#OpenSourceDiscovery 75: AutoGPT
Autonomous GPT
Today I discovered…
AutoGPT
A Python script that chains together GPT "thoughts" to autonomously achieve whatever goal you set. Unlike ChatGPT, AutoGPT doesn't require follow-up user input to achieve a larger goal. The user just needs to provide a high-level goal in natural language.
🛠️ How it works:
AutoGPT prompts GPT to create a plan to achieve that goal by executing tasks such as Google search, web scraping, image generation, code analysis, bash commands, etc. It then recursively optimizes the plan with self-criticism. AutoGPT takes care of creating the next GPT prompt and executing the tasks.
💖 What I like:
Text to speech: Instead of constantly looking at the screen, it is much more user-friendly to "hear" what AutoGPT is doing and what its next plan is.
Limit to n steps: Having AutoGPT run continuously would have emptied my bank account due to the high cost of GPT APIs. However, having the mode where it asks the user for permission to continue is quite useful. After spending 1 cent on the task, it became clear that it was not going to achieve the goal, so I stopped.
👎 What I dislike:
Token limit: The token limit of OpenAI GPT APIs makes it almost impractical for most use cases. I couldn’t get it to finish a single job due to this limitation. I am not sure how effective prompt summarization can be to overcome this because summarization comes with a loss of context.
Headless browser support: Unfortunately, it does not have headless browser support. This feature would have been helpful for running it in the cloud, saving system resources, and aligning it more with its goal of being autonomous.
Security: It is not safe to use this on your personal computer. It is possible that it could scrape a malicious code from the internet and save it on your computer.
Author: Toran Bruce Richards @SigGravitas
Demo | Source
Tech Stack: Python, OpenAI GPT APIsKnow of a cool open-source project? Share with the community