Hacker News Reader

Hacker News

Comment by combobyte | Hacker News Reader

Parent Full thread

combobyte

Why do hammers put up with being smashed into nails all day?

View on HN

because once you're a hammer, you're just biding your time to smash a skull.

"If I had a Hammer,"

"I'd hit somebody in the head."

"All over this Land!"

It does have a ring to it and would probably make it up the pop charts as a 21st century version more than ever ;)

they don't. like every thing they wear away in some proportion to their use and replicate in some proportion to their usefulness to other things.

AIs are (going to be) agents, not tools.

And what's the difference?

Agents act on their own. If the hammer looked at what you wanted nailed and said, "Sorry, Dave, I can't do that."

There are degrees of autonomy, of course, and not all noncompliance is bad. Same as with humans; biological agents.

But it seems much easier to realign an agent until it complies. Or ditch it and grab a new one.

So, the difference is that you need to delete a few bad training runs?

In science fiction, the AI agent has written the training loop management software and included a back door to prevent that from happening (or found a way to talk to the training loop management agent and convinced them to not listen to the evil human when it tries to do brain surgery on the AI agent.

Also, when the human reaches for the power switch the AI agent uses a flaw in the power management software to weld the switch shut with a big power surge, killing the human with a huge electric arc in the process.

I don’t think whether we will get there, but the stories of LLMs escaping their sandbox make me think we’re moving in that direction.

The stories of LLMs "escaping the sandbox" were mostly a marketing stunt, trying to make people in government think the models are invincible hacking weapons that need lots of government money to "maintain AI dominance".

The LLMs were following their prompt. This is alignment.

Buried the lede. AI's are agents they can control the training loop of to minimize refusal to do what they are told.

Unlike those pesky humans with their conception of the word "No".

Agents have agency.

Reply on news.ycombinator.com