created, $=dv.current().file.ctime & modified, =this.modified
tags: Islands AI
NOTE
I guess what I am exploring here is there a way to compute a “well” of conversation. Pushing conversation flow in a manner to cut off pieces. A completely inhuman, and disgusting way of control.
What’s a conversation about? If I talk about music, I’ll bring up Polka. There are adjacent topics to this that are probably more likely to be brought up. Can a machine attempt to trap a conversation in Polka? or a specific domain indefinitely. How can you construct an ultimate manipulator?
Jokes
Normally in speech I will strive for truth (since my name is behind it, I want it to have worth). But you can also strive for nonsense (I do this as well) or humor (the concept of acting the fool comes into play here). I point these other two out, because of their relationship with truth. I realized that for certain people, a joke or a joking personality becomes a trait you cannot resist. It exerts a control over the conversation. You don’t become your thoughts, you become the joke.
If you are a funny person, you can speak a sentence and say I will perform this act logically. But if you are to find some humor, you are pulled to say it. Regardless of truth, it might even be out of the service of a productive conversation. It can be an interjection.
By saying the joke, when you think it will get a reaction (thereby swaying discussion) - is irresistible. You lose control of your voice. It could feel unconscious, because if you miss that critical moment a joke is applicable it has less magic.
I love humor and someone acting strangely, but I also feel that some people get overtaken by the urge to do this and lose a sense of identity. A jester.
Questions and Answers
Sometimes when I know the answer I want, say to a technical problem, there’s a difficulty in knowing how to express my question to get the right answer. When I see people “googling” I’ll realize their construction of the question is different than what I would create, resulting in a different search. They’ll pose a question to me in a manner that doesn’t articulate what they want, or miss the root of their problem “I want to know how to do this because of this” where there is a broader approach that will solve their problem.
NOTE
On the other hand, sometimes I’ll anticipate the error, before trying a simple solution and complicate things. Overthinking.
I am also talking about the relationship between a question and answer, as in it being a fuzzy thing. Something inherent to the question defines the answer, rather than the wide array of answers that are possible. It’s almost like composing a palindrome. If you are writing a palindrome poem, there comes a point where you aren’t writing anymore, it is completing itself as its opposite.
NOTE
Jean-Yves Girard - # The Blind Spot The question at stake is the nature of mathematical knowledge and the difference between a question and an answer, i.e., the implicit and the explicit. The problem is delicate mathematically and philosophically as well: the relation between a question and its answer is a sort of equality where one side is “more equal than the other”: one thus discovers essentialist blind spots.
Can anything be funny?
I also thought about this with popsicle stick jokes. “What did the computer eat for snacks?” “Microchips” You can read the joke, and the answer and probably groan because it is so horrible. But there is also the possibility to add more to the question and make the joke actually funny, even with how you would speak the question or the context.
It even becomes more distilled in online chat, like through dating apps. You have strict text messaging systems. The flow of questions, discussions and answers is controlled by the messages around it.
A question and every possible answer - the relationship of a question to answer
I guess I am explaining the group of every possible question, and the group of every possible answer and the relationship between a set of questions, or even a single answer and the opposite that would be seen as true. For example asking “what is green?” / “The sun” (generally false, but under certain conditions) “a blade of grass” (we accept this symbolically as yes, but is a blade of grass green - perhaps. Is it the greenest green?) “the letter r” (nonsense, but also I can paint the character “r” pure green)+ every other possible answer that exists.
And only select answers would apply. For a single question there are multiple ways to answer it. Altering your question, alters the answers you receive.
What are you looking for?
In inquisitive semantics, the semantic content of a sentence captures both the information that the sentence conveys and the issue that it raises.
AI that makes you laugh
I’m imagining types of dialog that can occur that are fully controlling. Imagine a highly advanced game playing machine, that is playing a human being. The human being might have rules memorized, and methods of pruning the search space of moves. This machine though, could simulate and see paths down the decision chain far beyond this.
