GitHub - swax/NAISYS: Node.js Autonomous... | AI Programming And Chat | Page 1

edgy jungle Mar 9, 2024, 12:50 AM

#

looks cool. which model/s did you try with it?

#

i built something different but also terminal based. i ended up with wording like, "non-interactive only commands"... to cover vim, nano, apt -y, etc

echo creek Mar 9, 2024, 1:26 AM

#

edgy jungle i built something different but also terminal based. i ended up with wording li...

Gpt, Gemini, Claude, and minstral. Opus was probably the best. Right I told it like read/write entire files. I also made a wrapper for lynx that is non interactive but still allows following links

edgy jungle Mar 9, 2024, 1:43 AM

#

using nightwatch.yaml agent?

#

looks like it needs claude3 and gemini... ```shellModel: claude3opus
webModel: gemini-pro

#

pretty cool we get $5 free credit with claude

#

this is what autogen wanted to be 1 year ago

echo creek Mar 9, 2024, 2:05 AM

#

edgy jungle using nightwatch.yaml agent?

yea you can configure the agent to use gpt3/4turbo, claude3opus/sonnet, gemini-pro, or local for your own llm. the web model is used for reducing large web pages so you probably want to use something cheaper like gpt3 or gemini

edgy jungle Mar 9, 2024, 3:26 AM

#

cool cool

edgy jungle Mar 9, 2024, 4:17 AM

#

well my $5 free credit is gone. didnt get far either. but probably just as productive as a secops junior

echo creek Mar 9, 2024, 5:00 AM

#

edgy jungle well my $5 free credit is gone. didnt get far either. but probably just as pro...

yea im really interested in like creating an 'infosec' fine tuned model that could run locally, or at least on something cheaper like gemini pro or gpt3

edgy jungle Mar 9, 2024, 5:07 AM

#

there's a bunch of good deals right now. world be interested to know if it works with groq, as that is free mixtral and llama2

#

should work with open ai uri I already got it working with open interpreter which is similar

#

but I recon big business will eat this up at even this rate

#

might be able to proxy this through to free Google collab and or kaggle. I done it before with other ai apps

#

tin foil hats want private so need to find best models for that too

#

maybe point it at a knowledge base type deal so it's not always rediscovering

echo creek Mar 9, 2024, 5:51 AM

#

groq seems to work fine, these were my settings

#

in the agent config you set the shellModel: local

#

it's not very good though, probably less than gpt3 level

edgy jungle Mar 9, 2024, 6:25 AM

#

maybe try the llama model

#

although stock llama2 was pretty lame

echo creek Mar 9, 2024, 6:25 AM

#

yea neither were great, maybe w fine tuning

edgy jungle Mar 9, 2024, 6:26 AM

#

i might point it at my perplexity endpoint, it supports codellama70b-instruct

echo creek Mar 9, 2024, 6:26 AM

#

sometimes you gotta learn the model a bit and tweak the prompt for its specific behavior

edgy jungle Mar 9, 2024, 6:27 AM

#

yeah true. although for stuff that does function calling i suspect the fine tuning is worthwhile

#

like i'm deploying a model now to my home-assistant that is zephyr 3b with fine tuning. fit for purpose

echo creek Mar 9, 2024, 6:28 AM

#

the weaker models tend to output their own prompts, and guess the output of the commands they're putting in

edgy jungle Mar 9, 2024, 6:29 AM

#

true

echo creek Mar 9, 2024, 6:30 AM

#

i actually have some special handling for that where it'll automatically detect a prompt in the output and cut off the rest from there before processing

edgy jungle Mar 9, 2024, 6:30 AM

#

yeah i've seen a few unhinged detection workarounds

echo creek Mar 9, 2024, 6:30 AM

#

so you actually ran naisys and had it working?

edgy jungle Mar 9, 2024, 6:31 AM

#

yeah used up all my free opus creds

echo creek Mar 9, 2024, 6:31 AM

#

sorry about that

edgy jungle Mar 9, 2024, 6:31 AM

#

¯_(ツ)_/¯

#

is free

echo creek Mar 9, 2024, 6:31 AM

#

but thanks for trying it out, let me know if there's anything i could update in the docs

edgy jungle Mar 9, 2024, 6:31 AM

#

i had it on 60 second delay and the burn rate was fast, but still cheaper than an analyst

echo creek Mar 9, 2024, 6:32 AM

#

sonnet actually is pretty decent

edgy jungle Mar 9, 2024, 6:32 AM

#

i like the semi-autonomous UI design. is something the other cli's dont seem to get right

echo creek Mar 9, 2024, 6:33 AM

#

ive really only been building it to auto build a php website, didnt try the security stuff until a couple days ago

edgy jungle Mar 9, 2024, 6:34 AM

#

ai is great at all sysadmin cli stuff

#

i just read github copilot cli beta is available

#

if its like this plus unlimited gpt4 then wow

echo creek Mar 9, 2024, 6:35 AM

#

ah nice im gonna try that

#

im interested in getting agents to work together on a large project, like a web dev, db admin, content person, quality all working together - but they end up all talking to each other and not getting any work done, costing me $$$ lol

#GitHub - swax/NAISYS: Node.js Autonomous...