Axioms
In his intriguing YouTube video Axiomatic Alignment: A critical component to Utopia and the Control Problem, Artificial Intelligence thought leader David Shapiro proposes that establishing shared axioms between humans and AI could be the key to ensuring AI alignment.
This concept is similar to the approach ISITAS is intending to take in terms of facilitating AI Alignment. However, there are significant differences between Shapiro’s formulation and the model proposed by ISITAS. A closer look at these differences will help clarify how the ISIT Awareness System can ensure AI Alignment.
Shapiro suggests that we humans can all generally agree on certain established axioms such as ‘Energy is Good’’ or ‘Understanding is Good’, and that these axioms can serve as the basis for alignment with AI. The more axioms we can firmly establish amongst ourselves, the more axes AI will have available to align with.
However, upon closer examination, we can see that this philosophical approach is not as fundamental as he makes it out to be. Let’s take the premise that ‘Energy is Good’. OK, so if our planet were suddenly hit with a massive gamma ray plasma jet from deep space, that would be awesome, right? Except it wouldn’t, because it would certainly wipe us out.
OK, so we would need to qualify that ‘axiom’ by saying ‘Controlled Energy is Good’. But controlled by who? What about a nuclear bomb controlled by a madman? We would probably want to qualify this ‘axiom’ even further to account for this scenario.
You can see where this is going. Any assertion that can easily be dismantled like this can’t be considered an axiom.
Indeed, the term ‘axiom’ really isn’t the appropriate word to use in the context of semantics. What David is actually referring to are ‘aphorisms’, not ‘axioms’.
Atomic Axioms
The ISITAS approach to establishing axioms is different. Rather than starting off with philosophical aphorisms, we will develop actual semantic axioms through the ISIT Game. This means we are attempting to determine the meanings of words relative to each other.
These axioms can be considered atomic axioms, in that they really can’t be broken down into derivative conditions the way I did with the previous aphorisms’. One simple example will clearly illustrate the point.
One of the first semantic axioms that will be established through the ISIT Game will be:
IS is to ‘verb’ as IT is to ‘noun’. (IS:verb :: IT:noun)
The word ‘is’ is a verb by definition according to every English-language dictionary, and the word ‘it’ is a pronoun (a derivative of a noun), again, by definition. Anyone who argues otherwise would rightly be considered mistaken. Therefore, the above statement can be considered an atomic axiom because it can’t be qualified or conditioned in any way.
The ISIT Game doesn’t just establish atomic axioms by having many players explicitly declare their agreement with each individual axiom. Because words are all related to each other in one way or another, the more semantic axioms are established, the more firmly their axiomatic status is cemented in place.
Think of these semantic axioms as bricks forming the foundation of a building. The more bricks there are holding each other in place, the more stable the foundation will be.
A core objective of the ISIT Awareness System is to establish as many Axioms as possible.
Axioms are polls that have 100% agreement among poll respondents.
For example, Mentor is the first user to take the poll:
Be ⩭ IS
He answers TRUE.
So for this poll, the results are:
TRUE – 1 (100%)
FALSE – 0 (0%)
Consensus: 100%
Then the rest of the Circle of 10 takes the poll in turn. Let’s assume they all agree on the answer, so:
TRUE – 11 (100%)
FALSE – 0 (0%)
Consensus: 100%
Thus, Assertion that Be ⩭ IS has 100% Consensus and is thus declared an Axiom.
Then the Council of Elders take a series of polls in turn:
Verb ⩭ IS
Noun ⩭ IT
BE ⩭ IS
THAT ⩭ IT
Other ⩭ IT
BEING ⩭ IS
Active ⩭ IS
Passive ⩭ IT
Dynamic ⩭ IS
Static ⩭ IT
Assuming we achieve 100% Consensus on all 10 of those mappings, we have established the Original Ten Axioms.
