Eliezer S. Yudkowsky <sentience@pobox.com> On September 22, 1999 Wrote:
>The only reason I believe them [Penrose & Hameroff] at
>all, is because quantum coherence is a probable prerequisite for any
>number of ineffabilities, not just the ones they propose.
I don't know exactly what ineffabilities they propose because all that they say boils down to just "There's magic going on in the microtubules". There is no way to prove them wrong because they don't tell us what we're supposed to find, they could shrug off a thousand years of negative results and say "you just haven't looked hard enough yet". That's not science.
By the way, microtubules aren't limited to neurons, every cell in your body has them, plants have them too. Are plants conscious? When I slice a tomato is there a silent scream?
>I'm not sure I credit the stochastic dogma of neural nets. It's such a
>tremendous amount of wasted computational power. It seems to me, on
>purely evolutionary grounds, that every little quaver will be exploited.
Evolution doesn't need to come up with the perfect design, it just has to be better than the competition. In addition, even if an organic quantum computer were possible there is no reason to think evolution could get there from here.
A few of reasons for natures poor design, the most important is the last.
Once a standard is set, with all its interlocking mechanisms it's very difficult to abandon it completely, even when much better methods are found. That's why we still have inches and yards even though the metric system is clearly superior. That's why we still have DOS and QWERTY. Nature is enormously conservative, it may add new things but it doesn't abandon the old because the intermediate stages must also work, sort of like Microsoft. That's also why we have all the old brain structures that lizards have as well as new ones.
3) Lack of Genetic Variation: Mutation are random and you might not get the
mutation you need when you need it. Feathers work better for flight than the skin flaps bats use, but bats never produced the right mutations for feathers and skin flaps are good enough.
4) Constraints of Costs and Materials: Life is a tangle of trade offs and
compromises.
5) An Advantage on one Level is a Disadvantage on Another: One gene
can give you resistance to malaria, a second identical gene will give you sickle cell anemia.
6) Evolution has no foresight: This is the most important reason of all.
A jet engine works better than a prop engine in an airplane. I give you a prop engine and tell you to turn it into a jet, but you must do it while the engine is running, you must do it in one million small steps, and you must do it so every one of those steps improves the operation of the engine. Eventually you would get an improved engine of some sort, but it wouldn't look anything like a jet. You just can't get to there from here.
If the tire on your car is getting worn you can take it off and put a new one on, but evolution could never do something like that, because when you take the old tire off you have temporally made things worse, now you have no tire at all. With evolution every step (generation), no matter how many, must be an immediate improvement over the previous one. it can't think more than one step ahead, it doesn't understand one step backward two steps forward.
>Neurons are not remotely the simplified little blobs of "neural nets",
>or the flashes of excitation we record.
I've heard that said a billion times and I think it's an exaggeration. I hate exaggeration.
Sure neurons are ridiculously complex, but nearly all of it involves the basic metabolism of staying alive not signal processing.
>If you can establish that a subtle effect would establish a lot of 5-QB
>computations all over the brain, with a tweak that comes with no major
>evolutionary costs, that'd be enough to win the day for ineffability.
What's the point in that? If you're a 1000 Qbit quantum computer then you're God, if you're a network of a billion 5 Qbit quantum computers then you're a worm.
>In particular, who needs explicit qubits?
We do in order to talk about such things.
>>Me:
>>I'm a neural net and I'm pretty good at predicting what another neural
>>net that inputs handwriting and outputs ASCII will do.
>That's because you and the neural net are optimized for the same task.
I wasn't born that way, I optimized myself to do it.
>Can you think of any NP-complete tasks that people are optimized for?
I've tried to optimize myself to solve non computable tasks but it hasn't worked worth a damn. If I'm a neural net then my failure is easy to explain, if I'm full of weird mysterious ineffable stuff then it's a mystery.
>Remember, I started out as a Strong AIer. I still am a Strong AIer at
>heart, just one who's been forced to believe in noncomputability.
What forced you? It seem like a strange fit for someone who wants to program a AI. Personally I think it unnecessarily confuses an already confusing situation. I don't need that hypothesis, it serves no function.
John K Clark jonkc@att.net