• 22 Posts
  • 295 Comments
Joined 1 year ago
cake
Cake day: June 9th, 2023

help-circle





  • Apple is doing insane stuff with iMessage making SMS useless with Android. It is intentional manipulation. I’ve only seen it on some Apple kernels with the latest iPhone. My old man’s phone absolutely will not send or receive an SMS unless I message is manually turned off and then it won’t get iMessages. Apple has always been a shitty company, but this is next level insane. Just get and use the Signal app and ditch all the proprietary garbage and manipulative bs IMO.



  • That is all model loader code based stuff. You basically just need a bit of Python code that anyone can write, even you. You need to have a look at how the model must be addressed in format, then you find how the model loader code you are using creates the structured text and just insert your own modifications.

    It is easier to add a second prompt window where you keep persistent information and just change it yourself. If you want the model to change that stuff, you need to look into building an agent.

    I hacked around like this for awhile too, but evolved into using the notepad tab of Oobabooga Textgen WebUI and a Mixtral 8×7B model with a 32k context window. That model can deal with almost any detail you add in the context.

    The thing is, the model may not feel satisfied with mundane and redundant details. Most of the time, believe it or not, those changes that you perceive as errors or inconsistencies are not actually errors. The model is fully aware that it is making those changes. There is usually an instruction about creativity, inspiration, or initiative that will cause the model to shift details like this.

    My best recommendation is to openly discus the error in dialog with the model. They usually will talk about the issue if you call one out. Talking with AI is very different than with humans. Everything they say has meaning and functions beyond what is initially apparent. There are a lot of behaviors that involve certain patterns and keywords. You can even alter these behaviors by banning certain tokens/keywords. A good place to start is banning paragraph starting words for filler material. “So_” would be a good starting token to ban, as with AI, no sentence that starts with that token is worth reading. Other important words are twist, chuck, and cross. These are more flexible in their troublesome meaning and can manifest in multiple words like chuckles or across but still function the same inside the model.

    If you get to this level of familiarity, the persistence of a storied interaction is more of a negotiation with the AI.

    Another point of understanding is that the model does not know the difference between roleplaying characters it owns and you own or who is who. You or the model loader code is creating the name of the next character that the model generates for. Everyone in any story can be the model if it is asked to create the prompt.

    So if I say, “Jake is super creative” or imply it somewhere in the prompt, and the model later thinks ‘Jake is rather bored and needs to be creative’ it is going to change something that satisfies that character profile even if it has never generated the output for character Jake.

    Hope that helps some



  • Hike sound like 1st date PC to anyone? I mean, it depends on the definition of hike but someone I barely know dragging me to an isolated place in the woods somewhere sounds like a legit way to end up being drug through the woods with a much higher probability than the typical coffee shop date.

    Also, who doesn’t have work-mances without deeper meaning? I always have beer after work bros and lunch/break dates with my coworkers but that has no meaning beyond. In fact, I’ve always had a strict rule of never shit in my own back yard. If I was so inclined as to date someone from work, I would not do so until after finding a new job.








  • I’ve had this happen with AI stuff that runs in a Python venv. It only happens with apps that use multi threading, and usually when something is interrupted in an unintended or unaccounted for way. I usually see it when I start screwing with code stuff, but also from changing the softmax settings during generation or crashing other stuff while hacking around. There may be a bug of some kind, but I think it likely has more to do with killing the root threading process and leaving an abandoned child that doesn’t get handled by the kernel process scheduler in the standard way. If this happens I restart too.