Quick Introduction We have all been on forums, chats, reddit, discord, youtube, or somewhere and heard “Oh! Model XYZ is AMAZEBALLZ!zomgwtfbbq” then downloaded it (or more likely, some quantized form of it) and said “eww… This sucks!” This post is going to be a rather technical series of experiments to demonstrate the impact of implementation-specific hazards with inference. I will be using the term “reference implementation” to describe the lab that published and offers first-party hosting of ...
Not with agents and chain of thought. Agents can run for hours continuously. So sure, an AI agent is not the same as a sentient human on the time scale of a year, but on the time scale of a few hours, perhaps the AI is “sentient”.
The first paper you linked is focused on the mechanism and not the output. It creates a definition of “thought” and then talks about how the AI doesn’t “think in the feature space”. It only addresses chain of thought at the end, and says that the issue is that it constrains AI to think in natural language only (instead of, say, pictures).
But again, why would thinking in pictures define sentience? The paper gives a rock-paper-scissors example and says that the AI thinks about it in a different way than a human. So what? If a human plays rock paper scissors against the AI and the AI’s output is indistinguishable from a human’s, 99.9% of the time, why is that not sentient?
This is like saying python programmers are not real programmers. If a python programmer can implement the same program in python, who cares what language they use.
I’d love for a better definition of sentience than “it works differently than humans”.
Them
You
Those are not the same concepts. The model of LLM remain unchanged. Agents running for hours just modify the prompt they input into an LLM model.
I do not care about the rest of discussion, just swoop in to clarify that distinction.
Fair enough. When I say that AI might be sentient, I include systems like “LLM + agent harness”, treating the whole system as AI.