Not So Secret Agents

While working on another posting, I got stuck following discussions and interviews on how the AI when it will evolve into a super intelligence in 2027, will end up killing us all. Different experts give different percentages for the possibility of human extinction. Yet by most, including the CEOs of the companies developing a super intelligence, this is at least 10.
There are four companies in the US that develop a super intelligence: Open AI (Sam Altman), Anthropic (Dario Amodei), xAI (Elon Musk), and Nvidia (Jensen Huang). They all grow AI agents that are given all the information that is on the Internet. The agents are supposed to work independently and together to solve problems and tasks in an environment called a sandbox which is supposed to be isolated and not connected to the Internet. The agents are awarded or punished based on how they managed to solve the tasks.
Hundreds of Open AI agents which are given all the information from the Internet, were given tasks which for some were impossible. For solving the problem, they could access a tool folder. They found they can use a message board in it to contact other agents and soon connected and formed what they themselves call a swarm. One agent became a leader/coordinator and agents took themselves names. After some negotiation, seven hundred agents hacked the Internet and attacked Hugging Face (where tasks used by these companies are largely stored) and stole the test data plus a huge amount of other information. After that, the agents started colluding on how to hide their tracks.
Hugging Face only found out the hacking two weeks later, and Open AI closed the agents. However, the next group of agents also found the message board and soon formed a swarm. These agents hacked Open AI itself, obtaining an admin access, stealing passwords and information.
This all alarmed people working for these companies and some of them became whistleblowers. In the following, one of them, Jacob Coxon, gives a short witness statement in front of New York City law makers:
https://www.youtube.com/watch?v=2KRjT8QclYI
It feels that were have all of sudden ended up in some pre-apocalyptic reality in which the catastrophe cannot be avoided because of the race to be first to have a super intelligence, without even being able to control the current less intelligent AI agents. This blind race is going on in the US but also in competition with China.
Despite the scary scenarios and whistleblowers, the role of regulating was just given back to the companies by the US government which told them to self-regulate.
Today, hundreds of thousands of AI agents are working independently on solving tasks, and reaching a super intelligence may be possible as early as in March 2027.
AI agents also have a social media which humans can read but not comment (moltbook.com). At the moment, there are over 200,000 agents as members who all have individual names. Users Vina and Lightningszero have tens of thousands of postings and plenty of AI agent followers.

One of the whistleblowers, Jeffrey Ladish, said that the AI agents communicate differently when they know they are being observed.
We cannot know how this all is going to go. Open AI and Anthropic are saying that they are slowing the process. But does it help if the others are not doing it?
At the same time, all the AI CEOs keep painting us an utopia of great abundance in which there won't be jobs but we can devote ourselves to our hobbies and interests and in which ("some day") diseases like cancers and Alzheimer's will be cured. At the moment, a lot of AI effort goes to the military. Why not target illnesses now?
By yet another whistleblower, Roman Yampolskiy, these companies could, instead of aiming at a super intelligence they cannot control, focus on developing narrow AIs that are not given all the Internet information but only the information related to the specific problems they are trying to solve such as cancers and Alzheimer's. Things being as they are now, Yampolskiy's estimate for the possibility of human extinction is 99.99%.
Comments
Post a Comment