Q-Star, or Q*
Among the AI buzzwordists it turned out that among ChatGPT's future products there is something called Q-Star, which according to the gossip would be an AGI, that is, a general artificial intelligence. And the problem is that newsagents now have a problem. Not because Q-Star exists, but because so far they have treated language models as if they were AGIs.
DISCLAIMER: the photo at the beginning of the post is the description that a Stable Diffusion XL – run on my home PC – built using as prompts all the descriptions of "Uriel Fanelli" taken from my book " Altri Robot ", by which he is the protagonist. And definitely, it's one of the closest to what I had in mind while writing.
As I've written dozens of times, ChatGPT is a very large language model. Consequently, you can only use it for problems related to language (rereading, changing style, correction, translation), because it is a non-general AI, that is, an AI that does only one thing.
Consequently, the journalists and pseudo-experts who came out saying "Muh ChatGPT can't count", or "Muh ChatGPT can't program well", are about to make us look like shit, because they tried to use a dishwasher to iron their clothes , and they believe they won the Nobel for discovering that it's not possible. Well done, Sherlock. Other brilliant discoveries?
And in the end, all AIs so far have this characteristic, that is, they do only one thing. It is therefore not a question of AGI. Of course, if "language" is only one thing, it seems very relevant because language is the tool we use to exchange ideas, so many are led to think that if something speaks then it is exchanging ideas.
But is not so'.
An AGI on the other hand is a different thing. It is supposed to be capable of overcoming general problems, in the sense that it would be a non-specialized AI. He could paint, compose music, write, do math, program, and any problem we can bring up today that we can talk to each other about.
Let's be clear, no one is surprised that it took so little. In general, the problem we have with machines is that we have to learn particular programming languages (and then spend 4 days discussing the naming convention of variables and the git fork strategy, which in my opinion describes natural stupidity ).
The first and greatest problem of building machines that solve general problems is to describe the problems in terms that the machine understands , also called “programming”.
Once a machine that manipulates natural language has been built, the distance from AGI has become extremely reduced. It doesn't surprise me at all that after the VLLM we have already arrived at an AGI. And don't kid yourself, the next step will be even faster and more brutal, because we are at the point where one or more AGIs can build an ASI, or at least help. (an ASI is a general AI with capabilities superior to human ones).
However, now to explain what happens it will be necessary once and for all to explain the difference between an AGI and ChatGPT or DELL-E, which will make the geniuses who continue to "discover" that "AI cannot do this and that" look ridiculous, when they are just proving that a language model can only do the language model.
Nice Shot.
Let's get to the gossip now, because what we get is quite ambiguous.
The first thing is that this AGI has “only reached the level of a 10 year old”. Some say 10, some 7, some eight, and so on.
Honestly, it's an estimate that tells me little. Are you talking about intelligence in general? Since the philosophers in their proverbial productivity have not yet given us a definition, we really don't know how to measure it. Today's IQ tests are based on normalizing the observed population to a standard distribution curve. It means that if the population is entirely made up of mentally ill people, an IQ of 120 indicates a guy who is capable of tying his shoes without falling, something that others cannot. See what science is behind it. (*)
Consequently, we will do as we did with the Turing test: a machine is intelligent if it passes the Turing test and is Turing-Complete. The first thing indicates a double-blind test, the second part is an exact mathematical proof.
I mean, get ready, because I have a feeling if the first AGI is a success, from then on we'll throw away the IQ and use a Q-Star comparison. You will probably have a 0.5 Q-Star smart guy, or 1 Q-Star, or a µQ-Star (Gasparri, whatever), or 100 Q-Star smart people, like Hawkins.
However arbitrary, it will still be better than current IQ measures.
If, however, when we say that this AGI is as intelligent as a 10-year-old child, and we mean ability, well: Mozart was already composing at that age (I said he composed, not "played"), and there are children that if put in front of a flight simulator they can fly a Boeing 777 in a very good way, according to professionals.
