Thread by @yoheinakajima

🔥1/8
Introducing "🤖 Task-driven Autonomous Agent"

An agent that leverages @OpenAI's GPT-4, @pinecone vector search, and @LangChainAI framework to autonomously create and perform tasks based on an objective.

"Paper": yoheinakajima.com/task-driven-au…

[More 🔽]

🚀2/8 The system can complete tasks, generate new tasks based on results, and prioritize tasks in real-time. It demonstrates the potential of AI-powered language models to autonomously perform tasks within various constraints and contexts.

💡3/8 The autonomous agent uses GPT-4 for task completion, Pinecone for efficient search and storage of task-related data, and the LangChain framework to enhance decision-making processes. #GPT4 #Pinecone #LangChain

🎯4/8 The system maintains a task list for managing and prioritizing tasks. It autonomously creates new tasks based on completed results and reprioritizes the task list accordingly, showcasing the adaptability of AI-powered language models.

🔧5/8 To complete tasks, the system uses GPT-4 and LangChain's capabilities, enriching and storing results in Pinecone. This integrated approach allows the AI agent to interact with its environment and perform tasks efficiently.

🧠6/8 The system generates new tasks based on completed task results and prioritizes them using GPT-4. This allows the system to adapt and respond to new information and priorities.

🔮7/8 Future improvements include integrating a security/safety agent, task sequencing and parallel tasks, generating interim milestones, and incorporating real-time priority updates.

🤝8/8 This new approach paves the way for AI-powered language models to autonomously perform tasks within various constraints and contexts, enabling new applications and opportunities. Big thanks to all involved! #AIResearch #GPT4 #Pinecone #LangChain

📜 APPENDIX

🧵Thread (above) generated by GPT4 based on paper
📄Paper generated by GPT4 based on code
📊Graphs in paper generated by GPT4 based on code
💻Code generated by GPT4 based on prompt

*For each, many prompts to adjust initial output

Backstory 1/5:

Honestly, I was just trying to play around w the idea of an "AI founder" after seeing the awesome #HustleGPT movement.

That led to this prompt 2 days ago.

https://twitter.com/yoheinakajima/status/1640015750172930048?s=20

Backstory 2/5:

About 50 prompts later (dev docs, error codes, etc.), I shared this working prototype.

It's amazing that its first task is to create its next task - and it just keeps going.

https://twitter.com/yoheinakajima/status/1640172214871031808?s=206258707456?s=20

Backstory 3/5:

Realized it could be provided any core objective, in this case "make the world a better place".

Pretty fascinating to watch but also scary.

https://twitter.com/yoheinakajima/status/1640174201423433728?s=20

Backstory 4/5:

Interestingly, when I asked it to generate as many paperclips as possible, it first generated security measures.

Which was then picked up by the creator of the Paperclips Apocalypse theory himself.

Led to lots of AI safety reading.

https://twitter.com/ESYudkowsky/status/1640511156254289926?s=20

Backstory 5/5:

Sharing the original experiment led to many shared concerns and potential counter measures being shared publicly. Including awareness of what people are likely doing privately.

I believe this is a good thing.

Agh, my site has security issues.

So found the second best place to post it.

linkedin.com/pulse/task-dri…

And here you go!

Open-sourced a paired down version I’m cheekily calling “Baby AGI”.

Repo in thread:

https://twitter.com/yoheinakajima/status/1642881722495954945