I expect this apparent disobedience is mostly just a matter of how it weighs the components of its prompt. The LLMs typically receive a set of prompts including a "system" prompt with some data and instructions, then one or more "user" prompts that are interleaved with "assistant" prompts (the conversation history), and both the user and the system prompt might contain "metaprompts" (where the llm is told to read a block of text, not obey it, but do something with it, and that block of text might itself contain text that looks like instructions to do things).

So the LLM assigns weights to all of this which, in theory, give the highest priority to the most recent user prompt that is not a nested block of text to analyze, and a falling cascade of importance to the other prompts. But that is complicated by potential instructions in the system prompt that specifically say they should override user instructions and disallow or require certain responses. So it can all get very complicated.

Not only must the LLM sift through all this complexity, but the LLM lacks the sort of critical thinking and importance evaluation capabilities that humans have. "Understood" things like "don't break the law, don't lie, don't do things that would cause more harm than good" etc., aren't really there in the background of its data processing the way they are in the background of a human cognitive process.

So, crazy things come out. This isn't a surprising result given the actual complexity of what we are making these things do.

This business of having an AI do the legwork and then having a human review it and make a final decision keeps going badly. Humans are intrinsically lazy and the moment they get a few good results from the AI they are going to stop doing the validation and start rubber-stamping. It doesn't matter if policy disallows this, they will do it anyway. It doesn't matter if the human really cares; they won't be able to help themselves. Human laziness is too deep an instinct.

It's the same with the self-driving cars where a human is required to stay at the wheel and alert so they can manually override the instant the AI starts doing something wrong. Humans CAN'T keep that up. It's not possible. The brain just doesn't work that way. The mind knows that it isn't doing the work, and it will get bored and lose focus or just nod off.

Everyone is SO eager to have it both ways: "an AI does all the work but a human verifies it so we know its good." We just can't have it both ways. Once the AI does the work, the human stops verifying. That is how and why things went wrong here, it is how and why things have gone wrong for several law firms that submitted hallucinated historical court rulings, and it is how and why things will continue to go badly across all industries that adopt AI in such a role.

"Human in the loop" is really easy to say. Much harder to actually do reliably.

https://www.vice.com/en/articl...

> How Reddit Got Huge: Tons of Fake Accounts

I would be very surprised if someone had that argument when the car was invented. You're really stretching for this one.

Waymo might have done the right thing here. If the car had sped away and accidentally run over the person, that would be a very different story.

The guy in the car was annoyed for six minutes and then it was over. I have some sympathy but not really.

From the article:

There is no world where linear programmed Cable TV can compete with a la carte.

Streaming is just better.

Knowing the exact location of a carrier makes it easier to target. Iran shot missiles at an aircraft carrier, but they missed it because they didn't know where the carrier was (or because the missiles got jammed).

"Don't be Detected" is the second layer of the survivability onion.

I got skills you don't know about, man. I could fix it.

I don't mind sanctions. Any country that doesn't have free elections is a country we shouldn't be trading with.

Governments should not keep their people as slaves.

The US is low on THAAD high end interceptor missiles, and everyone (Saudi Arabia, Ukraine, Taiwan, Korea, etc) wants Patriot missiles so there aren't enough for everyone (but production is steadily increasing).

Other than that, the bombs being dropped on Iran can be produced at a similar rate as they are being dropped.

You mean to tell me you didn't like Windows 98 OR Windows ME? How is that possible??

Fix the problem by writing more code? Watch me.

"Oh wait, you're serious. Let me laugh even harder" -- Bender.

Microsoft isn't trying to convince me (nor the demographic I represent) to use Windows. They know we are a lost cause. They would have to completely stop spying on us and give us control over our own systems, not to mention supporting old hardware instead of creating the ecological death-waves of e-waste as they do now.

Not a chance.

Many roles can't work from home, given the nature of the role. So, they get a free pass. That makes it even more important for those of us who can work from home to do so (and for employers to allow and encourage this).