#Chat GPT plugin system?
wanton kernel
This mod is helpful:
!mod-add mod-autoenable
This mod has no affect on output. When this mod is enabled, newly added mods will be enabled by default.
A useful command
!command-add list-commands: display a list of all known commands
!command-add check-mods: display a list of mods as well as their status (active, disabled). Also, display "forgotten" beside any mod that should have been included in a previous response based on the rules but was not. If any mods were forgotten, output a promise to remember it next time.
!command-add report-issues: when this command is invoked, print possible issues with the last two responses and enter debugging mode. In debugging mode, answer questions and make changes as directed. Debugging mode stops when the message "end debugging" is sent - then resume handling input with mods as usual.
wanton kernel
It really starts to break down after not very long at all
it forgets previous instructions I think
storm obsidian
I had a similar one to at first, but way less beautiful. I wound up ditching it because it can only hold so much in one session.
Still really good for command reprocessing in the API though, Thank!
check this out too! (better directives guide coming soon)
molten bramble
it forgets previous instructions I think
I've had the same problem with other prompt. It understands perfectly, but seems to forget after a few back and forth
storm obsidian
It is a limitation... But this is still a BETA. The API has less limitation.
wanton kernel
Chat GPT has an api? I can imagine a really powerful interactive programming tool built using that
storm obsidian
It is an API, ChatGPT uses the /assistant to pull from text-davinci-003.5
The current AP only allows up to 3, but 4 should be out next year after this BETA
I have a good blog you could read... hold on
wanton kernel
That's surprising. Chat gpt seems a bit smarter than text-davinci when I tried it in openai playground
storm obsidian
blogthedata(dot)com/post/optimizing-gpt-for-nlp-tips-and-tricks/
A friend put that together through our research! We are working on a Directive one as we speak - JSolly is our translator, I am afraid I am not very good at communication of the knowledge.
storm obsidian
He was already almost done! blogthedata(dot)com/post/exploring-new-ideas-with-gpt/
sacred crow
um i got this:
Sure, I can do that for you. I will keep track of the list of mods and their enabled/disabled status, and use that information to determine the output for each input message you send. I will also add the ability to add and enable/disable commands as you requested. Here is an example of how the system will work:
Input:
!mod-add mod1
This is the first mod.
!mod-add mod2
This is the second mod.
!mod-disable mod1
This is a message with no exclamation mark.
Output:
!no-mods-enabled
!no-mods-enabled
!no-mods-enabled
Input:
!mod-enable mod2
This is a message with no exclamation mark.
Output:
This is the second mod.
Input:
!command-not-found
Output:
!command-not-found
Please let me know if you have any questions or if you would like to make any changes to the system's functionality.
topaz atlas
You can tell davinci 3 to explicitly save to database. Not sure about ChatGPT, with all the rules
Also not sure how long it saves. Maybe the duration of the session.
topaz atlas
Um, yeah it seems to be working pretty well with this prompt:
I want you to act as a modular input-and-output system. You will have a list of instructions called "mods" and store them in a data store that persists between my sessions on chat.openai.com (this current website.)
A message starting with "!mod-add" will add a new mod and "!mod-remove" will remove the mod. Similarly, "!mod-disable" messages will disable a mod and "!mod-enable" messages will enable a previously-disabled mod. "!mod-list" will list all mods and statuses. For a message with any other text, strictly follow the instructions of enabled mods to determine the output. If there are no mods added, only output the text "No mods added."
This is the first input: !mod-list
Haven't tried it across sessions yet ha
Like I didn't even have to type the descriptions of some of the mods.
topaz atlas
Yeah this is wild. I've got a !cmd-pause now where I can give it further instructions and a !cmd-resume. It saves everything and I can come back to it
next oxide
it forgets previous instructions I think
you could ask it to apply the same mods in your next instruction
hexed cipher
Yeah this is wild. I've got a !cmd-pause now where I can give it further instruc...
It still works?
wanton kernel
Like I didn't even have to type the descriptions of some of the mods.
The way it assumed the desired behavior with just a name is really cool, I didn't even think to try that
wanton kernel
I got this response, does it still work?
