Structural Limitations of AI - Possibly does not function at theoretical bests

[BigMan13]

For your optional reading pleasure:

Was thinking about VR and how that failed. And put a thought experiment that even if the tech was now optimized to just being glasses (or less), no-controllers, perfect mo-cap, you still wouldn't have the promise. You'd just have another internet situation where people think they're communicating perfectly, but information is being dropped, resulting in miscommunication and unnecessary conflicts.

Then I applied it to AI. Same idea of a perfection. For information, say there was a text-based AI ('chatbot') that had perfect information, perfect articulation, never hallucinated, always understood what you meant, and there weren't any particularities (e.g. situations where what is correct depends on context) in reality to account for. You'd just have the ever-present problem of people knowing precisely what to do, but being unable to do them. "I should but...", "My brain says...", etc. Asking the AI on what to do on that would result on the same issue of not knowing how to fix not knowing how to do something, and so on.

AI might also de-value informational economics. If AI can code perfectly, serve as a perfect expert, then the second and third world can match the first, especially if this perfect AI can teach people how to use it. The corporate performances, such as the 8 hours worrying about an email, are also valued for their human component, rather than the act itself. If AI could simulate all those performances for free, then it would lose worth both on no longer being a human performance, and it no longer being unique. I suspect this has already happened, and is contributing to the ongoing corporate (especially in tech) recession.

Then on productivity, a perfect AI currently running robots to produce things, work factories, and create even more robots, again removing all particularities of reality (e.g. non-computable physics changes in environments). Then we do have the promise achieved, but then what? Robots take over? Managers (who would be the first to lose their jobs) manage robots while everybody gets UBI? This line of thinking popular with people seems to be more about them than the technology. A severe depression and desire to end the world, this being more a pop-science fantasy to fulfill that desire.

While AI as informational assists will plausibly stay as technology, we've already known that robotics can't automate industry in the best of conditions since the 2000s. The linear and binary (or cubit) nature of code is the main problem, more of it via 'neural networks' does not fix the limitation.

I also talked with someone very familiar with the market and R&D recently who said something interesting. Even with AI as here-to-stay technology, there may be a dark age of its development after this initial rush. Young people are increasingly technically inept, and a significant amount openly hate AI and (to a lesser extent) broader tech. The person mentioned that companies are struggling to find young people properly invested into AI, and that the few they find are often ones who just used AI to trick companies into thinking they (the hire) knew how to use AI.

 

[XL xQgg?QcQCaTYDMjqoDnYpG]

I agree that you're addressing a real limitation, but I'm seeing you bundle together several different things under the term "AI" and that inevitably results in weaker conclusions
As in, you are covering at least three distinct domains:

• Informational systems (like chatbots, coding assistance, knowledge retrieval)
• Embodied automation (robots operating factories, machines interacting with the physical world)
• Economic and social effects (how labor, expertise, and status are valued in human organizations)

Those domains behave very differently

say there was a text-based AI ('chatbot') that had perfect information, perfect articulation, never hallucinated, always understood what you meant, and there weren't any particularities (e.g. situations where what is correct depends on context) in reality to account for.

That scenario is really about information, and not action.
Even if a system could answer every question correctly with perfectly articulated solutions, that still would not remove the gap between knowing and doing. Human beings have known many correct things for hundreds of years and still fail to execute them. The bottleneck here consists of motivation, coordination, and physical execution, not really informational clarity. As in, the same constraints apply just as well to a chatbot as they do to a book or conventional search engine.

a perfect AI currently running robots to produce things, work factories, and create even more robots

A separate issue, for controlling physical systems in open environments is a completely different problem than predicting text or generating code. In factories and real environments, you've got noisy sensors, wear, friction, unpredictable disturbances, on top of physical constraints. Making these models more sophisticated does not automatically solve those engineering problems.
So when people talk about everything getting taken over by "AI", they usually collapse these two things into one. That is, wrongly imagining that a system that's good at information processing automatically becomes good at acting in the physical world, and that leap is doing most of the work in the argument.

On the economic side, what you said about value isn't too shabby. Consider that a lot of what is rewarded in organizations is not merely the informational output (like writing a mail or producing a report), but reliability, coordination, responsibility, and signaling inside a social structure. If a machine manages to produce the informational artifact for free, then the value simply shifts to other aspects of the process rather than disappearing altogether.

On a related note, a pet peeve of mine is that the term "AI" is used as a blanket label for all of these things, but most current systems in practice are better described as statistical prediction systems. Might sound pedantic, but it actually explains a lot of the things we're seeing. Such as why these things are good at tasks like text completion or pattern recognition while struggling with things that require stable control of complex physical environments.
To me the interesting question is not whether everything will be solved by "AI" in the abstract, but which class of problem is actually being discussed. Like, information prediction, physical control, human organization, they're different domains with different constraints, and improvements in one of those domains does not automatically carry over to the others.

 

[cigint]

I also talked with someone very familiar with the market and R&D recently who said something interesting. Even with AI as here-to-stay technology, there may be a dark age of its development after this initial rush.

This is just a truism, and in fact all breakthrough developments have this characteristic, including past breakthroughs in AI. That's right, this isn't the "initial rush" in AI. It has had many similar past periods of fast development followed by speculation, then failure to live up to speculation and return to special-grant level development (what you call a dark age).

 

[peachlover669]

i think the biggest structural limit to ai that will eventually lead to its downfall is its inability to be controlled, even on the most basic level of an ai writing text it is not possible to accurately control what it actually gives you, anything created by an ai will always have a tinge of ai to it since its not a representation of a humans effort it is representative of a humans input and then a computers computing. the biggest example of this would be that of its image and video creation, unlike how a director can physically direct a scene to his ideal an ai can only guess what you want as there is no way for an ai to create from your mind unlike how a human can create. the given examples of video and pictures may seem small in the grand scheme of things but i think this concept of it not being able to truely create what the human intends will end up ruining it at any high level as with more inputs this issue is compounded and ultimately for anything with need for a fine amount of detail it will fail. of course companys do not care about this and most likely it will just mean that the quality of everything goes down as the adoption of ai increases but i believe eventually this will be realized and humans will gain back their jobs with ai rightfully knowing its place as an assistant, i see this happening with coding the soonest since if actual important code is replaced by ai such as that of operating systems and important infrastructure software then it will be noticed very quickly and the companys will either fall or realize their mistake and return to human coding.

also i dont really think vr failed it just was a pointless endeavor to begin with, it never reached its peak but was perfectly adequate for the things it sought out to do. it only "failed" because it just wasnt neccesary as we already have figured out the best way of interfacing with technology through screens and basic computing peripherals which i assume is what your getting at with miscommunications and unnececary conflicts.