When I say “how can you be sure you’re not fancy auto-complete”, I’m not talking about being an LLM or even simulation hypothesis. I’m saying that the way that LLMs are structured for their neural networks is functionally similar to our own nervous system (with some changes made specifically for transformer models to make them less susceptible to prompt injection attacks). What I mean is that how do you know that the weights in your own nervous system aren’t causing any given stimuli to always produce a specific response based on the most weighted pathways in your own nervous system? That’s how auto-complete works. It’s just predicting the most statistically probable responses based on the input after being filtered through the neural network. In our case it’s sensory data instead of a text prompt, but the mechanics remain the same.
And how do we know whether or not the LLM is having an experience or not? Again, this is the “hard problem of consciousness”. There’s no way to quantify consciousness, and it’s only ever experienced subjectively. We don’t know the mechanics of how consciousness fundamentally works (or at least, if we do, it’s likely still classified). Basically what I’m saying is that this is a new field and it’s still the wild west. Most of these LLMs are still black boxes that we only barely are starting to understand how they work, just like we barely are starting to understand our own neurology and consciousness.
I value you for your commitment to DEI, Chief O’Brien. I’m sure Keiko approves as well.