0
$\begingroup$

I have no specific knowledge of the AI field, but I heard that AI systems get better the longer they learn.

So, I was wondering: could it be possible that AIs will learn how to create better AIs (Or assists humans to create a better AI), and then those better AIs will learn to create an even better/faster AI, and so on? Wouldn't this mean that the AI would get exponentially better/faster because, after each successive generation, a slightly faster AI will do the job?

I also heard that "Google is using AI to design processors that run AI more efficiently". Wouldn't this be the same? AI designs faster CPU => AI get's faster and can design an even better CPU to run on.

Is something like that possible? Would this mean that at some point there will be a breakthrough in AI that will significantly increase the speed of AIs because of those loops?

$\endgroup$
3
  • 1
    $\begingroup$ No. Your premise that AI systems get better over time (without any other input) is not accurate. $\endgroup$ Sep 7 at 14:37
  • $\begingroup$ The learning process isn't well addressed in this question, but from my understanding the premise is only that an AI (or some AI) can learn (without caring how/from what data it learns). So the premise is indeed flawed in the sense that an AI can't learn indefinitely (there is a limit to how much it can learn, and the auteur should understand it) but, IMHO, it isn't enough to totally discard the question as it could (at least theoretically) learn enough to create a "better" version of itself. $\endgroup$
    – kirua
    Sep 7 at 15:14
  • $\begingroup$ Just want to say that I think your question is justified and I believe it's possible in the future. If we could create AGI, and if it has more combined info than all organizations or any single person in the future, it could theoretically have the ingredients to construct something novel by synthesis of an enormous amount of disparate phenomena and ideas that no AI researcher can on their own. No AI researcher has studied every field under the sun. If you have an AGI that had done that, it could pull inspiration and remix like nobody's business. $\endgroup$
    – Hmm
    Sep 7 at 17:45
0
$\begingroup$

The first thing to consider is that (as of today) there is no general AI (AGI) and most if not all research don't try to create an AGI. An AGI can be described as an AI able to perform well (at least at a human-like level) on a wide range of tasks. If you are interested in the question I recommand you to do some research about the term AGI as there is no consensual definition, but I believe that the one I gave you will be enough here.

This means that an AI can actually do just one (or a few) task and so, if we created an AI creating new AIs it would just create new AIs better to either do something else (so there is no loop of improvement and it's not what you are interested in) or better to create a better version of themselves which in itself is useless as this better version can't do anything useful to us (except maybe to show that we can do it).
I don't know if there are examples of research creating an AI improving themselves, but there are examples of AI improving other AI such as this one where an evolutionary algorithm is used to design a better neural network.

Concerning your example about AI improving the hardware, it could indeed lead to the improvement of a wide range of AI, but it's, in my opinion, different than an AI improving themself as the loop take several years before being closed (from this source it take several years to produce a new cpu) where an AI directly training an AI could do so in a matter of days (or weeks or hours depending on the technology, but anyway it's way faster)

But, if we had an AGI able to create better, more or equally general AI, it could indeed lead to a breakthrough in AI*, or this is at least the opinion of some people including Nick Bostrom which wrote the famous book "superintelligence" in which he described how an AGI could very rapidly improve itself and described it as a possible singularity (meaning that we are unable to foresee what this AGI could become and how fast it could improve itself).

*at the condition that a better AGI is able to further improve the next version, which is not so obvious (thanks Neil Slater for this remark). Adding this condition it means that your metrics for "better" must include the ability to create a better AGI and that you can effectively always (or for long enough to obtain a breakthrough) improve this ability (As far as I know, we don't know if it's indeed the case)

$\endgroup$
4
  • 3
    $\begingroup$ Another thing worth considering: An AGI that can improve another AGI does not necessarily lead to that second AGI being able to further improve the next generation (as if the "capability to improve" is just some real number that can be manipulated through wishful thinking). The idea of exponential improvements to AGI is based on some very fuzzy understanding of what an AGI is, let alone an "improved AGI" or some kind of measure that could grow generation to generation. $\endgroup$ Sep 7 at 15:37
  • $\begingroup$ Indeed, it's not so obvious that an improved AGI can further improve the next iteration, I added a point about this. As for the idea of exponential improvement, I wrote that it's the opinion of some people, not that there is a consensus about it. As far as I know this is an idea that I would judge as plausible, and I see no argument against it. From what you say this is due to a fuzzy understanding of what an AGI is, could you be more precise on what misunderstanding of an AGI could revoke this idea ? $\endgroup$
    – kirua
    Sep 8 at 11:42
  • $\begingroup$ The misunderstanding is about measuring capabilties. In terms of D&D-like stats, lets call the stats INT (for general intelligence), CraftINT (for ability to create another object with a specific INT). The sigularity starts with an AGI with INT 10 and CraftINT of 11, then goes on to assume that when the INT 11 AGI is built it will itself have CraftINT even higher. However, we have no theory that can measure or predict anything for CraftINT - or anything similar. We could link INT to something like IQ, but what we know so far is that no human, whatever their IQ, has any CraftINT ability. $\endgroup$ Sep 8 at 13:09
  • $\begingroup$ . . . unless you include biological reproduction, in which case human CraftINT is essentially the same as their INT in the short term, and evolution may cause both to drift up or down depending on the environment. Either way, we are presupposing a relationship between INT and CraftINT when we cannot define or describe the latter in anything other than the most general terms. Applying any kind of maths to it is extreme conjecture. $\endgroup$ Sep 8 at 13:10

Not the answer you're looking for? Browse other questions tagged or ask your own question.