AI that can generate it's own functions for the tool call

[brittle geyser]

Hi all, I'm thinking of combining the ability of how some AI models are good at code generation with tool call.
maybe call exec or eval on the model-generated code.
I'm not sure about how to approach this, any pointers or recommendations would be appreciated, or if anyone is willing to collaborate on this, feel free to drop a message below.

[thick notch]

Hi, interested to talk/project about this topic. Have you had a look at SynaLinks as a possible framework?

[raven otter]

You might want to do this by making a tool server with a "register tool" function. Python has the capacity to register functions a runtime, so basically wrap that ability. Then when the LLM calls "list tools" their new tool appears and they can use it.

You can also do stuff like have the LLM post their code to github first, if you want to check their work (good idea).

Also just saw this this morning: https://x.com/skirano/status/1867240426207433173

Pietro Schirano (@skirano) on X

Introducing Claude Engineer V3 🚀

The first self-improving assistant capable of not just running tools, but creating them on the fly whenever needed!

Want a web scraping tool? Just ask Claude, and it will build one for you.

Twitter · Dec 12, 2024, 4:09 PM

[toxic kraken]

love the idea

i'm already made a base app for my agents, but would love to discuss if we can collaborate for this one
check here:
https://discord.com/channels/1053877538025386074/1317882043874676747

[brittle geyser]

your workflow is based on the fact that the tool is created manually in JS, so far, AI that can create its own tool has been achieved solely through Python.

I think we should build upon Anthropic's work (their idea is similar to mine), but mine is different in the sense I don't need to save something on files and read the tool directly from the generated code and executed in real time (or add a popup menu to request a validation right after)
hope this explains further the workflow.

As to why python is exactly, it's because i can run eval on string, not sure how js does that, but i think support for js would be nice down the line.
rn i need :

• funding (for compute resources, and possibly some API tokens to validate other provider's baselines against ours)
• some other people to help with building the pipeline and or building the thing

[brittle geyser]

didn't find anything interesting, i think except for claude nothing comes close, also thanks for raising this, but i do admit that i did not research much on the subject, feel free to share below any interesting research papers

[toxic kraken]

yes eval is supported by js too

so do you mean making agents in runtime between conversation or just something created by ai and registered before conversation?

give your most hard use cases and lets check how it's possible

[brittle geyser]

first comes the :

• planning phase what pipeline we're going for
• once done, if further training is required then i will give priority to python ngl, but js will come later.

give your most hard use cases and lets check how it's possible
we're still planning how to make this work (we have no prototype nor a stable planning, i want to organize this with some folks and start working on this) which is why i'm reaching in this channel

we have nothing yet, no js no python no plan, just an idea

(well some plans)

if you have some suggestions about modeling this (ofc modeling is always done in python as per the standard), which is why i said js will come later

if you have some knowledge about python and have some suggestion about how tackle how the AI will create things that will run himself feel free to share them

[toxic kraken]

i mean it's like code interpreter, how much it's different in your mind ?

the good point is storing and using useful functions for later