LLM With Straps : An LLM Client That Bootstraps Itself

Have you ever wondered what happens if you let a Large Language Model client modify its own source code ? No need to wonder anymore with LLM with straps, a tiny shell utility that can modify itself to tailor it to your own specific needs. In other words, this LLM shell client can bootstrap itself.

Bootstrapping ? What do you mean ?

Bootstrapping refers to the ability for a compiler to compile itself. In this context, it means that this utility can fully understand its own source code and modify itself. This lets you change the utility only by using it.

Little demo

$ python llm-straps.py --boot "Modify the version of chat gpt used to do requests from 4 to 3.5."
"Function 'query_model' modified successfully to use GPT-3.5."
Output written to llm-straps__1.py

$ diff llm-straps__1.py llm-straps.py
3a4
> 
145c146
< def query_model(messages, model='gpt-3.5-turbo', temperature=1, stops=[], debug=True):
---
> def query_model(messages, model='gpt-4-1106-preview', temperature=1, stops=[], debug=True):
150c151
<     model (str): Model identifier, default 'gpt-3.5-turbo'.
---
>     model (str): Model identifier, default 'gpt-4-1106-preview'.

$ cp llm-straps__1.py llm-straps.py # we now use the generated llm-straps as our main one, and it now uses chat gpt 3.5 to do requests !

Funny enough, this change has been incorporated in this git commit : e37254d

Another demo with a beep command line utility :

$ python llm-straps.py --boot "Add a simple command --beep in the 'main' function that prints boop to the screen"
Function 'main' modified successfully.
Output written to llm-straps__3.py

$ python llm-straps__1.py --beep test
boop

$ python llm-straps__1.py --boot "Remove the --beep command line argument in the main function"
Function 'main' modified successfully.
Output written to llm-straps__1__1.py

$ python llm-straps__1__1.py --beep test
usage: llm-straps__1__1.py [-h] [--boot] [--debug] prompt [prompt ...]
llm-straps__1__1.py: error: unrecognized arguments: --beep

Setup

To run it, you must create a virtual environment :

> python3 -m venv .venv
> source .venv/bin/activate
> pip install -r requirements.txt

Then, you must provide your open ai API key :

export OPENAI_API_KEY=<your api key here>

And you are ready to rock !

> python llm-straps.py What is the meaning of live \?
The word "live" can have multiple meanings depending on the context. Here are a few common meanings:

1. Adjective: In the state of being alive or living, as opposed to being dead.
Example: "The live performance of the band was electric."

2. Verb: To have an active existence, to be alive, or to carry out life activities.
Example: "They live in a house by the beach."

3. Adverb: Happening or taking place in real-time, without any delay or recording.
Example: "We watched the live coverage of the football match."

4. Verb: To reside or stay and make one's home in a particular place.
Example: "Many retirees choose to live in warm climates."

5. Verb: To experience or undergo something.
Example: "She had to live through the pain of losing a loved one."

These meanings can slightly vary depending on the context and usage.

History of the idea

This idea was born in a Compiler laboratory at the University of Montreal after a great talk by Ian Arawjo. The discussion went something like this :

Ian : We have decided to integrate completion tools powered by ChatGPT inside ChainForge. This is a bit circular as ChainForge is used to study LLMs and now we are using ChatGPT to complete part of the prompts used to help us analyze ChatGPT and other models.

Marc : In compilation, we are used to it, it’s called bootstrapping.