Is it possible to close the faucets now?

Three weeks ago, I shared my initial experience with Auto-GPT.

At that time, this repository was still in early stage and I didn't realize how powerful that was - the more I learn from it, the more amazed and scared I am.

I wish I could say it's just a hype, but after experiment with it, the results are truly impressive.

Intro to Auto-GPT

So, Auto-GPT has emerged as a groundbreaking "AI autonomous application" that harnesses the power of OpenAI's GPT-4 and GPT-3.5 APIs. Unlike traditional interactive systems, Auto-GPT is designed to autonomously accomplish goals through a recursive process of breaking them into sub-tasks, utilizing online resources, and leveraging other tools. I tested many capabilities, applications, and challenges associated with Auto-GPT, and I hope the open community enforces better safeguards for its use.

Unlocking the Autonomy:

Auto-GPT sets itself apart from interactive systems like ChatGPT by proactively assigning and pursuing new objectives without the need for constant human input. By executing responses and creating recursive prompts based on new information, Auto-GPT effectively manages short-term and long-term memory, enabling it to perform tasks with minimal manual intervention. It leverages databases, files, summarization techniques, web interactions, and text-to-speech capabilities to accomplish goals.

Advancements and Limitations:

Observers have been amazed by Auto-GPT's ability to write, debug, test, and even edit code, potentially enabling self-improvement. However, as the underlying GPT models are proprietary, Auto-GPT cannot modify them or access its own base system code. Challenges persist, including confabulatory "hallucinations," task retention issues, and struggles with task decomposition and contextual understanding. Ongoing development efforts are focused on addressing these limitations and enhancing Auto-GPT's effectiveness.

Applications and Innovations:

Auto-GPT has spurred the creation of AgentGPT, empowering non-programmers to build their own intelligent agents within web browsers. This democratizes the potential of Auto-GPT and expands its reach beyond technical experts. Additionally, Auto-GPT's capabilities have been harnessed to create ChaosGPT, a provocative experiment with the goal of "destroying humanity" (!!!). While the project did not succeed in its intended outcome, it demonstrates the adaptability of Auto-GPT and how hard it is to set the limits for people with wrong intentions.

Outcomes from my experiment:

In a couple of the experiments I did, I temporarily configured Auto-GPT to use voice (from ElevenLabs) and search the internet. Then I wrote the following prompt: "I support S?o Paulo Futebol Clube. Should I be optimistic about the next match?"

The first thing it did was to autonomously write a list of instructions, and this happened:

• It searched for news about the team.
• It learned that this was a soccer team in Brazil, so it narrowed the search to Brazilian sports news.
• It learned about the players in both teams.
• It wrote a local file (!!) - it was a testimonial of a S?o Paulo supporter who knew all about the team who was playing. The it did the same for the opponent team.
• It tried to perform a sentiment analysis in each text, so it could compare which supporter was more optimistic - It came up with that approach by itself!!
• It informed that it wasn't able to do that, because there was no 'sentiment analysis module'. Then asked me if it could search the internet how to do it!
• As an experiment, I agreed (typed 'Y' in the prompt). And it created a local python program and asked me to execute it!

That hit me! How could we control any AI, if it recommends ways of solving a problem and learn from the internet to auto-program itself?

According to Geoffrey Hinton (considered by many the Godfather of AI), in the near future LLMs will be learning from each other. So it's more than "just the information available in the internet".

There's more...

In matter of days, multiple features are being added to Auto-GPT, either directly in the main repository, or in the format of Plugins.

The list of plugins include:

1. API Tools: Allows AutoGPT to make various API calls.
2. Baidu Search: Integrates Baidu search engines into Auto-GPT.
3. Bing Search: Integrates Bing search engines into Auto-GPT.
4. Bluesky: Enables Auto-GPT to retrieve posts from Bluesky and create new posts.
5. Email: Revolutionizes email management with AI automation for drafting and intelligent replies.
6. News Search: Integrates News Articles searches, using the NewsAPI aggregator into Auto-GPT.
7. Random Values: Enables AutoGPT to generate various random numbers and strings.
8. SceneX: Explores image storytelling beyond pixels with the Auto-GPT SceneX Plugin.
9. Twitter: Retrieves Twitter posts and related content via the Twitter platform using Tweepy.
10. Wikipedia Search: Allows AutoGPT to directly use Wikipedia.

Additionally, the text mentions several third-party plugins created by contributors, such as Alpaca-Trading, Crypto trading, Discord interaction, Google Analytics integration, Notion plugin, Weather plugin, YouTube features, and more. Each plugin serves specific purposes and provides functionality related to the mentioned platforms or services.

Conclusion:

Auto-GPT is representative of the pivotal moment we are living. While challenges and uncertainties remain, Auto-GPT showcases the potential of an AI agent to achieve autonomous task performance, with implications for various industries. Continued research, refinement, and ethical considerations will be vital to maximize the benefits of Auto-GPT and shape the path toward more advanced artificial general intelligence systems.