Ride on the AutoGPT ride

Also you will need vector storage for long term memory of the bot, I used Pinecone – registration is very easy.

To use Google, you will need to get the keys from the api.

1. Let’s go to console and create a new project

   

2. Next to the API dashboard

   

3. Create an API key in Credentials

   

4. Next you need go and click on enable button

   

5. Turn on your custom search

   

Setting up the project

• Create a folder for the project

  mkdir autogpt
  cd autogpt

• Create an environment

  conda create -n autogpt python=3.10
  conda activate autogpt

• Clone the repository and install the necessary libraries

  git clone -b stable https://github.com/Significant-Gravitas/Auto-GPT.git
  cd Auto-GPT
  pip install -r requirements.txt

• Next, you will need to rename the file .env.template just in .env and edit to suit your preferences:

  OPENAI_API_KEY=sk-O2sOu...N9E4 - ключик апи OpenAI
  TEMPERATURE=t - можно выставить свою температуру, по дефолту 0
  
  PINECONE_API_KEY=76b90....6  - ключ к векторному хранилищу
  PINECONE_ENV=us-west4-gcp - регион(смотрите ваш)
  
  GOOGLE_API_KEY=AIzaSyBms7...tgmwRHc - ключ API на этапе 3
  CUSTOM_SEARCH_ENGINE_ID=b2daab24c13324faa - айдишник на этапе 5(<script async src="https://cse.google.com/cse.js?cx=b2daab24c13324faa">)

In general, everything is set. You can also tweak a bunch of other features: the twitter API or voice dialing.

We start

Before testing, you can see the list of available settings:
python -m autogpt --help
You also need to create a file auto-gpt.json in the root directory

Let’s run in mode gpt3only
python -m autogpt --gpt3only

You will now be prompted to select a name and role for your AI. And then they will ask you to write goals. Let’s go.

In the process, he will ask you for confirmation of his actions.

• y – you approve the current step

• y -3 – you approve 3 steps at once

• write your prompt

The agent saves its fabrications in the auto_gpt_workspace folder

In the process, you can observe how he works with information. And here you need to be on the alert, because each request is paid, and it can quickly gobble up your API budget.

After an hour of work, about 50 cents were spent, and the AI ​​​​created several artifacts of its activity.

But then the agent became stupid. I moved on to the stage of searching for engines for creating courses and began to perform the same cycle. Even a direct instruction did not cool his ardor. At this point, I had to turn it off. The career of a successful infogypsy is now in jeopardy.

Conclusion

AutoGPT is a pet project that has come to fruition. It works quite clumsily, and it can be used mainly for entertainment purposes. But the very idea of ​​agents is very cool and will obviously be developed. There will be a lot of work for AI developers to create systems of control and feedback that allow agents to learn and develop, but not go beyond the allowed parameters. In the last article I wrote about the framework LangChainin which the functionality of agents can be elegantly customized to suit their tasks, these solutions are the future.

I write about AI and NLP in telegram.