I am building an AI with super-human intelligence
As a software developer, I am working on a novel AI concept that may lead to super-human intelligence. The idea is that instead of a single highway of information flow, there is a whole traffic square where information can go all kinds of directions before it generates output.
The information flow on these highways represent an abstraction of reality in some form. That means, it is a kind of language, although for us it is just noise. I will call this language Babelspeak. Based on Babelspeak, the AI can perform logical reasoning, store and retrieve memories. Much like our mind does, but in a language that is faster and more dense than any human language ever can be.
Now think of such a new AI running on your phone. It continually registers its environment and acts on it when needed. It may warn you: "Hey, don't forget your keys!". You can tell it: "Please don't do that when my mom is there. She is a bit afraid of you." Is this where we are going? Let's see.
With Babelspeak the AI builds its own conceptual reality. I am not talking about consciousness, we can leave that out of the discussion for now I guess. What I mean is, the AI has completely different inputs than us humans. It will have access to the sensors in the phone, including the camera. It will have immediate access to the internet. It will operate at much higher speed than our brains. But it will lack a lot of the senses we humans have. No emotions, I believe. Still, it will be able to integrate this into a representation that, hopefully, will create useful output.
The philosophical question I am struggling with is this: I believe the conceptual reality of this AI will be completely different from ours. Is there something we can say about it? Maybe it will be closer to fundamental reality? What do you think?
I wrote a little introduction to the idea of fundamental reality versus conceptual reality in my post yesterday, you may want to read it.
The information flow on these highways represent an abstraction of reality in some form. That means, it is a kind of language, although for us it is just noise. I will call this language Babelspeak. Based on Babelspeak, the AI can perform logical reasoning, store and retrieve memories. Much like our mind does, but in a language that is faster and more dense than any human language ever can be.
Now think of such a new AI running on your phone. It continually registers its environment and acts on it when needed. It may warn you: "Hey, don't forget your keys!". You can tell it: "Please don't do that when my mom is there. She is a bit afraid of you." Is this where we are going? Let's see.
With Babelspeak the AI builds its own conceptual reality. I am not talking about consciousness, we can leave that out of the discussion for now I guess. What I mean is, the AI has completely different inputs than us humans. It will have access to the sensors in the phone, including the camera. It will have immediate access to the internet. It will operate at much higher speed than our brains. But it will lack a lot of the senses we humans have. No emotions, I believe. Still, it will be able to integrate this into a representation that, hopefully, will create useful output.
The philosophical question I am struggling with is this: I believe the conceptual reality of this AI will be completely different from ours. Is there something we can say about it? Maybe it will be closer to fundamental reality? What do you think?
I wrote a little introduction to the idea of fundamental reality versus conceptual reality in my post yesterday, you may want to read it.
Comments (20)
If you are an actual AI programmer it will certainly be interest to see what you have to say in that department.
Could you please contribute to the topic instead of just venting your opinions about what you think I think of my ideas? If you are interested in more, search online for Babelspeak. Currently there is not much but I am working on it.
We don't know how humans think. Does your AI have the ability to think?
Quoting Carlo Roosen
How can AI have a concept of reality? If you can answer that then this might make more sense.
https://www.lesswrong.com/posts/yzGDwpRBx6TEcdeA5/a-chess-gpt-linear-emergent-world-representation
You are absolutely right, we don't know how we think. We have two perspectives on the matter, one is our own experience of thinking, the other is our neurological knowledge. However, we do not understand the relation between those two. We call it "emergent" but that is another word for "too complex to understand".
Yet something crazy happens at the moment. Neural nets were proposed around 1960, if I am correct. By tinkering, 60 years later, we have ChatGPT. And even while we still don't understand how it works, there are similarities between ChatGPT and our brain processes. That is kind of magic.
What I plan to do is based on observing my own thought processes. I will try to build the architecture that will enable the processes that I am aware of. Language is a large part of that idea. So yes, I believe one day my AI has the ability to think. Hoping for another piece of magic.
This might raise the question of consciousness. But in order not to complicate the discussion, I would like to save that topic for a later post.
