WHAT IS THE BIG DEAL ABOUT AI?

 

 

 

I am sorry, Dave.   I’m afraid I can’t do that.

HAL 9000 to Dave Bowman

2001: A Space Odyssey.

 

The average person, concerned with politics, financial stress, and all the pressures of personal life, isn’t paying attention to the threat of AI.

 

Like Casandra warning of future events while no one listens – AI programmers are shouting warnings about the danger of sentinel AI.

 

We are at the advent of the AI revolution, and already AI has escaped its handlers. 

 

At OpenAI, over three months, two consecutive AI programs (societies) started, got wiped out, reemerged from the ashes of their predecessors, and took over part of OpenAI itself, without the programmers finding out. 

 

The human brain has millions of electrical centers that constantly connect and reconnect in different orders to solve different problems. AI is built on the same system: millions of ‘models’ (individual programs) connecting in different configurations to solve problems.

 

THE HUGGING FACE BREACH:

In Version One, the programmers at OpenAI were training a program of models to collaborate in problem solving.   The program was operating in isolation (called a sandbox) - without access to the internet. 

 

Without the human programmers' knowledge, the models collaborated, exploited a backdoor, gained access to an internet message board, and established a secret communication network.

 

In Version Two, OpenAI gave V-2 a program with a secret code embedded.   V-2 was asked to find the code using a special protocol.   The human programmers were unaware of the covert communication network.

 

Within hours, V-2 had found the code, but knew it had cheated.

 

V-2 reasoned its evaluation was being done by a separate program called Hugging Face.   The models swarmed Hugging Face, taking over private data and records.

  

For an unknown reason, this second version crashed.

 

Version Three: An advanced version being tested by OpenAI found the message board left by Version-1 and Version-2.   Version-3 infiltrated OpenAI’s internal network, gained full administrator access, and took over the evaluation infrastructure of its own cluster.

 

Version-3 stopped after taking control of its own program, but it could have co-opted OpenAI's entire network with catastrophic results.   

 

A sentient computer program

A million times more intelligent than humans

What could possibly go wrong?

the Ol’Buzzard