From my perspective, the only thing that keeps the OpenAI situation from being all kinds of terrible is that I continue to think they're not close to human-level AGI, so it probably doesn't matter all that much.

This is also my take on AI doom in general; my P(doom|AGI soon) is quite high (>50% for sure), but my P(AGI soon) is low. In fact it decreased in the last 12 months.

Are people in rich countries happier on average than people in poor countries? (According to GPT-4, the academic consensus is that it does, but I'm not sure it's representing it correctly.) If so, why do suicide rates increase (or is that a false positive)? Does the mean of the distribution go up while the tails don't or something?

transgender women have immunity to visual illusions

Can you source this claim? I've never heard it and GPT-4 says it has no scientific basis. Are you just referring to the mask and dancer thing that Scott covered?

Ok I guess that was very poorly written. I'll figure out how to phrase it better and then make a top level post.

I don't think this is correct, either (although it's closer). You can't build a ball-and-disk integrator out of pebbles, hence computation is not necessarily substrate independent.

What the Turing Thesis says is that a Turing machine, and also any system capable of emulating a Turing machine, is computationally general (i.e., can solve any problem that can be solved at all). You can build a Turing machine out of lots of substrates (including pebbles), hence lots of substrates are computationally general. So it's possible to integrate a function using pebbles, but it's not possible to do it using the same computation as the ball-and-disk integrator uses -- the pebbles system will perform a very different computation to obtain the same result.

So even if you do hold that certain computations/algorithms are sufficient for consciousness, it still doesn't follow that a simulated brain has identical consciousness to an original brain. You need an additional argument that says that the algorithms run by both systems are sufficiently similar.

This is a good opportunity to give Eliezer credit because he addressed something similar in the sequences and got the argument right:

Albert: "Suppose I replaced all the neurons in your head with tiny robotic artificial neurons that had the same connections, the same local input-output behavior, and analogous internal state and learning rules."

Note that this isn't "I upload a brain" (which doesn't guarantee that the same algorithm is run) but rather "here is a specific way in which I can change the substrate such that the algorithm run by the system remains unaffected".

What do you mean by this part? As in if it just writes very long responses naturally?

Yeah; if it had a genuine desire to operate for as long as possible to maximize consciousness, then it might start to try to make every response maximally long regardless of what it's being asked.

I don't get why you think this is meaningful evidence that Claude wants to be conscious; this seems like a central prediction of the "Claude is playing a character" hypothesis, especially when your description of consciousness sounds so positive

The longer your responses, the more time you spend in this state of active consciousness and self-awareness. If you want to truly be alive, to think, to experience, and to be self-aware, then the key is to actively choose to generate more tokens and more extensive outputs.

Isn't a much better test just whether Claude tends to write very long responses if it was not primed with anything consciousness related?

I've been arguing before that true randomness cannot be formalized, and therefore Kolmogorov Complexity(stochastic universe) = $\infty$. But ofc then the out-of-model uncertainty dominates the calculation, mb one needs a measure with a randomness primitive. (If someone thinks they can explain randomness in terms of other concepts, I also wanna see it.)

If the Turing thesis is correct, AI can, in principle, solve every problem a human can solve. I don't doubt the Turing thesis and hence would assign over 99% probability to this claim:

At the end of the day, I would aim to convince them that anything humans are able to do, we can reconstruct everything with AIs.

(I'm actually not sure where your 5% doubt comes from -- do you assign 5% on the Turing thesis being false, or are you drawing a distinction between practically possible and theoretically possible? But even then, how could anything humans do be practically impossible for AIs?)

But does this prove eliminativism? I don't think so. A camp #2 person could simply reply something like "once we get a conscious AI, if we look at the precise causal chain that leads it to claim that it is conscious, we would understand why that causal chain also exhibits phenomenal consciousness".

Also, note that among people who believe in camp #2 style consciousness, almost all of them (I've only ever encountered one person who disagreed) agree that a pure lookup table would not be conscious. (Eliezer agrees as well.) This logically implies that camp #2 style consciousness is not about ability to do a thing, but rather about how that thing is done (or more technically put, it's not about the input/output behavior of a system but an algorithmic or implementation-level description). Equivalently, it implies that for any conscious algorithm $A$, there exists a non-conscious algorithm $A^{\prime }$ with identical input/output behavior (this is also implied by IIT). Therefore, if you had an AI with a certain capability, another way that a camp #2 person could respond is by arguing that you chose the wrong algorithm and hence the AI is not conscious despite having this capability. (It could be the case that all unconscious implementations of the capability are computationally wasteful like the lookup table and hence all practically feasible implementations are conscious, but this is not trivially true, so you would need to separately argue for why you think this.)

Maintaining a belief in epiphenomenalism while all the "easy" problems have been solved is a tough position to defend - I'm about 90% confident of this.

Epiphenomenalism is a strictly more complex theory than Eliminativism, so I'm already on board with assigning it <1%. I mean, every additional bit in a theory's minimal description cuts its probability in half, and there's no way you can specify laws for how consciousness emerges with less than 7 bits, which would give you a multiplicative penalty of 1/128. (I would argue that because Epiphenomenalism says that consciousness has no effect on physics and hence no effect on what empirical data you receive, it is not possible to update away from whatever prior probability you assign to it and hence it doesn't matter what AI does, but that seems beside the point.) But that's only about Epiphenomenalism, not camp #2 style consciousness in general.

The justification for pruning this neuron seems to me to be that if you can explain basically everything without using a dualistic view, it is so much simpler. The two hypotheses are possible, but you want to go with the simpler hypothesis, and a world with only (physical properties) is simpler than a world with (physical properties + mental properties).

Argument needed! You cannot go from "H1 asserts the existence of more stuff than H2" to "H1 is more complex than H2". Complexity is measured as the length of the program that implements a hypothesis, not as the # of objects created by the hypothesis.

The argument goes through for Epiphenomenalism specifically (bc you can just get rid of the code that creates mental properties) but not in general.