Universal Artificial General Intelligence Standard
UAGIS is a non-profit organization that systematically researches and standardizes Universal Artificial General Intelligence (Universal AGI).
In 2022, we established the theory of "induction", which is lacking in modern science.
In 2024, we plan to release sample code for artificial general intelligence free of charge.
To get to AGI, We felt that modern AI was missing something crucial, but it wasn't clear.
Even if we refer to the brain, we unconsciously make up for the parts that are lacking in programming.
For example, when making inductive inferences, programmers use their intuition to determine the prerequisites for what population to collect statistics from.
Therefore, we decided on the criteria for inductive inferences, such as what kind of data is best for statistics.
With that criterion, you don't overfit and you don't need a lot of data to train.
In modern times, there is an image that "intelligence" = "learning", but "thinking" and "learning" are different.
Modern neural networks output "guessed results" when information is input.
Conversely, inductive inference criteria tell us what information is needed to get a good "guess result".
Since the information you want changes depending on what you want to guess, the information learned in advance is not enough.
"Thinking" is searching for the information necessary for what you want to guess now without relying on prior learning.
Criteria for inductive inference tell us the likelihood of inference from a single sample rather than the statistics of multiple samples.
Even if you don't have a lot of past experience, you can explain the "thinking" that allows you to make good decisions from a little information.
The biggest parts missing in artificial general intelligence are now available.
Rather than parts, "induction" is the foundation of science alongside "deduction".
All inference, including the brain, can be explained by "deduction" and "induction."
If you can't investigate the brain completely, you can interpolate with the theory of induction.
Rather, without consulting the brain, you can simply create a program that gives the optimal solution to any problem.
There are no longer any technical barriers, as we no longer have to investigate the brain completely.
An artificial general intelligence sample program is expected to be completed in 2024.
Since it is necessary to agree on what is the optimal solution for inductive inference, we will promote standardization.
In conventional science, credibility is expressed by "probability", and the credibility of "probability" is expressed by "variance".
However, the "variance" is distorted by chance "bias" when the sample size is small.
For example, if all the heads come up at several coin tosses, the variance is 0, so we mistakenly think that we will always come up with heads.
Therefore, we have introduced a new value, "unknown", which represents the degree of unknown.
The quantity and quality of the sample determines the "unknown" proportion.
The quality of the evidence refers to its suitability as evidence, that is, the degree of "bias".
Conventional science and AI assume that given information is correct and do not doubt it.
For a particular application, it's fine if the programmer adjusts it so that the best data is entered.
But "general" doesn't allow programmers to adjust on the fly.
It is necessary to have a mechanism to correctly "suspect" the data to be used.
The completion schedule was earlier than expected, so the human resources and funds were not in time.
Please accept our apologies for the delay in disclosing information, as research is our top priority.
If anyone agrees, please help us.
2022/12/31 (planned): Pseudo code of Level 1 Universal Artificial General Intelligence will be published
2022/10/01: Completion of the theory of induction
2022/09/01: Established, opened website