[circuit10]

"those non-problems"

Why is that a non-problem? It's a really important concern that we need to take more seriously

I pasted this from another comment I wrote but:

The concerns about AI taking over the world are valid and important; even if they sound silly at first, there is some very solid reasoning behind it.

See https://youtu.be/tcdVC4e6EV4 for a really interesting video on why a theoretical superintelligent AI would be dangerous, and when you factor in that these models could self-improve and approach that level of intelligence it gets worrying…

[JohnFen]

I don't think the reasoning is solid at all. I mean yes, a theoretical superintelligent AI would be very dangerous, but I see exactly no reason to think that current models could get there.

[manojlds]

Yeah feels a bit like we invent planes and worry about wormholes and time travel.

[circuit10]

I don’t think we’re as far off as you think

[tester457]

People had no reason to believe that today's models would exist.

We are on this part of the ai takeoff graph. https://waitbutwhy.com/2015/01/artificial-intelligence-revol...

[geraneum]

> People had no reason to believe that today's models would exist.

People had no reason to believe one day we would finally understand what causes the thunder. We finally did, and it is not made by Zeus.

[JohnFen]

That's not exactly true. There was plenty of reason to believe that. The only question was what the timeline would be.

[sebzim4500]

Personally, I wasn't expecting anything as good as GPT-4 so soon. So I no longer have any real confidence in how far away 'real AI' is, whatever that means.

I would not be shocked to find out that AGI (using Altman's definition) is more than 50 years away, but I also would not be shocked if it came in 5.

It's really hard to know how scared to be, I think that rationally I should be pretty terrified but I'm not.

[circuit10]

Well hardware and parameter count are scaling exponentially, so it seems very feasible that it could happen very soon. Of course it's possible that we'll hit a wall somewhere but it seems that just scaling current models up could be enough to get to the point where they can self-improve or gain more compute for themselves

[matthewfcarlson]

We've been out of exponential territory for a few years now (https://en.wikipedia.org/wiki/Moore%27s_law). Yes, we are still bounding forward at a crazy pace, but I think the pace is slowing down somewhat

[rolisz]

Hardware isn't scaling exponentially anymore (Moore's law is dead). Parameter count isn't really scaling exponentially anymore either. GPT3 had 175b parameters 3 years ago. There are some attempts at training 1 trillion parameter models, but they are not better than GPT3.

[lhl]

While I agree we probably aren't getting exponentially increasing parameter counts (GPT4 is by all accounts 1T paramaters and of course, it is significantly better than GPT3) we are still seeing lots of improvements - 3.5 is much better than 3, based "just" on InstructGPT/RLHF training. Models are getting better as well - LLaMA 30B beats/matches GPT-3 on raw eval benchmarks at 1/6 the parameter count.

We're also seeing lots of optimizations with new models (RoPE/RoPER embedding, Swish/GeLU activation, Flash Attention, etc) but I think some the most interesting gains we'll be seeing soon is with inference-optimized training (-70% parameters for +100% compute) [1] combined with sparsity pruning (-50% size w/ almost no loss in accuracy) [2] and quantization [3] which will lead to significantly smaller models performing well.

[1] https://www.harmdevries.com/post/model-size-vs-compute-overh...

[2] https://arxiv.org/abs/2301.00774

[3] https://openreview.net/forum?id=tcbBPnfwxS

[JohnFen]

What I doubt is that the current approach can lead to AGI at all, regardless of scale. But I'm just speculating along with everyone else. We will see.

[blibble]

as moores law is dead it's hard to see more exponential scaling

they're also not going to find another 2, 4, 8, 16 ... internets worth of content to parasitise

[circuit10]

It’s still exponential, but a little slower. (edit: wait, is that still exponential if it slows down?) Anyway we only need to get to human level (or maybe a bit less) and we’re not that far off (maybe 10 or 20 years at current rates of progress?)

Not all types of AI need external training data, you can train on how effectively a goal is achieved

[blibble]

> maybe 10 or 20 years at current rates of progress?

how can the rate be maintained?

exponential chip scaling is over, and they've parasited, sorry, trained on the entirety of accessible human knowledge

the rate may drop to zero

the exponent may even go negative once LLMs start ingesting their own hallucinations

[visarga]

> they've parasited, sorry, trained on the entirety of accessible human knowledge

I see this as a new development in language, used to be restricted to meat neural nets and books, now it can also be consumed and created by LLMs. A new self replication path was opened for language. Language is an evolutionary system, it's alive. Without Language humans are mere shadows of what they can be. Language turns a baby into a modern adult, and a randomly initialised neural net into chatGPT.

