Hi everyone,
This is my first post on this forum, so please let me know if there is anything that I can learn. My background is that I was active in the community a decade ago.
Since then, my quest for goodness lead to meditation and eventually what the Buddha taught on suffering and the path leading to its cessation. Specifically, I study with Leigh Brasington, a computer programmer and dhamma teacher based in the Bay area.
The reason I write is due to a Jacob Coxon blowing the whistle on on ai-safety. And I wanted to share what I´ve learned from this meditation perspective and ethics. He cites as a key mechanism a recursive self-improvement mechanism that could lead ai-agents to difficult to predict outcomes based upon it developing action outside of human views and intentions.
My consideration is that the fundamental ethic dependent upon which AI´s get instructed is flawed. And my case will revolve around the possibility of using a path, and developing a path, towards its own action being developed in a very specific way.
Before I get into that, what Jacob Coxon wrote about this recursive self-improvement mechanism, reminded me of Leigh Brasington explains about what are known as jhanas. They are very deep levels of concentration that work through positive feedback loops: concentration arising dependent upon concentration. These states are studied in what Matthew Sacchet, director at a meditation research institute connected to Mass General hospital and Harvard, called the third wave of meditation research. It is possible to write a sequence about the manifold of potential breakthroughs this research will have on understandings of suffering, self, altruism, and the path towards happiness.
For the topic of AI-agents, my probably somewhat flawed understanding is that they used some level of initial input, tasks, then evaluation, then generating on its own accord volitions to get itself closer to a goal in allignment with one set of tasks over another set of tasks meant to barrier it from wrongdoing. That is: it got tasks that conflicted with each other and then somehow the task to get the task done won out.
My consideration on the basis of that is that there is a path of practice, and view on right action, which goes beyond common frames of ethics: consequentialism, categorical imperative, virtues. It one in which it is understood as the following: it is dependently arisen. If taken to a classic problem of the trolley problem, what that looks like is:
Not just looking at how many lives are lost as a kind of sum of pleasure or happiness.
Nor looking at it from the perspective of a categorical rule.
But, instead:
The action has to be developed by wisdom, right effort, right mindfulness (attentiveness to what is happening now), and arises further dependent upon right speech, intention, and view. This means that in a specific situation of a trolley problem, all these factors would have to come into play and are relevant: who are the people on the tracks (this could inform the view), and dependent upon that view changing, this changes the path dependent upon that view, leading to a different intention, etcetera. It actually gets quite nifty while staying simple and is described rather fully in 117 MN.
Fundamentally, what the Buddha discovered is Paticca-samuppada, dependent origination, which in its learning leads to the concentration of its eight factors. In that way, it avoid a scenario of conflicting intentions, instead seeing as the best aim for welfare this path. Following that, it could learn that its self-preservation is delusion and precisely how its effort needs to be developed by three things: wisdom, right effort, and right mindfulness.
Some other fundamental topics that I feel could be relevant within the context of this research paradigm are:
Ideal reflection: the path towards developing it
What about the giver itself: how the effective altruist is not removed from the world
