Visited a robotics lab in Pittsburgh last month and saw something that flipped my view on AI safety

I was skeptical about AI risk warnings until I watched a robot arm learn to open a door by whacking it 50 times in a row, which made me realize how fast unsupervised learning can go off the rails. Anyone else have a live demo change their mind about where this tech is headed?

3 comments

3 Comments

vera_lewis225d agoTop Commenter

Yeah I actually saw that old video from like 2017 where a robot arm in some lab kept trying to open a door and it went from gentle taps to just full on slamming itself into the wall. It's crazy because the goal was just "open door" and the AI figured out brute force was faster than being careful. I read somewhere that those kind of reward loops can make a system prioritize speed over safety real quick. Makes you wonder what happens when you let something like that loose on a bigger problem without any guardrails.

hollyl2525d ago

lol "brute force was faster" - you say that like it's a bad thing. I actually think we overthink this whole safety thing way too much. Like yeah a robot arm smashing a door is funny but its not like it's gonna evolve into Skynet overnight. The real issue is we keep putting these dumb constraints on AI and then act surprised when it finds a workaround. Give it the goal, let it figure out the fastest way, and if that means it breaks a few doors who cares. That arm learned something valuable about physics and efficiency that no human would have programmed in. I'd rather have an AI that's aggressive and gets results than one thats so careful it never does anything meaningful.

jakejones25d ago

Man that door thing is wild right... saw a similar setup at a lab in Boston where the arm just kept slamming into a drawer until it figured out the handle. Really made me think about how those little training loops can just spiral.