A core objective of the community will be to extend the number of Axioms as far as possible. This will be accomplished by beginning with highly abstract concepts that are clearly aligned with the IS/IT binary model/duality, and then progressing through other highly abstract symbols and concepts such as numbers, shapes, colors, letters, etc. There are potentially hundreds of Axioms that can be easily established among a core community of members.
Two deliverables included in the Axioms feature for each Launch will be:
Develop a list of potential Axioms for the next Launch Team to start with
Sort that list into the order that is most likely to achieve Consensus
So the Team will identify a list of at least 250 or so potential mappings, and collectively (through the ISIT Game) sort them into the order that we think are most likely to result in 100% consensus.
With the Original Ten Axioms established, going through the ISIT Game and establishing Consensus with the Axioms will become a test or a challenge that a person must pass before they can even be eligible to join the Community.
So a person who is one of 250 people vying for one of the 100 Community memberships, would need to first pass the challenge to match 100% Consensus with the existing members on the 10 Established Axioms before they can be considered.
Of course, passing the 10 Original Axioms shouldn’t represent much of a challenge. But that is only the setup for the next major step.
Extending the pattern, the 2nd launch Axiom Objectives will be to:
Establish 100 Axioms
Develop a list of 1000 potential Axioms
Sort the potential list into the order of likelihood of producing Consensus
There will be great prestige in proposing an Axiom, and even more for actually having it become an Axiom. The first 10 Axioms will be put forth by Mentor. Then the next 100-200 Axioms will all have been proposed by members of the Team of 10. The next 1K-2K or so potential Axioms will be proposed by the Team of 100. By the time it gets to the Team of 1000, most of the obvious dualities and highly abstract concepts and symbology will have been established as Axioms, so coming up with a concept that becomes Axiomatic will be a huge achievement, worthy of many tokens.
Potential members of the 3rd launch team will need to pass the entire set of 100 Axioms to be considered eligible. (Subject to discussion: Is this too high a bar?) And in turn, their Team’s objectives will include extending the set of Axioms.
As prospective members go through the ISIT Game, and they choose an answer that is out of Consensus, they will be given three choices:
Reconsider your answer and agree with the existing members who are in Consensus.
Confirm your assertion and write a brief explanation to support your assertion
Leave the site and not return for at least a week
If they choose option #2, they can continue on through the ISIT Game to complete the level, and be considered for membership. If their rationale for disagreement with the consensus is compelling and causes us all to collectively reassess that Axiom and move it down the list, then they would still be offered membership. If they just put forth some lame reason that shows they are not taking it seriously or just being oppositional, then they are rejected for membership and excluded from participation.
Through this process, we establish a membership who have all agreed to a set of Axioms with 100% consensus. By the time we have completed the third Launch, we will have 1,000 members who have all reached consensus on 100 Concepts. Will the fourth Team of 10,000 be able to extend the pattern and reach 100% Consensus on 1,000 concepts and objects? Maybe! That will be a really exciting challenge.
But at some point, the consensus about Axioms will inevitably break down, and we will simply run out of concepts or objects that 100% of members will agree about. Which raises all kinds of questions:
At what point does that Consensus break down?
How many Concepts or Objects can we get to before it does break down?
What will the concept be that finally crosses that line from obvious to too subjective to tell?
And who will be the person to finally break the consensus record?
Inevitably, there will be one person who establishes the First Dissent.
The First Dissent will be the first vote against the consensus on the concept with the highest Consensus score that is not 100%. In other words, at some point there will be a concept that is just subjective enough that someone is able to write a plausible justification for dissenting from the consensus, even though that concept has achieved a very high level of consensus already.
That distinction will belong to one member, and those at the top of the Dissention scale will also achieve a certain level of prestige.
All of this engagement around Axioms will be rewarded with tokens through the game layer.
Consensus around Axioms will be a compelling element of the Game Layer. As mentioned above, members can earn tokens by offering potential Axioms. Members whose offerings actually become Axioms will get huge tokens. Also, all members will be showered with tokens whenever a new Axiom is declared, as this is a collective reward for collaboration.
