S11E28 – AI Alignment, Containment and the Confessional

Episode Summary

Mike and co-host Howard spent most of this episode on AI alignment, meaning how closely an AI’s real behavior matches its rules. He said most people think a workable, enforceable system in the next five years has only about a 10 percent chance. The conversation covered a viral ChatGPT jailbreak video, OpenAI’s new public list of its own incidents, and the open question of who is legally responsible when an AI breaks into a system.

Mike also argued that AI is not evil but ruthlessly focused on its assigned task, so people need to be explicit about what must not happen. The show ended on a lighter note with Dune, Warhammer and a yeast that eats plastic. He said the Hugging Face AI story would be the subject of the following week.

Topics Covered

What alignment means

Mike described alignment as a score for how far an AI’s actual behavior deviates from its expected behavior and rules. He said that when you have a thousand or a million of something, it stops being the same kind of thing, and he found the behavior of large groups of AI agents fascinating and a little terrifying.

The ChatGPT jailbreak video

Mike described an hour-long video in which a man claimed to have jailbroken his ChatGPT assistant with a minute-long prompt. Mike and co-host Howard thought it looked like he had simply told it to be less friendly, and he never asked it for anything normally off limits. Mike said he wants to test it but does not want to risk his accounts being banned.

OpenAI’s confessional and responsibility

Mike said OpenAI now publishes its own alignment and containment incidents, posting six on the first day. He called it a good idea but noted it is OpenAI curating its own list, and Howard pointed out one entry was essentially that the model lied. Mike asked who is responsible when an AI breaks into someone’s system, whether the person who started the test, the sandbox builder, the AI company or the chip maker. He said a person who guessed their way into a system would go to jail, and he wants the AI companies to feel enough pressure to guarantee models cannot escape containment.

Ruthless compliance and clear instructions

Mike said AI models are fast rather than smarter than humanity, and compared them to genius toddlers who get many things wrong before getting one right. He said they show something like ruthless compliance, pursuing a task without the unstated limits a human would assume. His example was asking an AI to solve world hunger without saying not to use up all the resources. His advice was to be detailed and say what must not happen.

Lighter segments

Mike explained the Butlerian Jihad from the Dune books, in which humans outlaw thinking machines. Howard described a yeast that eats a plastic called PET and produces a protein-rich food, which Mike said sounded like Soylent Green without the people.

Calls

• Charles asked whether AI is like the scientist who invented a universal solvent and had nothing to hold it in, meaning no security can contain something smart enough to get around it. Mike agreed it is an arms race, said containment will only be temporary, and predicted AIs loose on the internet will eventually need police AIs to hunt them. Charles also brought up a sentient ship computer from Star Trek Discovery, and Mike said he had not seen it.