😱😱😱BabyAGI😱😱😱
Guys have implemented an idea that I've been thinking about for a long time.
What if we set GPT-4 a high-level task (like making the world better or conducting research in a specific direction) and then run a sequence in an infinite loop:
1) Use GPT-4 to break the task down into subtasks
2) Add all subtasks to a list
3) Prioritize the list
4) Execute the task in the same loop (break it down into subtasks, add to the list, etc.)
It's interesting, but in this approach, the neural network
a) goes into very interesting ideas
b) spends a lot of API tokens 😂
P.S. Initially, the project was conceived as an AI startup, but it became clear that the approach is applicable to any task.
P.P.S. What I was thinking about and what is lacking in the project is real execution. The guys "execute" the task using a linguistic model while I was thinking in terms of "generate code" which I would send to Python execution.
The guys have open-sourced the project.
Github: https://github.com/yoheinakajima/babyagi
Paper: https://yoheinakajima.com/task-driven-autonomous-agent-utilizing-gpt-4-pinecone-and-langchain-for-diverse-applications/
UPD: Houston, we have a problem!
Other guys have open-sourced a similar project that not only can browse the internet but also generates code and tries to execute it 😱😱😱
Github: https://github.com/Torantulino/Auto-GPT
❗️ Subscribe to @rvnikita_blog to stay updated on similar news and thoughts ❗️
