engaging brain, and then being in a situation where one has uttered a bunch of nonsense - basically backed oneself into a conversational corner. A human might catch themselves with "err, never mind, forget that!", but the LLM's only recourse is to continue extrapolating the nonsense the only way it k…

As the query prompter you are in control of the feedback loop. You can ask the AI to re-examine it's output to catch errors just as a human would do for himself. Practically speaking this does work to a limited extent. Sometimes the AI just sticks with it's guns and runs with it just like a human might.

Tip: If you are adding a self-critique step, show the model the initial output as if it is evaluating something from someone else (i.e. "grade this answer from a student") as opposed to from itself (i.e. "you wrote this, is it really correct?").

As you correctly note, humans have a problem with admitting fault. Especially the case online. But humans online are very ready to correct others.

That's exactly the kind of larger abstract pattern in the data a model would emulate.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#96

Post by kromem » Wed, Dec 13, 2023, 10:57 PM UTC

MeImCounting wrote: post #39

As I see it the AI schism is more about the debate between functionalism/computationalism and the idea that the chinese room thought experiment was an argument for, "biological naturalism". There is a lot of effort dedicated to showing that AIs dont have some innate quality called "consciousnes" or "sentience" or what have you. There is not just a lot of effort to show that, but also to show that that is somehow a li…

The sentience thing is such a red herring that gets too much time spent on it. The topic is a pariah in neuroscience even, but we are going to discuss it as nauseum for AI when there's barely any research on its mechanics in humans?

The far more interesting topic is not if sentience is occurring (it's almost certainly not yet), but if it is being accurately modeled by a non-sentient agent.

A LLM may not have a subjective experience of emotions, but it very likely is modeling some kind of emotional tracking from the input of massive amounts of emotional language similar to how Othello-GPT modeled an Othello board from the input of legal moves.

To me, that's a far more interesting nuance to explore than the red herring binary of "sentient or not."

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#97

Post by 1vuio0pswjnm7 » Wed, Dec 13, 2023, 10:58 PM UTC

Why would anyone think "AI" would excel at innovation. Honest question.

Innovation is defined as "introducing something new".

"AI" is autocomplete and autoselection based on _past_ input.

"AI" is regurgitating, rehashing or recombining something that has come before. It might do this in a new (=yet untried) way, but it cannot "introduce something new". Only man can do that.

Make no mistake, recombining can be useful. With AI we can screen a large number of possible recombinations that would be practically infeasible without it.

But the only way something new can be "introduced" is for man to produce new data for "AI" to process. "AI" is always one step behind.

There is no way for "AI" to have "new thoughts". All "thinking" by AI is always just regurgitation, rehashing or recombination of man's past thoughts.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#98

Post by gamacodre » Wed, Dec 13, 2023, 11:53 PM UTC

corethree wrote: post #84

sidlls wrote: post #80

Earlier quoted context omitted.

All innovation is composition plus random generation. That’s a fairly bold claim. What supporting data do you have to justify it?

There are two types of statements that can be made in this world. One is data driven based on evidence. The other is logic driven based on axioms and the logical implications of said axioms. My statement is derived from the later. Therefore evidence is unnecessary. It is niave to blindly faith in all truth in the hands of data without understanding nuances between the relationship of data and logic. If you have a pur…

Couple of thoughts:

You are assuming that a pure function has access to some source of randomness. I think that's a bit less than pure unless a source of randomness is one of the inputs. Doesn't change your point at all, but it distracted me from what you were saying.

Speaking as someone who has occasionally invented new things, I think you underestimate the variety of outputs that can arise from a lifetime of accumulated experiential cruft and a situational fitting function (e.g. "we need a working teleporter to stay competitive"), without bringing in randomness at all.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#99

Post by DrSiemer » Thu, Dec 14, 2023, 12:21 AM UTC

1vuio0pswjnm7 wrote: post #97

Why would anyone think "AI" would excel at innovation. Honest question. Innovation is defined as "introducing something new". "AI" is autocomplete and autoselection based on _past_ input. "AI" is regurgitating, rehashing or recombining something that has come before. It might do this in a new (=yet untried) way, but it cannot "introduce something new". Only man can do that. Make no mistake, recombining can be useful.…

The current generation of language models is only rewarded for how well it can mimic human communication. There is no incentive for innovation.

Give it a few generations and we'll add that. A current model could already run a chain of "what if" scenarios and grade the results for how unique and how beneficial they are. Supercharge that and you'll probably end up with something useful.

Everything is a remix.

Re: Artificial intelligence systems found to excel at imitation, but not innovation

#100

Post by corethree » Thu, Dec 14, 2023, 2:36 AM UTC

gamacodre wrote: post #98

corethree wrote: post #84

Earlier quoted context omitted.

There are two types of statements that can be made in this world. One is data driven based on evidence. The other is logic driven based on axioms and the logical implications of said axioms. My statement is derived from the later. Therefore evidence is unnecessary. It is niave to blindly faith in all truth in the hands of data without understanding nuances between the relationship of data and logic. If you have a pur…

Couple of thoughts: You are assuming that a pure function has access to some source of randomness. I think that's a bit less than pure unless a source of randomness is one of the inputs. Doesn't change your point at all, but it distracted me from what you were saying. Speaking as someone who has occasionally invented new things, I think you underestimate the variety of outputs that can arise from a lifetime of accumu…

Right and in this case you describe you'd be composing ideas to form new ideas right? And you'd do it within the parameters of the situation. So you would only select ideas that are in the bounds of "teleportation".

But then out of that set of ideas how do you choose which ideas to compose to formulate the new idea?

So you have idea A and idea B. And you randomly formulate a new compositional rule as an idea: A + B. Because A + B didn't exist as an idea before, it was randomly generated by you.

>I think you underestimate the variety of outputs that can arise from a lifetime of accumulated experiential cruft and a situational fitting function

Well it's actually possible to mathematically calculate the total amount of compositions If I know the total amount of experimental cruft you have accumulated in your lifetime. It's a combinatorics problem and the output of that is, you're right, extremely huge. We have fitting functions that reduce it, of course, but within this fitting function we're just iterating through all the remaining possibilities or aka "randomly generating" the new idea.