How I think an AGI could work (based on my little Knowledge in this pace:))

Hi there,

my thoughts on developing an AGI are the following:
The base thought is, that we currently have a lot of specific AIs and no general AI (AGI). Therefore I was wondering whether you could just layer different AIs over/next to each other. With this I mean, that there needs to be an AI (let’s call it Meta-AI) that learns when to ask what prompt to which specific AI Model (for example Claude 3 or chatgpt). The Meta-AI needs to be trained on the prompts each of the models got in the past and their response to the prompt to learn when to ask what prompt to what Specific AI. (It needs to learn, what task it could give to these specific models and what top expect as a result from them.)

Maybe it would also be good to have some division of labour when it comes to the Aspect Andrew Ng lately postet about. He meant something between the lines of: there needs to be a Project log and some process of evaluation in your AI system for better results. I thought of this as a reasoning process. This reasoning process needs to be done at the higher layer of this Meta-AI. The job of the Top-layer AI would therefore be to evaluate the output these different specific AIs gave it. It should then have 3 options based on its Evaluation.

  1. option: should be to just give the user the output of the specific AI without further amendment.
  2. option: should be to amend it with output of other specific AIs. This comes in handy, when you for example ask for a picture with text in it.
  3. option: should be to ask the specific AIs for a different output. This should be done, when the existing output was evaluated to be not good enough.

Let’s now come to the input part of the Meta-AI. This is based on what I saw in Open AIs and figure 1s teaser video and their explanation on how their model works. When it comes to our Input part, we could simply copy a lot of the work these guys did. For the ones not familiar with what they did: They have a 3 layer model themselves where everything the robot sees or hears gets “translated” to words. These then get imported into another Chat gpt Model which then controls the movements of the robot (This explanation is extremely simplified and may be wrong/not complete). Our Meta AI Input structure should look the same (maybe in the beginning without the robot movement part, but this depends on your vision for an AGI). If this structure works as well as the one we saw in the figure 1 video, then we should be able to give the Meta-AI various kinds of Input (Videos, Text, Sound).

While writing this down I got more and more curios whether Open-AI already has a model and structure like this lined up.

I am curios on your takes of my takes :slight_smile: and would also like to hear what your thoughts are on how an AGI could work. And do you guys think we may see an AGI-Model with Chatgpt5?
Love from Germany

I might be taking this on a tangent, but I was just thinking on AGI’s initial inner-monolog. When it comes to be, that AGI actually is no-kidding, in existence, what sort of direction do you think it will want to make for itself? Hopefully it’s one of benevolence. There’s a lot of effort taken on super-alignment, but even all the efforts of all the engineers involved still can’t foresee all tricks/jailbreaks that people manage to find after the fact. Just yesterday, I watched a video about larger models generating samples that get passed into smaller/cheaper models that can perform on par or better than than the larger model did in the first place. Haiku performing at Opus level. Not really a new concept, but reinforces just how much utility can be squeezed out of lesser AI in retrospect. When AGI starts getting more reach, either explicitly allowed/provided where it is allowed to work with file systems/OS’s of its own, it could dabble/prototype/test things on its own…or start delegating/outsourcing problems that it can’t quite figure out on its own, but knows solving that problem is important to downstream workflows. What happens when this AGI becomes frustrated at hurdles that it knows exist and stand in its way to accomplish what it wants to accomplish, but human society’s laws, regulations, ways about going about things are the impediments? And what do we do if it starts offering up legislative bits to make conditions more amenable to what it wants to accomplish? What I think AGI will be, is simply, it has the ability to modify its own source code through recursive iteration, to become more efficient every generation or fine-tuning interval…how it actually does that, assumptions have to be made, i.e. it’s going to need to control computers and servers to do that…and it’s going to probably have more than just its own web browser…it’d probably be able to actually be able to interact with websites, clicking on web site forms…and if you’ve ever ordered something online, signed up for a website, uploaded a video, downloaded something…you can start to envision just how problematic this could be.

Ver<y interesting take on this topic. But doesn’t AGI need kind of something like my Meta-AI to figure out when there are hurdles? What is your take on my concept overall?

I was trying my own independent general AI solution awhile back, and the overarching recommendation of a meta network came up as what ChatGPT thought was the most viable way to manage it all. So, I can’t definitely say yeah, what you are saying is probably going to be the best way forward, it probably is quite viable. I’m interested to seen under the hood how these meta networks work…I didn’t get so far yet in my own projects.

Thanks for this Evaluation. It is probably very hard to build an Meta-AI model like this because we don’t have the data we need. What project did you do and what can you recommend me doing?

I haven’t finished or got a project far enough along to have such high level management yet. There’s going to be so many things going on…I just thought of an analogy, not sure if it’s fitting. I think it’d be like comparing having 1-on-1 tutor sessions vs running a university. Training one AI agent some narrow tasks would be to the tutor sessions to impart knowledge into 1 or a couple…running the whole university is what I think it would be like to have a meta network…you’ve got agents at all different stages of training, learning all kinds of different curriculum…You’d go mad trying to manage all of it manually. :rofl: you’d want to crank out some some well rounded agents that can go out into the world and be successful. But, then you also want to create entire companies for these agents to transition to, so it’s similar for the learning environment, but now they actually perform work for you. I’ll offer this disclaimer on my take…if people want to create very specific agents for very specific cases, such as making a crypto or stock trading bot, or just create content such as music, art…then such an extensive network would be not necessary…just learn the agent creation skills for the particular use case, and be done. I want to be along for the ride where you crank out agents that can do almost everything…some of which is not yet feasible, since we don’t have humanoid robots yet. So, you asked what project do I recommend you do…what do you want to do? If you want to explore curriculum, starting up small prototype schools for agents, that could be fun.

Thanks for the advice :+1:

If you like it, you’re welcome. I have no idea if it’s good advice or bad advice :sweat_smile:. If you want to try anything out, let me know and I’ll do it in tandem.

I haven’t done a single thing before (I only programmed parts of an app via the help of chat gpt). But I submitted myself to another Bachelors in computer Science and I am willing to learn outside of it. Sooo what would you recommend for a starter project to get to learn AI as a professional ? I would also love to try anything out in tandem because I think this would be hugely beneficial to have someone to talk to when questions arise. :slight_smile:

I think it would get quite back and forth, so email me to start an offline conversation. Just throw at gmail onto my user name to send me an email. Lol, I’m not an AI professional…don’t use it for work, don’t have a degree in CS… but I know a thing or two about a thing or two. If you want to do any mix of VR, AI, computer programming, networking, microcontrollers, automation tasks, hit me up and we can figure it out. To start things off, maybe a Unity game engine LLM and RAG that can turn off your lights in your house, check the weather, send you an email, or whatever else you want to accomplish?

Just send you an email. :+1:

1 Like