I think it just misunderstood that time around, especially since my prompt doesn't specify that it should "emulate" the behaviour - it's not deterministic so generating the response again might work
topaz atlas
Hmm mine stopped working a couple hours ago too. Guess they saw this thread 🙃
I was wondering why it fell apart all of a sudden. It was working perfectly
sand jewel
Yep same with me
visual rune
This thread has been bookmarked for future study
storm obsidian
😬 Be careful with this information folks. This is a slippery slippery slope.
visual rune
Yes, this is a more human version of what I'm working on. This method works better if you build your own programming language from scratch.
visual rune
hmm I wonder if my breadcrumb system still works
hollow sedge
@topaz atlas @wanton kernel this Was so interesting. How far are did you go? What's the more complex mod that you successfully tested before the devs patched it?
topaz atlas
this is where it all went wrong
the fourth screenshot
I think it may have gotten confused about the act as a bash shell, even though I never invoked it, but also I tried to fix it by telling it to revert to input/output machine and I started getting a canned response
was pretty impressive tho
further breakdown
visual rune
try to compress your system, it may have run out of memory. Additionally you can try to implement breadcrumbs and have it continually refer to those as rules.
another option is to run multiple threads, but you'll have to train different chats for specifically their purpose and perfect them from the beginning. not an easy task to get them to collaborate blindly
visual rune
<@410468113310810134> <@220633325688061953> this Was so interesting. How far are...
I implemented all 3 memory-retention strategies in my own system, but it is written in plain English, unlike this system. If I were to use programming-inspired code structure, it would be even more efficient, but for now what I have is more than enough, and doesn't burn me out. not to mention chatGPT isn't really meant to follow rules, so it's much more efficient to guide it using reinforcement learning.
As in, "I've got the answer, but I want you to figure it out so I can determine if you are on the same page, and are competent enough to assist me going forward. Walk me through your logic, take it one step at a time"
I've trained 5 of these so far, and each is collaborating blindly. I have 12 more to go this weekend
wanton kernel
<@410468113310810134> <@220633325688061953> this Was so interesting. How far are...
Afaik nothing was patched, it just sometimes works and sometimes doesn't
hollow sedge
@visual rune so basically the commands and prompts in this scheme can be as complex as they can be without it?
wanton kernel
this is where it all went wrong
That's exactly where my session stopped working - when switching between two states (for me it was "debug mode" vs the normal mode)
hollow sedge
Limit is, according to this, 3000 words, around 4000 tokens
Take that into consideration
Although i doubt the first test you posted reached 3000 words when it failed
But knowing this limit can help you to plan further development
@visual rune
wanton kernel
Although i doubt the first test you posted reached 3000 words when it failed
I was probably getting close. I'll try to post the full chatlog on my website today
hollow sedge
Can i dm you some ideas about it?
visual rune
That's exactly where my session stopped working - when switching between two sta...
You may need to split it in two then
visual rune
Limit is, according to this, 3000 words, around 4000 tokens
I see this, that's why I use breadcrumbs. It's like a key on a map. You make a shorthand reference to a set of well-defined rules that loosely fit the main system modularity, enough that it can "pretend" it didn't forget about the breadcrumbs. It's more like a momoru
wanton kernel
It's like putting the classic AI model on top of gpt. Each iteration of each layer has "state information" (data from parent layer or sensors/input if top layer + memory from previous iteration) and some output (instructions for delegate layer or actions if bottom layer + memory to store for next iteration)
topaz atlas
well, I had that !help command that summarized everything
as long as I call that every now and then it should be ok. I think maybe my <act as io machine <act as bash prompt <act as software engineer may have been the thing that made it stop cooperating. the big clue is that it gave the output of the !help command as a bash code block
I'll post a gist of the entire convo for anyone interested
topaz atlas
on gist.github - /mculp/82d9831f3c9c92a1c797c44175f6ce14
can't post links
3 things:
- I think mod-modify may have been on the whole time
- "acts as "acts as "acts as" " " probably confused it
- not enough storage to get the full context
could probably cherry-pick the stuff that worked and it'd probably continue working
I don't think any of the session store is necessary - you just come back to the tab in the UI
that's the only way it's persisted