×
✍🏽 writing 🌎 travel ⚒️ projects 📬 email me 📝 resume

BrowserBot

2 min read — #Tinker #AI #Code #Technology

Built at Scale's Generative AI hackathon — an action transformer for automating common tasks through a browser

h/t karpathy

h/t karpathy

What is this?

This is a fun project that I worked on with some folks at Scale’s Generative AI hackathon. The intention was to build a lightweight “action transformer” that could help automate simple and repetitive tasks to the user by leveraging the power of LLMs.

From our official GitHub repo:

Our project is a grounded action transformer that uses language models to decompose complex tasks into simple instructions. We were inspired by recent developments in action transformers and action generation, but noticed a lack of open-source options. We chose to use natural language as the medium for communicating tasks to the model. Our website is built using Node.js, Puppeteer, Express, and a vanilla frontend. We use GPT-3/OpenAI API for our prompts. We are proud of the ability for our product to work out of the box for a limited set of tasks and look forward to the potential for community-created tasks to expand the set of actions.

^ Here’s a demo — view it in fullscreen on your browser, my monitor is WAY to huge but I love it.

We essentially created a set of low-level instructions that we “trained” a model on through prompt engineering and taught GPT to generate these instructions for itself, given a high-level task. We then used Puppeteer to execute these instructions in the browser once these tasks were handled individually by GPT itself or custom-built adapters.