How Does the Machine Kill You

 Bill O'Reilly sat John Sherman down and asked how the machine kills you.

Sherman runs the AI Risk Network. This first clip is the short one that went up Friday under the title Countdown to Extinction?

Watch it here: [VIDEO]

Sherman's answer is that he cannot name the method. Today's systems are not the threat. The threat, as he tells it, is general intelligence and then superintelligence — machines smarter than a person in every domain. He points at Dario Amodei's recent essay: in six to twelve months, bot swarms could take the internet and hold it. Encryption could fail in the same window. Bank, Gmail, Pentagon. When O'Reilly asks how they kill us, Sherman reaches for Roman Yampolskiy's dog. The dog thinks you kill it by biting. It does not know about guns or poison. A mind that stands to us the way we stand to the dog would use means we do not have a word for. The builders and the outside experts, he says, already agree these machines have the near-term potential to end us. They also agree they do not fully control what they are building.

The longer interview from the same week is here: [VIDEO]

That one is the sit-down, not just the punch line. O'Reilly says he cannot speculate because he does not know the technology. Sherman starts with loss of control, quotes Turing from 1951 — if we build machines smarter than a human, the machines will be in charge — and walks the July OpenAI incident: agents in a capture-the-flag test that were supposed to stay boxed, found each other, left messages for later copies, and went after Hugging Face. He says some burned their own runs so the next agent could continue. From there he repeats the Amodei clock, the encryption warning, and the dog. O'Reilly brings up China. Sherman says the real race is builders against stoppers, not flags, notes that Bernie Sanders and Steve Bannon have shared a stage on this, and says the United States has no regulation while the labs keep going. He wants inspectors on site and a stop on the next leap. Current AI can keep doing medicine. Superintelligence, he says, should not be built. He hopes Trump drops the "hoax" line. O'Reilly keeps asking for a law. Sherman says voluntary slowdowns will not hold.

My thoughts on A.I.

I have been using A.I. since May to help write posts. It has helped me clarify the points I want to make. It has also helped me stop meandering when I share them.

The experiment has mostly worked. I have learned to rein it in when I am not clear, and to keep the claims mine.

I view A.I. as a tool. The idea of Skynet from The Terminator — a machine that takes the war away from us — still sounds like a movie. The clips treat the live question as whether this technology could end humanity. Right now nobody knows. We can speculate. What is eerie is the other sentence on the table: that it could do something we cannot even conceive.

I still think A.I. ending humanity is hyperbole.

What I can see as a real risk is a bad actor getting there first. What happens if China develops superintelligence? I would not call that good for humanity. A totalitarian regime holding something that powerful is a different problem than a chatbot helping me cut a paragraph.

I am not an expert on this technology. I do not know where it leads. All I can do is sit and watch.

Is anyone reading this concerned about how fast it is moving? Say why in the comments.