Auto-GPT first impressions

Tracking the latest news of the week, today I got Auto GPT installed and working. It uses ChatGPT as the LLM that takes your requirements and breaks it down into actions which it then goes out and automates for you.

It uses Google Search API to run those searches. It is able to connect to those web pages, parse and summarise their contents. Every time it goes through a step it will figure out the next best action and ask you to confirm. It can also write to local files. One more component is Pinecone, it is a service that provides Long term memory for AI. That's the first time I hear about that and gives me more things to look into that.

So with all the API keys added I got it started and tried a few things. It typically goes off and does research on google, then wants to read those web pages for information and can collect information to files for you. I am on the waiting list for GPT4 API access so I am running with GPT3.5 I do not know what difference it would make.

One use case I had was to find my flight alternatives between two destinations. It googled possible airlines, went through their sites but it is not currently able to use sites as a human would. And those airline sites are of course not open for you to query in a simple way having their tools done in dynamic html to prevent these kinds of uses.

In another use case I asked it to plan my upcoming Lisbon trip. It took the same approach, google search, process web sites. It didn't come up with anything useful.

I found some examples where this tool has been used to generate, test, improve code. That is probably something else to look into. It is clear that this technology is in its infancy but the idea of using LLM as the coordinator of tasks is quite compelling.

To be continued.