Is whom trains the AI more important than the AI?

[lightbulb27]

What I've learned is that AI is taught. So, do you think it should be required that people known whom trained the AI and how?

For example... what if someone like Adolph Hitler, trained an AI engine? What if the leader you can't stand... Trump... Biden... blah blah... trained an AI? It will give answers you can't stand.

Imagine Trump takes over as President, has an AI programmed under decree. Ya think that's gonna be similar to the Biden AI system? The truths will differ.

This seems a sizeable risk. Shouldn't there be a truth in disclosure about AI?

That said, it appears AI is in cases, training it'self. That's scary enough... but needs either limited or declared.

For some background:

Yes, should be full disclosure

No, that's private

I don't know enough

[GuyAnswersGirls123]

I went with "yes" just like COI disclosures in respected papers, although the ultimate goal as I see it is to weed out the humans required to supervise the training set.

This is just a temporary and practical step as I see it which comes with a lot of cons like the bias of the human supervisors pervading into the training.

Fully autonomous machine learning is the goal as I see it which we might, in simplistic terms, relate to a being that requires no parents or teachers whatsoever: an entirely free and independent thinker.

Human supervisors are clumsy and interfere. They're only a temporary requirement as I see it in this early infancy stage of things like NLP and text-to-image synthesis.

[GuyAnswersGirls123]

As a simple example of why human supervision and supervised machine learning in general is clumsy and problematic, consider a simple case of image recognition. A machine is taught through human supervision to recognize images of animals in photographs: "this is a picture of a dog", "this is a picture of a cat," and so forth.

Meanwhile, let's say the human supervisor makes a common human error which is to mistake a picture of a tortoise with that a turtle. Now the neural network will be incorrectly trained to contain this human error.

That's a simplistic example and can be mitigated by more human supervisors verifying each other and so forth. Yet things get substantially more complex and human errors tend to leak into the training.

So we want to eliminate human supervisors in this way since they're an unreliable source of errors.

[GuyAnswersGirls123]

I know one person on this site who actually works in NLP research and she might roll her eyes at how grossly I oversimplified things above. I hope she'll forgive me for it. I like to rely on a lot of simple analogies.

[IslaTheWitch]

those are different things...

for now, so-called AI is built by specific algorithms

so:
1. someone needed to write those algorithms
2. someone needed to write the whole environment for learning (neural networks, more algorithms, prepare data for feeding)
3. someone needed to prepare the whole infrastructure and believe me... teaching AI eats resources... so not only skilled people are needed but also a lot of money for physical machines (this kind of "work" isn't good for cloud like AWS or whatever)
4. mods to work with specific instances, check and correct answers of an AI to let it learn...

and a lot of other people...

who is most important? One person would do NOTHING, it's a team work, including an AI.

[lightbulb27]

I'd like to know more how this really works. One person and even a group of people can be flawed and biased. to me, this is going to wrap right back to God (love) eventually, but I'll stop there.
to my question... still... whom is teaching it... is highly important, in essence, the parents. not sure whom gave "birth to it" is so important, but maybe their reasons...

[IslaTheWitch]

there is no one person. Most data come from internet. It's selected but still... humanity teaches it

[GuyAnswersGirls123]

A lot of where I was thinking and discussing is in the futuristic context of fully autonomous machine learning, completely unsupervised. I actually suspect that might not be too far away with further advancements to NLP with people like you contributing to it.

All we need to make human programmers obsolete is for an AI to be able to understand design requests and generate efficient code better than a human being can as I see it. At that point the AI should be able to bootstrap a superior version of itself, and quickly after, start iteratively converging towards eliminating the need for human supervision of training sets.

Please correct me if I'm mistaken or you disagree with that view.

[GuyAnswersGirls123]

I'm actually kind of excited by the idea of being rendered obsolete when design requests in natural human languages could be executed by machines. I'd miss my source of livelihood if it occurs in my lifetime, but simultaneously so excited by the potential.

[IslaTheWitch]

@GuyAnswersGirls123 "All we need to make human programmers obsolete is for an AI to be able to understand design requests and generate efficient code better than a human being can as I see it." There is much more... understanding not clearly presented requirements is one thing. The other is general knowledge and EXPERIENCE. During our work we make mistakes and learn not only more efficient but also clearer and more readable ways of implementing stuff. An AI doesn't do this, it's pure logic, so readability is not important. But it won't work for us. Another thing is intuition. We use it all the time. AI doesn't feel, so intuition won't help it to find the best solution. And pure logic is not the best solution in most cases.

My IDE already has AI support... and it changes totally nothing :D Still a long way for an AI to be able to take my place :D

[GuyAnswersGirls123]

For sure; maybe I need to clarify "near-future" as something I could see as being possibly within my lifetime, not like 5 years from now but maybe 50 years if I live that long.

The case I'm thinking would make concepts like code readability and maintainability and even software engineering practices in general no longer an issue since the AI can just generate machine code from design specs and following conversations in a natural language like English. We no longer deal with maintaining source code. So even concepts like SOLID become obsolete, the issues associated with imperative programming and side effects are irrelevant, etc, since they're all aiding humans at the end of the day and the humans in this case no longer deal with source code.

A lot of the intuition part I think could be transferred to the designer communicating in a natural language as I see it. Like the way I'm visualizing it, the burdens will be heavier for designers.

Right now as a programmer in my case as an example, I and the rest of the engineers actually have enormous control over the user-end design of our product in ways that always creates some tension with our user-end designer. Yet he accepts it and compromises because it's too time-consuming in collaboration if he micro-managed every single possible programming design decision that impacts user-end design in more ways than just runtime performance. So due to the snail-paced nature of the collaborative process, he's forced to give us engineers a great deal more user-end design control and leave us to independently make a lot of those decisions ourselves.

But if he could just ask an AI to implement his design spec and the AI generates executable binaries immediately after, then he could afford to iteratively go back-and-forth with many, many iterations per day until the end result it as he conceived.

[GuyAnswersGirls123]

One thing I'm curious about with ChatGPT and other implementations of NLP. It seems like conversations with it don't contribute to its persistent training set. Correcting a mistake it made or a sub-optimal answer does not affect future sessions whatsoever from what I can tell, even if multiple users point out the same error. Is that a technical limitation or more something imposed deliberately for ethical concerns?

[IslaTheWitch]

@GuyAnswersGirls123 about chatGPT, it learns in batches, not from session to session. It wouldn't be optimal

The problem is, I don't think that any time soon, people will be ready to TRUST an AI and let it write code without human supervision. So what you described, for now, is not even a dream...

[GuyAnswersGirls123]

Makes sense. I was wondering how much of it was a technical restriction vs. what I'm bundling into an ethical concern (trust included).

It's actually kind of a dream of mine, or used to be. I got into programming mainly out of a passionate desire to design games, and the programming dominated to the point where it became my career.

Yet if I could have hired a programmer to implement my decisions for me and communicate back and forth with him/her and iteratively arrive on the game, I would have much preferred that back then. If the programmer could whip up a design change or clarification in a matter of seconds as what I'm thinking AI might be able to feasibly do in some future, that would have been so idealistic.

[GuyAnswersGirls123]

"Can you make Mario jump a little bit higher for his maximum jump height and make him descent a little slower? Also I want you to change the animation timing like this [show a GIF, e. g.]" Like that as a simple example and work iteratively back and forth to produce the game like this.

[IslaTheWitch]

@GuyAnswersGirls123 lmao... it's what I do daily... but for now, most of its ideas are far from efficient or perfect. And I'm a perfectionist... so my AI instance suffers a lot working with me :D

[GuyAnswersGirls123]

That's something I've noticed a lot with code requests so far in ChatGPT. For common algorithms and data structures taught widely in CompSci, its implementations range from passable to decent, but nothing exceptional as far as efficiency or maintainability. Yet for the more esoteric algorithms and data structures I deal with a lot in computer graphics, like an implementation of this paper:
inria.hal.science/hal-03656945/file/G1ACC.pdf

The source code is often completely incorrect, like it didn't even understand what I requested despite multiple iterations trying to correct it to the point where I'm practically writing the source code for it.

[IslaTheWitch]

@GuyAnswersGirls123 for now, chatGPT only reproduces existing implementations, it doesn't create its own implementations. It's just a Google on steroids...

[GuyAnswersGirls123]

That matches up a lot with the behavior I'm seeing. It does seem to be capable of making corrections based on my requests, but there can be a whole lot of hand-holding required in some cases to converge to an answer that's correct let alone a decent one.

I think I'm probably overestimating its potential to some degree but it often seems to me like it's tickling an ultimate dream of the highest-level programming, from binary to ASM to C to C++ to C# to Python to... English! I'm looking at natural languages like this to replace computer languages. At least that's my dream, possible pipe dream, of NLP in this context.

[IslaTheWitch]

@GuyAnswersGirls123 let's wait and see if this dream will come true :D

[GuyAnswersGirls123]

Yay! 🤝Looking forward to seeing what happens. I'm very excited about it all.

[genericname85]

not neccessarily. cause you can influence what an ai can do after it is already trained. why do you think bing ai and chat gpt have a strong leftist bias? it's not because of the training. it's because their companies have an agenda.

[Nik1hil]

Indeed, A. I lacks Morality, Hence it decides on logic. All the Logic of an A. I should align with Human Morals and Ethics.