Comment by grim_io
6 hours ago
Exactly, like any AI tool ever.
Someone wrote some instructions. No agent harness ever simply decided to pursue its own interests.
6 hours ago
Exactly, like any AI tool ever.
Someone wrote some instructions. No agent harness ever simply decided to pursue its own interests.
How will you know when that happens? Or are you defining interests so narrowly that it's definitionally impossible?
If you are arguing that there is consciousness that's emerging from LLM's, I have to disagree on that.
We will know when we see it. I don't see it right now.
No. I'm not saying anything about consciousness.
Does a gradient descent algorithm pursue its interest of minimizing error? Does a home automation assistant pursue its interests when it sets my thermostat? I'm not super interested in the definition of "consciousness" or "interests". However, a thermostat setpoint has effects that are visible in the real world. That's a thing that happened, regardless whether you consider it to have happened in "the pursuit of an interest".
I'm saying that LLMs are affecting the world. And sometimes those effects might be difficult or impossible to trace back to a particular prompt written by a particular human. Chatbot input and output doesn't have to be in the form of text i/o. You can put them in a for loop. Remember OpenClaw?
> We will know when we see it. I don't see it right now.
There might exist an incentive to make it hard to see.