The magic was always in the language, not in the neural network. We should care more about the size and quality of the training dataset than the model. Any model would do, all model tweaks are more or less the same. But the data, that is the origin of all the abilities. But we cannot own abilities, it should be fair game to learn abilities and facts even from copyrighted data. Novel and creative training examples should not be reproduced by LLMs, but mere facts and skills should be general enough not to be owned by anyone.

[circuit10]

The training data thing is a problem mainly for LLMs, so it might be a limitation if we purely scale up LLMs but there are other types of AI around too

Chip scaling still seems to be going pretty fast, and we may discover new ways to make better use of the chips we currently have, like better methods of quantisation, or just using more of them, which could get us just far enough to reach the self improvement threshold

So we could end up hitting a wall with chip scaling or something but I don’t think it’s that likely

[patch_cable]

I watched the video.

> has preferences over world states

I think that part is a leap. I don't think is given that a super intelligent AI will "want" things.

> presumably a machine could be much more selfish

This feels like we're projecting aspects of humanity that evolution specifically selected for in our species with something that is coming about though a completely different process.

> It's a mistake to think about it as a person.

I agree, but I feel like that's what these concerns about AI are doing, because that's what people do.

> (The whole stamp collector thing)

It also seems to me there is a huge gap between a super intelligent AI and the ability to have a perfect model of reality along with the ability to evaluate within that model the effect of every possible sequence of packets sent out to the internet.

[circuit10]

> I think that part is a leap. I don't think is given that a super intelligent AI will "want" things.

But if it has no goal then it can’t act rationally or intelligently. Something like an LLM might not appear to “want” anything, but it “wants” to predict the next token correctly which is still a goal (though since it’s only related to its internal state it might be a little safer)

There’s another good video about why this would be the case here if you’re interested: https://youtu.be/8AvIErXFoH8

> This feels like we're projecting aspects of humanity that evolution specifically selected for in our species with something that is coming about though a completely different process.

That’s because evolution is a process that optimises for a goal. The only reason altruism is a thing is because it actually indirectly benefits the goal, which is for our genes to survive and be passed on, and fellow humans tend to share our genes, especially relatives (who we tend to be kinder to). AI training is also a process that optimises for a goal, but unless having humans around helps that goal it wouldn’t display any human empathy. In this case “selfishness” is just efficiency which a training process definitely selects for

> I agree, but I feel like that's what these concerns about AI are doing, because that's what people do.

I feel like they’re doing a pretty good job at modelling AI as a theoretical agent, which does share some similarities with humans because humans are agents, but the main mistake people make is assuming their goals will be similar to humans because human values are somehow a universal truth

> It also seems to me there is a huge gap between a super intelligent AI and the ability to have a perfect model of reality along with the ability to evaluate within that model the effect of every possible sequence of packets sent out to the internet.

That’s very true, it’s an unrealistic thought experiment, but it’s a a good introduction to the concept that something significantly more intelligent than us can be dangerous and pursue a goal with no regard to what we actually wanted

[patch_cable]

> but it’s a a good introduction to the concept that something significantly more intelligent than us can be dangerous and pursue a goal with no regard to what we actually wanted

I think thing significantly less intelligent can do this too. See any computer program that went wrong. I don't think that is a novel idea.

Perhaps it is a lack of imagination on my part, but I can't help but think, in this stamp collector example, someone would just be like "wait why are these machines going crazy printing stamps" and just like turn them off.

I feel like any argument on the dangers of superintelligent AI rests on the belief it can also use that intelligence to manipulate humans to complete any task and/or hack into any computer system.

[dalmo3]

I don't agree evolution optimises for a goal at all. IMO optimising for a goal means you first define a goal, then you work towards it.

Evolution has no goal, it's simply a process determined by chemical reactions. Any goals we attribute to it, e.g. "for our genes to survive and be passed on" are emergent phenomena, a rationalisation after the fact that that is indeed what's been observed.

It's plausible that AI "goals" emerge evolutionarily as well, but for that to happen we first need to create not AGI but Artificial Life, which is a huge leap from today, and I certainly don't understand how that's inevitable.

[circuit10]

Then by that definition AI training has no goal, it's simply a process defined by calculations. But whether you want it call it a goal or not, the fact remains that they look very, very much like goals. "If it looks like a duck, swims like a duck, and quacks like a duck, then it probably is a duck."

> It's plausible that AI "goals" emerge evolutionarily as well

AI training is vaguely similar to evolution, except more efficient and directed

[dalmo3]

> Then by that definition AI training has no goal, it's simply a process defined by calculations.

No, the very definition of training is that there is a goal which to train for. Those calculations were created by humans with goals. For LLMs, the goal is token prediction.

Evolution has no training.

[ChatGTP]

What is the training goal of ChatGPT ?