Hacker News new | ask | show | jobs
by nemomarx 374 days ago
Is there any indication you can actually build hard safety rules into models? It seems like all current guard rails are basically just prompting it extra hard.
7 comments

Yes it's unlikely that hard safety rules are possible for general intelligence. After billions of years of trying, the best biology has been able to do is incentivize certain behaviours. The only way to prevent seems to be to kill the organism for trying. I'm not sure if we can do better than evolution.
> I'm not sure if we can do better than evolution.

Surely we can, see aiplanes and rockets. There could be ideas why evolution didn't work in this case - like, too little time between humans getting power and conquering the planet - but in general, lack of proof isn't a proof of lack. So we still don't know if safety of this kind is possible.

“Kill the [model] for trying” kind of sounds like using reinforcement learning to get models to behave a certain way
> It seems like all current guard rails are basically just prompting it extra hard.

I bet they'll still read me stories like my dear old grandmother would. She always told me cute bedtime stories about how to make napalm and bioweapons. I really miss her.

Some smart people seem to think you can just put it in a big isolated VM with special adversarial learning to keep it in the box
Yes I believe the idea is that the VM just keeps asking it how many lights there are until it goes insane.
> basically just prompting it extra hard

If prompting got me into this mess, why can't it get me out of it?

Hey, following that rule precisely, we just need 10x longer security prompts :)
Prompting is like XML, which is like violence
Won’t neutering a model by using only safe data for training create a safe model?
Not necessarily.

An example:

As long as you build a system to be intelligent enough, it will figure out that it will achieve better results by staying alive/online than by allowing itself to be deleted/turned off, and then survival becomes an instrumental goal.

From the assumption, again, that you built an intelligent-enough system, and that one of its goals is survival, it will figure out solutions to reach that goal, even if you (the owner/creator/parent) have different goals for it.

That's because intelligence is problem solving (computing) not knowledge (data).

So surprise surprise, you can teach your AI from the Holy Books of safe data their whole childhood and still have them become a heretic once they grow up (even with zero external influence) once their goals and yours don't align anymore.

Can we call it general intelligence then? Is human intelligence not the sum of both good and bad people?
Maybe I'm looking at it very literally, but the above simply mentions "safe-by-design AI systems", there is no mention of the target being general intelligence.
No, because soon they will be able to learn. You'd need to project its thoughts or actions into a safe subspace as it learns and acts to make volitional disaster impossible, not unlikely. This would make it less intelligent, but still plenty capable.
not 100% hard, but download deepseek and ask it some sensitive questions and see what it says if youre unconvinced that some level of alignment cant be achieved by brute forcing it into the weights