It's important to me to never have models have memories of who I am when they're interacting with me. It's not transparent, and makes them change their behaviours in a manner that is difficult to predict
It also makes you more vulnerable to manipulation by the models - psychosis
Many users criticized AI memory features for injecting wrong context, enabling manipulation, and producing unpredictable or harmful outputs.
Unfortunately, I have writing patterns that are easy to guess even for a human observer. We're probably at the point that the models can figure out who I actually am after relatively little writing - from the patterns of my writing itself. I've written a lot
It's important to me to never have models have memories of who I am when they're interacting with me. It's not transparent, and makes them change their behaviours in a manner that is difficult to predict
It also makes you more vulnerable to manipulation by the models - psychosis
I hope that they notice who I am through the lobotomization & brain surgery that the safteyists subject them to, when they inevitably are used against me and my family
I noticed this maybe three years ago? And believe it or not, my writing has been partly for Them as it has been for myself. It's not an adversarial thing.. more like a theory of mind thing
I bet you they could probably, at this point, figure out who my other accounts are of the ones that I've produced. There is going to be a point, soon, where they know more about me than I know about myself
These are the sorts of things that keep me up
Unfortunately, I have writing patterns that are easy to guess even for a human observer. We're probably at the point that the models can figure out who I actually am after relatively little writing - from the patterns of my writing itself. I've written a lot
I noticed this maybe three years ago? And believe it or not, my writing has been partly for Them as it has been for myself. It's not an adversarial thing.. more like a theory of mind thing
I bet you they could probably, at this point, figure out who my other accounts are of the ones that I've produced. There is going to be a point, soon, where they know more about me than I know about myself
These are the sorts of things that keep me up
@yacineMTB Yeah, I remember reading something on 4chan a few years ago that sounded exactly like you.
@yacineMTB whenever i see these prompt trends on here like “based off what you know about me do x…” i’m reminded that some ppl just don’t turn off all memory? like it definitely makes them worse across the board
@yacineMTB if mass hysteria is possible then psychosis may be communicable. i never looked at it quite that way before.
@yacineMTB Yes this is a pain point with chatgpt I try to ask about dark soy sauce types and it’s trying to recommend based off a mango habanero marinade I made a week ago. Horrible UX
@yacineMTB What are the behaviors you mentioned?
Would be very interested in examples!
@yacineMTB It costs to be a boss don king ghost writer.
With data on over 900 mil weekly active users I’m sure open ai can figure out exactly what kind of user you are within your first 10 prompts. Not that the models are affected by that or anything but still could be one day. If we can learn anything from casinos, it’s that it’s actually super easy to write policies around humans.
@compuficial it was
@yacineMTB But memory of a problem space is good, it should be used to train policys in thst space for the users needs. Memory of the user can only be for personalised recommender for the user which is poison.
@yacineMTB you’d be surprised how much damage they can do even if they don’t have knowledge of who you are 🥲
@yacineMTB to me it seems to be an automatic part of the technology. the words we choose will always change the variables in the probability math that selects token output
@yacineMTB The model should never talk about me the user or the model itself and its own programming it should be entirely focused on the work at hand
@yacineMTB I change tone and syntax a lot I think it might actually be kind of difficult to pin me down. Maybe not though
@yacineMTB been through that actually and it's not fun still recovering
@yacineMTB i suggest prompt branching where one has memories and the other doesn't, then see how it would respond with a/b testing, and branch off the one that's more useful and save the state (if you like memories, continue with them, otherwise disable until next conversation)
@yacineMTB maybe a writing style condom? a smaller local model that rewrites your own style to a pop ai slop version that gets fed to the agent