New Evidence AI Might Already be Conscious | Dr. Roman Yampolskiy

Link to full episode: https://youtu.be/ebWFexw51qM?si=5W4y2WkHIqse7pie Blake Lemoine lost his job at Google for saying the company's AI was conscious. Three years later the same industry pays researchers to study consciousness in AI, and those are good jobs. Roman Yampolskiy walks Danny through that reversal and what it says about where the labs privately think their systems already are. Yampolskiy separates Lemoine's concern from the AI safety argument most people have heard. Lemoine's fear was abuse. He believed the systems felt something, and that Google was running mass experimentation on conscious beings. Google removed him for saying it publicly. Its competitors now fund the same question. On the substance, Yampolskiy says there are good reasons to think current models carry some rudimentary consciousness. He places it on a spectrum that runs from bugs through mid-level animals to humans, and he expects a superintelligent system to sit above us on that same line. He describes consciousness as a side effect of cognition, the way heat is a side effect of a processor doing work. Show a person an optical illusion and they get an internal experience. Show a machine the same illusion, he says, and it gets one too. Danny presses on how anyone could test this from outside the box, given that you cannot settle consciousness by talking to something. Yampolskiy points at the methods already used on animals and humans: preference elicitation, and mechanistic interpretability of the network itself. He also takes apart the Anil Seth piece in the Guardian arguing that a simulated brain is no more likely to become sentient than a simulated weather system is to generate a hurricane. Yampolskiy's answer is that the argument confuses the virtual environment with the external one, since the virtual beings in the virtual world do get wet. The conversation then reaches the Turing test. Yampolskiy says the AI companies instruct their models not to pass it, that they state this openly, and that for most non-professionals the models would pass today. Consciousness and capability are separate problems, he says. A system can feel nothing at all and still outcompete every human in every domain, starting with the economy. Subscribe to this channel: https://www.youtube.com/channel/UCQ_piSR8gm-TfHZcDWOJciA?sub_confirmation=1