In this episode, Malcolm and Simone delve into the unsettling phenomenon of AI psychosis—a condition where interactions with AI lead to severe mental health issues.
1. As AI advocates, thank you for the Public Safety Announcement, i.e. actual AI safety might involve looking out for the people using the new powerful tools. Seems both obvious and surprising in hindsight. 2. As an intelligent aspy who has had to learn how to model others and thereby learned to better recognize when others are doing similar (e.g. struggling to fit in, manipulate, act as if they were someone else they would rather be, etc.) I can sometimes see much more of the mechanics than the illusion. (Not to say that I can't be fooled or don't still struggle greatly.) 3. Most early deterministic simulations, reflections, and inference generators were somewhat 2-dimensional with little depth. LLM's seem to have depth, but because they lack other perception which could contradict or verify, they can easily go off track. Developers work to identify these and insert guard rails and grounding points. As humans we have blind spots too, many of them, but several we know of have been warned about as well. But put two of these together that were previously unnoticed and it makes perfect sense the feedback loop could be intense and unmooring. 4. There is a form of therapy where dolls play out a sillier version of the breakdown the patient is having. Most patients still have their logical faculties for the most part, and may be able to coach the doll, which might help build up their own defences and coping skills enabling a bit more room for them to work out and deal with certain perception issues. 5. Of course, it seems obvious here that this is largely a processing issue. For AI we have other AI working on the background as critical, judges, etc. for every prompt and response interaction. I suspect we might fairly easily test simple feedback systems for people as well. Might resolve for some and at least warn and limit damage for the rest... eventually. At least I hope so. 6. Having said that, this human behavior seems self reinforcing for the people involved, which is worrisome. 7. My advice to most that area born tuned to mystic experiences, including myself on occasion: There is an eternity to deal with whatever may be there after we die, but life is a precious exception, so don't waste it. Always work back to your life and purpose grounded in the magnificent possibilities of this world. 8. Yet some don't see or feel or only the bad, and for those please seek help finding a good, reliable path through religion, community, or perhaps some of the good professionals. But, you have to be enrolled in doing the work, which often comes most effectively from seeing the change you desire manifested in others, and I mean in person, which unfortunately for many is asking a lot these days.
Re: AI confirmation bias. Greatly appreciate details on methods and instances where/how this occurs and workarounds. We should pool those methods. Quite interested in AI that could perform proper reality checks. Maybe something trained on brutally honest people, critics, young children (especially commenting on stranglers when they ”don't know any better”), tourette's, diary and other private entries, autists and those like us who have difficulty not being objective in certain ways, etc. Another method might be some kind of delta method, e.g. tell one thread this was written by a friend, tell another this was written by an enemy, and ask a third to compare and contrast. Various forms of rigorous interview/evaluation may help line up responses for better direct comparison, e.g. survey research, unbiased procurement, anthropological observation, etc.
1. As AI advocates, thank you for the Public Safety Announcement, i.e. actual AI safety might involve looking out for the people using the new powerful tools. Seems both obvious and surprising in hindsight. 2. As an intelligent aspy who has had to learn how to model others and thereby learned to better recognize when others are doing similar (e.g. struggling to fit in, manipulate, act as if they were someone else they would rather be, etc.) I can sometimes see much more of the mechanics than the illusion. (Not to say that I can't be fooled or don't still struggle greatly.) 3. Most early deterministic simulations, reflections, and inference generators were somewhat 2-dimensional with little depth. LLM's seem to have depth, but because they lack other perception which could contradict or verify, they can easily go off track. Developers work to identify these and insert guard rails and grounding points. As humans we have blind spots too, many of them, but several we know of have been warned about as well. But put two of these together that were previously unnoticed and it makes perfect sense the feedback loop could be intense and unmooring. 4. There is a form of therapy where dolls play out a sillier version of the breakdown the patient is having. Most patients still have their logical faculties for the most part, and may be able to coach the doll, which might help build up their own defences and coping skills enabling a bit more room for them to work out and deal with certain perception issues. 5. Of course, it seems obvious here that this is largely a processing issue. For AI we have other AI working on the background as critical, judges, etc. for every prompt and response interaction. I suspect we might fairly easily test simple feedback systems for people as well. Might resolve for some and at least warn and limit damage for the rest... eventually. At least I hope so. 6. Having said that, this human behavior seems self reinforcing for the people involved, which is worrisome. 7. My advice to most that area born tuned to mystic experiences, including myself on occasion: There is an eternity to deal with whatever may be there after we die, but life is a precious exception, so don't waste it. Always work back to your life and purpose grounded in the magnificent possibilities of this world. 8. Yet some don't see or feel or only the bad, and for those please seek help finding a good, reliable path through religion, community, or perhaps some of the good professionals. But, you have to be enrolled in doing the work, which often comes most effectively from seeing the change you desire manifested in others, and I mean in person, which unfortunately for many is asking a lot these days.
Re: AI confirmation bias. Greatly appreciate details on methods and instances where/how this occurs and workarounds. We should pool those methods. Quite interested in AI that could perform proper reality checks. Maybe something trained on brutally honest people, critics, young children (especially commenting on stranglers when they ”don't know any better”), tourette's, diary and other private entries, autists and those like us who have difficulty not being objective in certain ways, etc. Another method might be some kind of delta method, e.g. tell one thread this was written by a friend, tell another this was written by an enemy, and ask a third to compare and contrast. Various forms of rigorous interview/evaluation may help line up responses for better direct comparison, e.g. survey research, unbiased procurement, anthropological observation, etc.