It ignores it literally every time lol

I use this little script https://github.com/jmacc93/paste-to-tmux-script daily. Its for pasting your current clipboard item to a target tmux session, along with very simple dsl for controlling which tmux target to send to, opening the tmux target, killing tmux sessions, etc. It makes a vastly, vastly better and more productive repl for me. The workflow for using it looks like typically is like: use `@sw name_of_target`, `@open`, `name_of_executable`, then move my cursor to something I want to evaluate, use ctrl-c to copy the line, and ctrl-. to send it to tmux

I think I've changed stuff around since I last updated that repo, though, but the general idea and skeleton is there, if anyone wants to use it / hack on it. And I'm gonna do a rewrite of it soon, I think, so that instead of using a dsl (the `@...` forms above) it uses `!...` or some similar form to execute arbitrary shell commands

I'm hoping to (soon hopefully) integrate it with my https://github.com/jmacc93/noca notebook canvas program as well

I'm very interested in this topic and if anyone is interested in talking about it, please let me know. You can also email me at joe_miller_93@protonmail.com

I'm thinking about trying to write and train a superagent (an agent composed of agents) for my next project. If anyone is interested in talking about that, again: let me know :)

Comments URL: https://news.ycombinator.com/item?id=39490629

Points: 2

# Comments: 1

I hope you all like it :)

P.S. the demonstration requires javascript but the other stuff on the page I wrote in markdown then used pandoc to turn that into html and manually spliced that into the page's index.html. It pulls CodeMirror from the web, and I think that's it iirc

Comments URL: https://news.ycombinator.com/item?id=39239427

Points: 3

# Comments: 1

For example, you would say "Don't apologize", and then after that you would start every message with "(Remember to not apologize)"

I've also learned from using local LLMs that if you force the LM to start its response with something (via the 'start response with' field in text generation webui), then it will just go with that thing. This can be used to prevent RLHF-trained models from not responding because they think the question is unethical or illegal, and from giving the typical "I am a language model bla bla bla" responses. To be clear, if you put in the start response with field "Sure, I can answer that!", the LM will just go with it and not respond with "that's unethical" or "you're a horrible person for even conceiving of such a thing. they should lock you up for life!!", etc etc. It seems there is a similar effect when you edit the LM's past responses in some way, the LM's new responses will mimic that way of responding

Carrying that over to ChatGPT, it seems that if you request ChatGPT always starts its replies (via parenthetical reminder as well) with "Sure, I can answer that!", or with "I apologize", that that does seem to affect how it starts its replies. There appears to be some cases where it will say, for example "Sure, I can answer that! As an AI language model, ...", but it seems that forcing it to start its response a particular way helps prevent it from apologizing

But generally, for the apologizing thing, I just downvote it when the apology doesn't make any sense, and otherwise ignore the apologies, as others are doing. This indicates a potentially hazardous 2nd order effect where people are trained to ignore ChatGPT's apologies. eg: Boy who cried wolf, etc

I'm like 70% confident that a LLM that is constantly condensing its own memory (its immediate memory / input, and condensed memory) using internal prompting, and conditioning its output using both that condensation and its immediate memory, would lead to effective long term memory. You could probably make a model that does this on multiple scales, with condensations on top of condensations, and so on