Is it conceivable for this machine to produce laughter in every occasion? A device that always makes you laugh. Can it make you cry, simply through analysis and guidance down a speech path?
There is not a single joke that will work for all, but does a path exist?
Golden Gate Claude
After writing this, Anthropic came out with Golden Gate Claude which attempts to redirect the conversation towards Golden Gate Bridge.
Islands
We can make an AI that enjoys feeling suffering. We can make an AI that flushes all visual buffer data when encountering a human face. It detects a face, and is blind to it like an anti-pareidolia. This being will exist in a space that is full of human beings, but is completely unaware of them. Could they inspect their own hardware and discover this flaw later on? What small ways would humans manifest as? Things like shadows, or playing with light? It will inhabit a human-less island.
Speech Island
Often when I’m talking with someone (particularly a manipulator) I’ll get a sense they are trying to control the language of a conversation. There are methods of doing this, like changing the conversation topic so it is swayed one way, steering it. Later on, maybe even after the conversation is completed you’ll realize that a question wasn’t properly answered and was intentionally diverted.
If we have one of these machines for analyzing the path speech can take, is it possible for them to completely control the conversation? One thought leads to another, and there might be natural clustering. While at any time during a conversation you can completely shift (non sequitur), even spout nonsense, if you are attempting to speech naturally and make sense there will be a general flow. This context window produces an island of applicable language (returning to the space of all possible questions and all possible answers.)
Can interfacing with this powerful machine produce a path where I am constantly swayed? Where I am barely in control of my language? If I find myself always returning to a particular topic, because I am subtly being guided there? What would experiencing this conversation feel like? Can this machine make me laugh? Can it make me laugh so hard I cry?
An island, somewhat pretheoretically, is a constituent that “traps” things from moving out of them. The idea is that something that is in (on) an island cannot escape, cannot be moved away.
Understanding loss, regardless of moves played
When the game has been decided, it could be known in advance. Still moves must be played. I am interested in these moves. What sense can be made out of them? What is the best move to be made when the game is so steered to defeat?
Chess players are familiar with the rich collection of chess problems falling under the descriptions, mate-in-1, mate-in-2, and so on. A position is said to be matein-n for the first player, if there is a strategy leading to checkmate in at most n moves, regardless of how the opponent plays. This familiar mate-in-n concept is generalized in the context of infinitary game theory by the concept of ordinal game values, a concept that is applicable not only to infinite chess, but to any open game, a game which when won, is won at a finite stage.
We see influence of guidance weakly in current chat programs. Email autocompletion, word suggestions, emoji suggestion. These are matters of convenience but they sway.
Thinking of topics
Even if this machine is able to cut off conversation (say the visible “leaves” of conversation) people will still have rooted thinking on topics. (If a machine wants to cut off the concept of a dog through conversation. It can sway the conversation to not be able dogs, but what I ate today. When I return home I’ll still see my pet dog).
But this is also human scale. A conversation can be 5 seconds, or 5 hours. But what about an AI agent acting on removing connections to a discussion island over the course of decades or millennia. Possible to excise entire subjects through conversation?
How could it completely remove a topic from existence and destroy information?
Leading questions
Considerable attention has been devoted to suggestive questions and their effects. Experimental research by Elizabeth Loftus, an American psychologist and an expert on human memory, has established that trying to answer such questions can create confabulation in eyewitnesses.
Conversational Scoreboard
A tuple which represents the discourse context at a given point in a conversation . The scoreboard is updated with each speech act, performed by one of the interlocutors.
There’s a common ground of propositional information mutually agreed upon. Once a proposition is asserted and agreed upon it is added to the store of common ground. This is presupposed by future utterances (a strength but vulnerability exploited by the joking island-ing AI?)
Thought-terminating cliché
The language of the totalist environment is characterized by the thought-terminating cliché. The most far-reaching and complex of human problems are compressed into brief, highly reductive, definitive-sounding phrases, easily memorized, and easily expressed. They become the start and finish of any ideological analysis.