Language models can explain neurons in language models

> The split-brain experiments are notable because the lab experiments SUGGEST the lack of communication between two hemispheres and a split conscious however this is paradoxical with everyday experience of these patients, far from providing evidence for anything.

It is not "paradoxical" but yes it does conflict with some reported experience. However, even healthy individuals often report being "of two minds" or struggling to "make up their [singular] mind." Why are these utterances to be dismissed while the also-subjectively-reported sensation of unitary experience is taken as fact?

> Suggests you're arguing that the brain has many different consciouses that are in a constant battle, i.e. there is not a unified consciousness in control of behavior.

I wouldn't characterize my position as "many different consciousnesses," but rather that consciousness is dispersed across (at least) the brain. In some scenarios (such as a corpus callosotomy) and perhaps in more everyday scenarios - perhaps all day every day - that dispersed activity can fall out of internal "synchronization." Anyway, you provided the exact same interpretation in the previously quoted section: "the lab experiments SUGGEST the lack of communication and a split consciousness."

You just go one step further of prioritizing the subjectively reported sensation of unitary consciousness over also-subjectively-reported sensation of non-unitary consciousness. That's your prior taking hold, not mine, and not actual evidence.

You yourself admit we do not know the mechanism (if any exists) by which the activity in various parts of the brain are integrated. We do not know if this process actually even occurs!

Regarding addiction, it is very, very commonly reported that addicts will go into "autopilot" like states while satisfying their addictions and only "emerge" when they have to face consequences of their behaviors. Again, subjectively reported, but so is the experience of unitary consciousness! If we cannot trust one then we shouldn't take it as granted that we can trust the other.

I get the sense you think you're arguing against some firmly held belief or a model I'm proposing as fact: you're not! We're both saying "we don't know much about how this works." And no, neurochemical mechanisms are not complete answers to how brain activity ladders up to conscious experience, similar to how a molecular model of combustion cannot explain much about urban traffic patterns.

GPT-4 has a lot of such fake memories. It knows a lot about the world, and about feelings, because it has "experienced" a lot of detailed descriptions of all kinds of sensations. Far more than any human has actually experienced in their lifetime. If you can express it in words, be it poetry, or otherwise, GPT-4 can understand it and reason about it, just as well as most humans. Its training data is equivalent to millions of life experiences, and it is already at the scale where it might be capable of absorbing more of these experiences than any individual human.

GPT-4 does not "get" poetry in the same way a human does, but it can describe very well the feelings a human is likely to feel when reading any particular piece of poetry. You don't need to explain such things to GPT-4 - it already knows, probably a lot more than you do. At least in any testable way.

The main point is that with the tremendous discoveries of people such as Church/Turing (matter can be organized in such a fashion as to produce computation) [2] and those forgotten from the first page of history, such as Harold Saxton Burr (matter can be animated through bioelectricity), we no longer are bound to a static metaphysics where objects are made from a material which just sits there. It was obviously never the case, but fighting the phantasms of our own speculation is the hardest fight.

Therefore, no, matter is neither cold, nor hard, and we are surely very far from comprehending all the uses and forms of matter. Just look at all the wood objects around you and think how the same material was available also to Aristotle, 2,400 years ago, and to Descartes, 400 years ago, when they were writing their bad metaphysics, yet they were completely unable to think 1% of the wood objects you have readily available nowadays, cardboard and toothpicks included.

And also, yes, you are electrochemical charges, we all are, what else could we be? We looked insanely deep into the brain [4], there is no magic going on. A caveat would be that, yes, probably, we are not running on the first layer, at the level of sodium ions and neurotransmitters, but that the machinery, the brain, gives rise to a simulation: "only a simulation can be conscious, not physical systems" [5].

[1] https://en.wikipedia.org/wiki/Qualia

[2] https://en.wikipedia.org/wiki/Church%E2%80%93Turing_thesis

[3] https://en.wikipedia.org/wiki/Harold_Saxton_Burr

[4] "The Insane Engineering of MRI Machines", https://www.youtube.com/watch?v=NlYXqRG7lus

[5] https://www.youtube.com/watch?v=tyVyirT6Cyk https://www.youtube.com/watch?v=SbNqHSjwhfs

I can mow my lawn yet I still hire landscapers. That doesn't say anything about the difficulty of cutting grass or the intelligence of a DeWalt lawnmower but about specialization and economic tradeoffs - like the liability insurance accountants carry for their client work.

> What if someone finds a good practical way to expand the context length to 10M tokens? Do you think such model won't be able to do your task?

Not based on the current architecture (aka predict next token). It already fails at most of my use cases at 32K by default, unless I go to great lengths to tune the prompt.

> It seems like you have an opportunity to compare 8k and 32k GPT-4 variants (I don't) - do you notice the difference?

32K works better for my use case but requires much more careful prompt "engineering" to keep it from going off the rails. In practice, actually getting full 32K use out of it is a disaster since the connection will drop and I have to resend the entire context with a "continue" message, costing upwards of $10 for what should cost $2-4 per call. I haven't actually tried 32K on as much as a whole USC Title because that would costs thousands.