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.