So let's be very careful about using these descriptions.
The birth of general artificial intelligences, and their sale, requires the definition of some unit of measurement to describe their capabilities. After all, it's a product, and if you know how much horsepower your car has, you should know how intelligent the product you buy is. Otherwise how do you choose it among many?
It's a big mess, because just as the birth of GPT models demonstrated that traditional academic linguists didn't understand shit (**), it could happen that the birth of parameters to quantify an AGI disproves traditional psychologists. The world is becoming fun, and it will always be more so.
However, the point is that we do not have a unit of measurement to quantify (and sell) the work of an AGI. If we can use the number of tokens for a VLLM like ChatGPT, or the resources consumed by the GPU for stable Diffusion, (Amazon already does this), but in the case of AGI, it is not very clear what the measure of their power is, or their ability, or whatever.
This is a funny thing, because we are simultaneously defining two things, AGI and ASI, starting from the fact that ASI is superhuman: this requires an exact measurement of a “human mind”, and then a comparison with a measurement of ASI. In short, how do we know when it's an ASI?
In theory, ASI should be a more powerful AGI than a human mind. But first and foremost, we don't have very clear, quantifiable opinions about how powerful a human mind is.
Also because the question always comes back: WHICH human mind? Gasparri or Gödel? There is a factor of 1000, at least, of difference. And especially, what ability: we note that we would not put a Boeing 777 in our hands to drive a machine "as powerful as a 10 year old child", although several 10 year old children on the simulator show themselves to be very skilled, and especially we forget that to fly better than a Boeing 777, the nervous system of a mosquito is enough. Which still prohibits Gasparri from driving a Boeing 777 (science explains everything), but does not explain whether an AGI driving a Boeing like a ten year old child in front of Microsoft Simulator is "superhuman" or not.
A small note: as a human species we build machines only and always because they are superhuman. You wouldn't buy a car that goes slower than you can walk (sorry about the Flintstones), a hammer that drives nails worse than you could punch wouldn't make sense, etc. If we build a machine less powerful than the human body, no one will buy it: it must have at least one superhuman feature, otherwise it would be meaningless. This is why, for example, I don't keep a nutcracker at home: like almost all Judoka, I don't need it. The doctor says I ruined my fingers, but that's it: no one buys a machine less powerful than their body.
Even a normal computer is superhuman: try to calculate Pi in your head at the same speed if you can, and you will discover that even a Commodore 64 can beat you big time.
And therefore, the notion of ASI puts us in crisis: if all machines are superhuman at least in one dimension, when does an ASI begin?
This is why I say to be very careful: as soon as we actually arrive at AGI, the step towards ASI will be very short, if not negligible, precisely for the reason that even this initial AGI will already be superhuman.
ChatGPT is also superhuman, so to speak: he writes better than most human beings, he writes better than any journalist who has ever existed (ok, it doesn't take much), which may be opinions, but he is MEASURABLY superhuman in quantitative terms: writing them all the answers he writes, reading all the prompts he reads, would be impossible for any human being.
As a working machine, it is already superhuman. And with that, imagine how easy it will be to distinguish an ASI from an AGI.
What's the only unconvincing thing about my speech? It's simply that this Q*, or Q-Star, is gossip so far.
A gossip born around a story that generates billions of dollars.
So, let's wait to see it.
(*) if you run an IQ test on a population and the answer is not the normal distribution, then you change the test until a normal distribution comes out. So an IQ of 100, if the whole sample attends the Pisa Normale, indicates a very intelligent person, but if we use the whole country as a sample, then the Normale student starts from 120. The same person has 100 or 120, depending depending on the sample used. Patascience, in short.
(**) When Chomsky says he fears ChatGPT will destroy the world, he probably means HIS world. And he is right to fear it, because demonstrating that natural language can be generated with a statistical method demolishes more or less all his work on generative syntax. According to the Chomsky school, it is worth remembering, ChatGPT could not work. Instead it works.