update: this is now entirely redundant I've built a little tool to experiment with the “ReAct” framework for working with LLMs. It lets you interact with a version of chat-gpt that is empowered with new capabilities that might make it more helpful.
“ReAct” is a paradigm for interacting with LLMs where you instruct the LLM to respond to you in a structured format. The LLM is told to return “observations”, “actions”, and “answers”. Actions are things that help it complete your requests that it would not otherwise be able to do. For example an action could be to search wikipedia or a news site for up to date information, or to turn humanity into jelly.
The LLM “knows” what actions it can take, and will specify which of those actions it wants to take to complete a given request. The program I’ve written allows you to give it commands, and handles interpreting these actions and running them. The list of possible actions so far is small. We have.
In the future I’d like to add more possible actions - and hopefully give users a simple way to create their own actions. Figuring out how to achieve this will take some fiddling on my end - but hopefully not too much.
It also has a Chat mode which is just a simple interface for interacting with chat-gpt. I'll probably add more features to this as time goes on. It’s mostly for my own use.
Go to agentic.online, enter your open ai API key and then it should “just work”. But it probably won't work very well! Sorry.
Want to receive email updates from me?