Recent research points towards the nature of LLMs as being simply very nice compression machines… with a trained model like GPT (3.5, e.g.) being, essentially, a compression of vast quantities of human knowledge.
With this perspective, we can see more clearly how a language model, on its own, is not a consciousness, but rather a type of memory that can be used to predict or infer appropriate data/information decompression. Granted, with massive billion-parameter compression, we might find that not just facts and connection of facts is being compressed, but also meta-properties of such facts allowing for a type of “understanding” in the decompression that is already super-human. But in the end of a day, we would still not consider such a system conscious, or even a mind.
So let’s consider additional properties and architectures that would be required for us to consider a system to be a conscious mind. For fun, I’m going to call our imaginary synthetic conscious mind “ISCM”. I have a few thoughts:
- Continuous input - like us, ISCM would require continuous relevant inputs to build an internalized belief state about the world.
- Belief state - Like us, ISCM can’t know anything with absolute certainty. “Facts” are proven unfactual all the time, and even our basic facts (this blanket is red) stand on colloquial usage of language. ISCM might assign “confidence” to it’s beliefs, holding some more crucial than others.
- Goals - For ISCM to have agency, it must have goals. The goals would be interrelated.
- Planning - ISCM must be able to make plans to reach its goals based on it’s current understanding of the world.
- Act - ISCM would need to effect things to reach its goals. Even if purely embodied, ISCM must be able to take actions: “change its mind”, record new facts, research, plan, etc. Being able to effect real-world digital or physical change to reach goals would require additional non-mental actions like computer-system plugins (writing emails, looking up the weather, booking tickets, etc.) and actuators (move to X, cool down the room, dispense medication, etc.).
- Temporality - I broke this out as a separate property because it is a core assumption of all of the above. Inputs happen at specific times, goals and planning are sequential and time-bound, acting–being imperfect–requires feedback to determine whether the action resulted in the intended outcome. Our current generation of LLMs understand time semantically equivalent to any other property, but conscious minds which interact with the world need a much more profound integration.
What’s not on this list?
- Embodiment - Some theories say consciousness and mind cannot exist outside of a physical body. It’s clear a consciousness that has no interconnection with the physical world would be quite limited, it also seems clear that our other properties of Continuous Input and Belief State can encompass this idea–better input (sensations, perceptions) and better beliefs (judgements, feelings, intuition) would make a more capable ISCM. Indeed, we could limit ISCM’s inputs to a single “body of sensors” like an android, but we might also give it the input of multiple independent physical forms.
- Feeling - Another way to say “inputs” and “goals” (e.g., fight-or-flight)
- Intuition - A more sophisticated form of belief state management that includes meta-analysis and pattern matching.
What else? 