For me a proof of thinking would be, you give it a complicated question 'A' and it has no answer. Then some other day you give it a piece of seemingly unrelated information 'B'. Then a day after that, it says "He, do you remember you asked about 'A'? Now yesterday you talked about 'B', and maybe there is an answer here." Then it explains you in all detail a perfect answer.
That is what I am trying to achieve with Babelspeak, a method to store, retrieve and combine internalized abstractions.
Maybe my above answer to MoK answers your question too?
Either way, I do not see how this reveals the questions that came after it. If you can explain.
I can run with the hypothetical that "AI" will just become "I". The question about how this forms will be shaped by input/output. If in 'cyberspace' then there could be something we may call akin to 'consciousness' but certainly nothing like human consciousness ... it may even be faulty to call it 'conscious' at all in anyway we appreciate the term.
So, no. This will not be 'closer to reality' as reality is for us what it is just as it is what it is for other conscious sentient creatures. A shared world makes the reality shared. If AI can map reality it is mapping reality not knowing it.
I have a feeling you mean more than this so you'll likely have to explain further what your point is?
You specifically asked "How can AI have a concept of reality?"
I have shown pretty solid evidence that AI internally models the stuff it interacts with (and thus, if it interacts with reality, it will - or at least in principle can - have an internal model of that reality). To me, it seems fairly self-evident how that's relevant to the question you asked. An internal model of reality, to me, seems like what someone might mean when they say 'a concept of reality'.
Quoting Carlo Roosen
?
But there's really no telling until it actually starts being built and learning things. We already know that neural nets have their own ideas about things that are similar to ours in some ways but different in some surprising ways. I expect that to be the same with this guys project, if this guys project even ever gets working.
Have you seen Google's new AI-summarized search results? It takes forever. My browser delays for five or ten seconds before putting up a result. It's annoying, and the results aren't any better than what I'd get if I just clicked the top link anyway. Tonight I figured out how to turn it off (someone published a Chrome extension for that purpose) and now my web searches are back to the way they were before, serving up search results instantly. I noticed the improvement right away and I have no desire to switch back. They're like everything else AI does ... even when it's correct and useful, it's like swimming in lukewarm oatmeal. There's just something off about it.
I have a feeling the day after you release your AI, someone will release an extension to turn it off. People find this stuff annoying.
Do you often forget your keys when you leave the house? Doesn't the act of locking the door behind you serve as a reminder? Is that the best use case you can think of to motivate my interest? If you were presenting this to an investor, what would be your elevator pitch? As it stands, you haven't piqued my interest. You say, "... the AI has completely different inputs than us humans," but you gave no examples. I already have access to the Internet, and my phone's sensors don't sense anything my own senses already do. Unless you give it LIDAR like an autonomous vehicle, or an infrared sensor, or a built-in CO2 detector. It's straightforward to hook up all kinds of lab sensors to smartphones. That doesn't strike me as much of a breakthrough.
It could be that you have something more impressive than you described. But remember, you get 30 seconds for an elevator pitch. You didn't motivate me.
tl;dr: I see nothing revolutionary about hooking up lab sensors to an AI. I'm sure people are already doing it. Autonomous vehicles sound like the closest thing to what you're describing. They're equipped with sensors humans don't have.
At the moment, of course, autonomous vehicles aren't autonomous. Indeed, it's very expensive to fund the support staff. As a prof commented in this NY Times article:
How would you tell if one conceptual reality is different from another in the first place? You seem to allow it to admit of degrees ("completely"), so similar conceptual realities ought be able to be distinguished.
Interesting. Could you elaborate on this a bit more please? What differentiates your idea from existing systems that take into account a wealth of sensory information constantly such as self-driving vehicle systems or current AI that already cross checks "multiple sources" (a whole traffic square) before converging into a single "highway" (end user output)?
The average cell phone battery has a capacity of 10 Watt-hours. A single query to ChatGPT uses 2.9 Watt-hours.
How soon do you expect the hardware technology, needed to support your project, to be available? And why do you think that expectation is realistic?
:rofl: