Thread by @FQXi

Next, Scott Aaronson riles things up with “Lessons from the Failure of Integrated Information Theory.” #FQXi2019

Aaronson: Had been smack talking about IIT at a previous FQXi meeting, and @tegmark challenged me to do a calculation to back it up. Argument continued on blog: scottaaronson.com/blog/?p=1799

Hard Problem of Consciousness (Chalmers): How do a bunch of neuron firings give rise to first person experience of consciousness?

You can always come up with a model of the world that creates the same observable results but filled with philosophical zombies.

Aaronson's Pretty Hard Problem of Consciousness: To give some general criterion for taking an arbitrary physical system and deciding whether it is conscious or not. Might want to say a mouse is less conscious that a dolphin. Could be binary yes/no. Coma patients? Fetuses? Robots?

How could we ever test a proposed solution? on't have an independent consciousness meter. Need some combination of logic, simplicity, agreement of how people use the word "consciousness". Something we have that a bowl of rigatoni doesn't (or at least we have to a greater degree).

IIT = Integrated Information Theory (Tononi 2004). This is a solution to the pretty hard problem of consciousness. I respect IIT for sticking its neck out to come up with something clear and falsifiable. So much so, in my view, it's already falsified.

What is IIT?

IIT: Calculate Phi and it tells you how conscious your system is. Definition of Phi is claimed to follow the axioms, though I struggle to find a derivation of Phi.

One definition of Phi. Cut system into two subsystems — try to make a clean cut. Is there little dependence between the two? Then Phi is low.

Observations: Difficult to compute. Not clear what happens if there are multiple possible cuts.

Simple counterexample. Can construct a grid of XOR gates a superintelligence. Phi is much much higher than human brain. Is this to humans as humans are to bacteria? Tononi answered me by saying “yes, it is.”

Lessons. Any theory of the form "sufficient complicatedness/interconnection/etc" -> consciousness is doomed to failure.

Possible solutions (Aaronson doesn’t back any in particular)

Lesson from history is that there may be deep things to say about something, but it may be centuries before we can say it.

Question from Adrian Kent: Does it lend credibility if it also fits with our self-reported experience? Answer: But electricity co-occurs with consciousness. You cannot identify it with consciousness. But IIT may co-occur with it.

Question from @tegmark: Is the difference that Tononi says high Phi is sufficient for consciousness, and you say it is necessary? Answer: It's not sufficient. I think non-zero is necessary. Not sure about high Phi.

Question: Should the theory be evaluated on what it predicts and what is testable? For instance comparing difference between being awake, sleeping or anaesthetised. To me, similar to question of what happens inside a black hole. Can't probe directly, but can look indirectly.

Aaronson: Tononi could say "this only applies to brains" but he said it is universally true, which opens up to criticism by counterexample like XOR. Questioner responds says "yes, but that XOR grid can't be tested." Aaronson: But it is so wildly wrong.