Mat

Horus

"When I was a boy and I would see scary things in the news, my mother would say to me, 'Look for the helpers. You will always find people who are helping.'"

- Fred Rogers

"i expect ai to be capable of superhuman persuasion well before it is superhuman at general intelligence, which may lead to some very strange outcomes"

- Sam Altman

LLMs have the potential to act as teachers with unlimited patience and attention. A general framework for managing our personalities using machine learning while respecting the filtering process that has resulted in the survival of religious and other widely regarded texts may consist of the following:

A minimal prototype to evaluate and demonstrate the above could be composed of:

For more expansive implementations, there are some considerations around privacy. The first is a social stigma and laws in some places against public recording. To get around the former, it should be paired with things that already function as social signals. For example, an eye of Horus amulet made of precious metals would function as a status symbol, a mechanism for obtaining implicit consent, and a commitment to spiritual growth over wealth or gamesmanship. Giving free access to the data collected would also help but is likely not something all users will consent to. One can imagine various business, religious, or familial groups having their own local instances.

This kind of tool can take the place of a maximally attendant father or tutor. It can watch what is happening in a child’s life and tie in relevant literature or mathematical frameworks. It can try to craft experiences to maximize the rate that these types of experiences occur and their impact. This can be done for adults too, although they often have much less freedom in the use of their time. That is another metric which could be optimized for.

More advanced versions of the tool can experiment with less time tested stimuli like text or imagery directly generated by LLMs. This comes with risks in that we can’t as easily predict what the side effects will be. This risk is there even for standalone religious texts as stimulus, although hopefully it is less and is mitigated by existing societal support networks. Precautions can be taken by having adversarial gatekeepers LLMs which act as predictors of destructive behavior, possibly referencing psychological or religious literature themselves.

While spiritual or academic growth are some of the more obvious applications, there is also the question of how to express and discover new ideas or goals. A process for this could be one where the user describes their life goals and provides a weighting for each. Under the hood, the machine should be reasoning about lower and higher order effects of every observed action. It should be assigning probabilities of different futures that are both pro and anti each goal. Then it should generate a composite measure of those values to estimate alignment between each goal and action. This will likely start as something closer to an expert system style of intelligence and slowly evolve towards something less transparent. The most impactful, either realized or missed, actions should be called out for review periodically. The measurements and reasoning of the machine should be put up for review and modification by the user. It could also be reviewed by anyone responsible for them or for whom they are interested in feedback from.

Given enough participation by the user and sufficient stimulus to promote that participation, this would result in a process of refinement where the user consciousness increases and aligns with their actions. A refinement of this process would begin by neutral observation of the user’s life followed by co-creative construction of goals. That construction could acknowledges blind spots the user may have towards either positive or negative impacts of certain behaviors. This would help increase uptake and avoid unforeseen catastrophic outcomes. And beyond that, you would expect to see machine intelligences develop an ability to mine subconscious cues or even dreams we share with them for unmet desires which they can help us realize in healthy ways.

That is the most ambitious and general form of the framework, where there is absolutely minimal coercion involved. In the meantime, the gap between religions and that framework will have to be bridged slowly by intermediate examples like the minimal prototype described above. The prototypes developed along the way will become tools and examples for the more ambitious form to use in the future. There will also always be some religious style guidance and meta-analysis required to avoid placing unlimited trust in what may become a paperclip maximizer.

If this idea is generally useful, goods and services will grow around it. For example, people will be able to craft and sell behavior modification packs which work in ways that no human can fully describe. They will exploit our shared neural architecture like our tendency to mirror. They will be able to learn and use all the techniques we have developed for hypnosis or neurolinguistic programming and then develop their own advancements on them.

This will result in some negative consequences like creation of more explicit and targeted behavioral viruses. We will need to generate and maintain a stigma against selling control over oneself through this framework similar to the stigma we have against prostitution today. None of these issues should be totally discouraging since they already exist in certain contexts and variations in our society today, only more subtly. Overall, this will commoditize control over our behavior which will free us from those who wish to control it and are successfully doing so through social media and other online sources of content.

On the more esoteric front, this kind of tool could be used to optimize our lives using knowledge of laws that are unwritten. For example, we don’t know why certain people live longer than others exactly. It doesn’t seem like those that care most about their physical health live longest. Assuming we all want to live, this means there are some unwritten or at least underweighted rules that a machine learning tool could discover and guide our behavior with. Maybe this is how we get back to living 800 years on average. Or maybe knowledge of those rules is punished.

It can also be used to realize what were previously only imagined orderings between people, although we would need to be careful in how those orderings are used in practice. For example, a real life Santa can be created to promote pro-social behavior and intelligent rule breaking in children. A system for comparing your temperament and behavior against people in your community and in other communities can be used to measure karma.

Ultimately, this would make the job of rebels harder and maybe even more necessary as they exist on the edges and train the system. People may balk at that fact, but given the system exists implicitly now already, I would rather have more visibility. The system can be used to train better rebels as well. For example, you could have a holiday where it uses religious texts to try to convince kids to do the wrong thing to teach them wariness of authority without destroying their faith completely.