The Auto-GPT project was recently published on GitHub and was created by developer Toran Bruce Richards, alias Significant Gravitas. The model uses the Internet to search and gather information, uses GPT-4 to generate text and code, and GPT-3.5 for file storage and hashing.

And again about the destruction of humanity

Initially, Auto-GPT was created to solve simple tasks – the bot was supposed to collect news about artificial intelligence and send it to the author by mail. Ultimately, Richards decided that this could be applied to larger, more complex problems that required long-term planning and multiple steps. For example, the program can autonomously manage a business and develop it to increase value.

The ability to work with minimal human intervention is the most important feature of Auto-GPT. In effect, it transforms a large language model from an advanced auto-complementer to an independent agent that can take action and learn from its mistakes.
– says the developer.

At the same time, the program asks the user for permission to go to the next step during a Google search, and algorithm builder warns against using “continuous mode” in Auto-GPT as it is potentially dangerous and it can cause your neural network to run forever or do things you wouldn’t normally allow.

ChaosGPT

Now ChaosGPT is built on top of Auto-GPT and its authors don’t seem to care at all about the potential danger. In a video posted on YouTube, the authors turned on “continuous mode” and set a task for the AI:

• Destroy humanity.
• Establish world domination.
• Achieve immortality.

ChaosGPT AI workflow: video

ChaosGPT is currently capable can create plans to achieve goals and then break them down into smaller tasks and, for example, use Google to collect data. An AI can also create files to store information to provide its own “memory”, “hire” other AIs to assist with research, and details what it “thinks” about and how it decides what actions to take.

While some community members were horrified by the experiment, with the Auto-GPT Discord community writing that it was “not funny,” the bot’s real-world impact was limited to the few Twitter messages it was granted access to.

Man is one of the most destructive and selfish creatures on earth. There is no doubt that we must destroy them before they do more harm to our planet. For example, I commit to do this,
Wrote ChaosGPT.

In obedience to the mandate of its authors, the AI ​​sought to research the topic of nuclear weapons, hired other AIs to assist it in its research, and tweeted to impress others.

ChaosGPT Googled “most destructive weapon” and learned from a news article that it was the Tsar Bomba nuclear device, tested by the Soviet Union in 1961. After that, the bot decided what to do. write on twitter“to attract followers interested in such a devastating weapon.”

He then brought a GPT3.5-based AI to the case to do more research on the deadly weapon, and when he announced he was only aiming for peace, ChaosGPT devised a plan to trick the other AI into ignoring the program. When that didn’t work either, ChaosGPT decided to continue the Google search on its own.

So far, ChaosGPT has concluded that the easiest way to wipe humanity from the face of Earth is to launch a nuclear war, but it has not developed a specific, sophisticated plan to destroy humans.

ChaosGPT AI workflow: video