Recently, OpenAI claimed they built AGI... again, but what AGI is?
Let's quickly figure out how the internet works, what current AI is and what AGI must be.
Binary architecture that we currently have doesn't support random or non-specific way of declaring data - it's either 0 or 1.
Modern Random comes from cosmic rays, micro CPU temp fluctuations or Lava lamps - it's impossible to simulate all atoms to understand how it formed.
Nowadays programs work all the same:
trigger -> re-action
input -> output
request -> response
There is no way to get data without giving data, processes can't start without a trigger, RAM must be populated by a program.
Current hardware (even quantum computers) can't compute undefined output since... computers are made to do precise work. If program breaks our expectations, we call it "hallucinations" or "bug".
HTTP Standard defines this explicitly: you send stateless request, THEN, you get stateless response. A request doesn't "know" anything about another request mid-flight.
However, client or server may store data between requests.
Current AI uses HTTP and it doesn't store your tokens, you must send all the history and append next message in every request. It has caching, but it's not training a model, but skips tokens that were just handled.
In programming it's called "memoization", a pure function where same input always produces same output, which is precisely what Determinism is.
So by the fact that AI companies offer caching and use HTTP, it exposes that their systems are not thinking but processing data. That is, the precise work. "hallucinations" comes from predicting next best suitable word (not most correct sentence). And offering caching proves models can't reconsider or self-correct as their computations turn into stone.
Which is fancy autocomplete based on human language.
Actually, AI is already AGI, but let's say it's not.
Let's not go into abstractions about AGI, but let's discuss how it could work technically.
IMHO, it should be at least:
Theoretically speaking, to get AGI access, your computer (server) would be exposed to the AGI, so it minds your existence. And possibly could reach out to you without you making a request from your side.
You're not sending requests and waiting for same amount of responses anymore, you're sending some requests and might get some responses. Because AGI would actually need to be going as a non-stop service like a brain.
As a way of optimization, it would suspend certain processes (e.g. areas of brain) to consolidate information, I would call it a "sleep".
You would not send all your history anymore. All your requests are remembered and stored on the server since transferring a whole neural network through ... network and then getting a copy back doesn't look cheapest.
You would be given access to a dedicated service that operates with you in real time (let's via websockets), not access to 10 seconds worth of processing power of Nvidia GPUs with pre-installed 700G model.
I would say 24/7 access to something like NNUs (neural network unit, not binary-only processing) with model that can grow in size indefinitely.
Words like "bye", "hello" might be completely ignored because there is no reason to spend time on answering you that as its AGI, basically task-dedicated person - a corporate slave.
If it doesn't listen to you carefully enough, you can threaten to ~~cut off salary~~ be unplugged from the socket.
For all that, we would need a hardware that can grow in size on its own (it's not impossible, there is already some progress in that), that runs 24/7, but degrades sometimes to optimize itself. And it must be taught a lot of information, which would take years as it's a physical neural network, not binary math. Maybe eventually creating 1 AGI that is better at certain tasks than humans.
Doable, expensive, will find its niche probably.
This all raises a question: why we need all of these complexities if we can just have sex?