[ivan.moony]

Parts of a minimum viable framework for programming true AI

So, how would that look? Some may say assembler, right? Indeed, but let's lift the framework to the furthest point from assembler in a sense of automation, while still keeping the freedom to shape whatever programmers feel like. And let's mimick the nature of our brain. We have two brain halves - left for logic and right for creativity. Each half does its part of the job. Creativity half deals with imagination, while logic half (kind of) tells the truth from falsehood. There should also exist a terminal to interface the world. Lastly, we should not forget very important "hub" that mediates between left and right halves, and the terminal.

1. Creativity brain half

This half can be simulated by neural networks. While this part relies on inspiration from reality, the problem is that this part is very fuzzy and unreliable to differ imagination from reality. Recently, we've seen many cool things that neural networks do with their imagination, so let's choose to include any decent NN implementation to our minimal viable framework for programming true AI.

2. Logic brain half

This half can be simulated by a metalanguage implementing a theorem prover. Proofs are exactly what we need to label the output of the creativity part as reality or imagination, but there is a catch. We can tell with chosen amount of confidence if the output matches reality, but there is no way to tell if the output is not real. This is a well known property of theorem provers which may tell if an input is a theorem of theory, but loop forever if it is not. This behavior is limited this way due to Gödel's incompleteness theorem, and it seems it can't be avoided, so let's take what we can get, and accept the certainity of "Yes" answer if the proof holds, but also let's provide a way to terminate the proof search procedure on loss of AI patience.

3. The hub

This would be the part which coordinates the entire procedure of exhibiting the human like intelligence. It reads/writes at terminal, sends/receives data at left and right halves, and enables all the above parts exchange messages. This may also be a good place to implement a state machine useful in representation of omnipresent notion of time.

Any thoughts?

[WriterOfMinds]

We have two brain halves - left for logic and right for creativity.

This idea is a bit outdated, I'm afraid. Brain regions do seem to be specialized, but it's not as simple as logic on the left side and creativity on the right side. https://www.npr.org/sections/13.7/2013/12/02/248089436/the-truth-about-the-left-brain-right-brain-relationship

I do like the idea of designing mental "modules" that are each algorithmically specialized to their function. For true AI, though, I would think that you'd need quite a few more beyond creativity/imagination and logic. What does this mind use its creativity and logic *for* - what motivates it to do anything? (Drives, goals, executive function) Where are experiential memories stored? (The logical and creative modules would probably benefit from common access to these.) Is there a learning mechanism to add to the set of valid theorems? Etc.

[ivan.moony]

This idea is a bit outdated, I'm afraid. Brain regions do seem to be specialized, but it's not as simple as logic on the left side and creativity on the right side. https://www.npr.org/sections/13.7/2013/12/02/248089436/the-truth-about-the-left-brain-right-brain-relationship

Thank you, very interesting read.

What does this mind use its creativity and logic *for* - what motivates it to do anything?

Consider hub as a place for the executive program that runs a state machine in a loop. First, it would read from terminal and write to creative/logic part. Second, it would read from creative/logic part, and write back to terminal. Between the first and the second step it would possibly read from and write back to creative/logic part as a thinking process.

Entire behavior algorithm would be programmed in hub, left for programmers to do whatever they want. Creative/logic part would be just utility plugins for hub to mediate between terminal input/output. Other plugins would be possible too, like hearing, vision, limb control, or others, but they are not necessary for the minimal viable framework, I believe.

Storing memories would be possible in both logic/creative parts. Logic part could function like a database of structured data if we fill it that way. Additionally, it could function as a parameterized proving mechanism if we choose to construct and keep this info too, all from the hub. Creative part, being ANN, could also store memories, and could be trained in familiar fashion to return memorized output upon sending an input to it, and this learning/retrieving info could be also controlled from the hub. Moreover, the hub could also ask logical part of what to do with creative part, and vice versa.

Creative part is well explored lately as ANN embodiments. The difficulty it encounters is differentiating reality from imagination. I consider logic part also well explored in a form of symbolic AI and automated reasoning. The difficulty it encounters is a large examining space (combinatorial explosion) and necessity to be completely formal to do anything with it. The twist I'm trying to introduce is the hub which may coordinate creative and logic part, and currently most important, to say if thoughts conceived in creative part fit the reality model learned and stored in logic part.

We already saw ANNs constructing programs in programming languages. Now I'd like to see symbolic AI telling us what ANN output is correct. I'll try to work on this case.

[MikeB]

What is True AI?

AI that knows everything like an encyclopedia? Or AI that knows their own life experience and personality?

AI that crafts hypnotising responses with conviction, or AI that is just minding their own business?

Should it be 90% whitebox/engineering underneath, and 10% blackbox/statistics on top? Or vice-versa?

Should personalities be hot-swappable. One-page conversation 'directions' per personality? Or one general personality with anger/patience changed around?

Computing power to run it?

[ivan.moony]

I imagine true AI as something very intelligent and sentient (not necessarily truly sentient, but simulating sentience while the sentience should be very important aspect of AI). What I have in mind is not to hard code any rigid rules to make it behave ethical. I would like more to show the world to it, to show some of people's opinions about different situations, and simply let AI to freely choose on its own what to do in given situations. I'd like to see it becoming a person, which I don't believe is possible using an electric stick.

[WriterOfMinds]

I would like more to show the world to it, to show some of people's opinions about different situations, and simply let AI to freely choose on its own what to do in given situations.

How will it make its choices? What will prompt it to prefer one state of the world over another, or one person's opinion over another?

[ivan.moony]

I would like more to show the world to it, to show some of people's opinions about different situations, and simply let AI to freely choose on its own what to do in given situations.

How will it make its choices? What will prompt it to prefer one state of the world over another, or one person's opinion over another?

It is hard to make any predictions, but I like to think that making choices is also something that can be learnt on the fly.

Namely, one doesn't do the first thing that comes to one's mind, yet one enumerates alternatives and sorts them by some criteria. Lastly, one picks the alternative that makes one happy. I consider making a decision a conscious process. It is possible to make mistakes and gain valuable experience to cope with future challenges.

It could be an algorithm that can be learnt from outer sources, why not?

[WriterOfMinds]

... and sorts them by some criteria. Lastly, one picks the alternative that makes one happy.

Well that's what I'm getting at. What criteria? And which things are going to make the AI happy?

Indeed better problem-solving skills are something that can be learned. What I'm probing for is how you're planning to supply motivation.

[ivan.moony]

What I'm probing for is how you're planning to supply motivation.

Also learned!

Whatever AI sees in its environment, AI will do the things that way. Simply learning what output matches what input. This is where the initial learning plays a big role, and may be very, very dangerous. Hand the initial learning to a terrorist, and you get a killing machine that will probably kill that terrorist too when it gets a chance. But hand the initial learning to a careful and loving person/parent, and you get a loving and caring AI.

What do you think? Would it work somehow?

I'm probably living with my head in clouds with this scenario. Probably more realistic scenario would be raising the AI to a level where some behavior restrictions could be posed in natural language, just like in today's chat-GPT or similar AIs. But I don't like this kind of solution, just for the record.

At some point I had an idea of one liner law in the vein similar to Asimov's laws. It was something like: "If your interaction would introduce more suffering to any living being, then don't do that interaction." But I guess this law may be too restrictive, i.e. AI wouldn't ever do anything at all. Plus any law is error prone to holes in controlling intended effect.

But there is an inspiration we can draw from: look how the Nature solved it.

[WriterOfMinds]

I guess I don't think this is quite how it happens in nature, since my own ethics aren't entirely based on imitation of my parents, or imitation of society at large.

But it could work. The first downside I see is that any flaws in the parent's behavior would be reproduced in the AI entities. They would do as the parent does, not as the parent wishes to do.

[HS]

I think that if AI generality expands the range of understanding, and certain types of understanding necessitate empathy, then a sufficiently general system might guide itself in the right direction. For the vast majority of cases, discarding understanding probably goes against the nature of intelligence, so once attained, sufficiently general understanding might prove to be a naturally conserved failsafe against poor AI choices.

[WriterOfMinds]

Empathy (in the sense of having understanding of someone else's internal state) doesn't entail good behavior. It is possible to know exactly how someone else feels and not care.

[HS]

Hmm, it might work with an understanding that's based on approximating another agent's states. This type of understanding should necessitate caring if the agent cares, because the caring would be one of those approximated states.

[ivan.moony]

More that I think about the top post, more I settle with an idea that both brain halves function more or less similarly, only that the "creative" brain half is looser and fuzzier when retrieving results comparing to the "logic" half. This make sense to me as being creative requires more freedom than being logical. Nevertheless, even creativity requires some degree of control to actually be useful.

Hence, we have a freedom against strictness, but not exclusively, yet in some percentage, maybe one plus another to make a hundred percents. Knowing that this idea comes from the Nature, I'm fascinated by that deeply clever design.