Hacker News Reader

Comment by naasking | Hacker News Reader

> If you built an android that had every if-statement imaginable coded into it so that it behaved exactly like a human

Except this robot is inconceivable, because it would exceed the size of the universe. The conscious android becomes conceivable when you say its on the human scale and yet still exhibits the same behaviours.

> Another issue is the purposelessness of consciousness. Why does there have to be experience tied to our reactions to stimuli, if the resulting behaviors are identical. Isn't there quite a bit of wasted energy and skull space that could be better used for more survival oriented if-statements. All of our behavior seems possible without the associated feelings.

This theory does explain the purpose of consciousness and qualia, as this poster explains:

https://news.ycombinator.com/item?id=11853240

View on HN

>Except this robot is inconceivable, because it would exceed the size of the universe.

Actually doesn't have to exceed anything at all.

Parent described it as being able to react to every possible situation ever, but you can get the same effects by hard-coding reactions to just a few millions/billions of possible events/discussions that can happen inside a room (and with coarse definition of distinct events), and just let some person with it in a room for 1-2 hours.

Heck, people used to feel Elisa was real...

Ok, ask it "If Kim Kardashian was debating Bernie Sanders on whether machines can think, how do you think that conversation would go?" I bet that wasn't on your pre prepared question list.

No, but it can always answer: "I don't really care for the Kardasians" or "This is a silly question" -- same as I would have in real life actually.

> Parent described it as being able to react to every possible situation ever, but you can get the same effects by hard-coding reactions to just a few millions/billions of possible events/discussions that can happen inside a room (and with coarse definition of distinct events), and just let some person with it in a room for 1-2 hours.

Honestly, I've always found any such construct absolutely trivial to identify. I've been able to stump every chatbot with my first question. No agent that doesn't actually understand would be able to pass a Turing test issued by a trained computer scientist, even if limited to 5 questions.

Would a person with Aspergers/autism/etc pass that test?

If not, how could you tell them from the "non understanding" robot?

I also think most robots would pass the "I couldn't tell its a robot" test with flying colors if they responded just like real people in the same settings as everyday conversations -- because those interactions are usually limited in scope, and you don't purposefully try to unmask those people with tricky questions.

> Would a person with Aspergers/autism/etc pass that test?

Sure, my first question is always simple direct semantics test. Perhaps something like, what's the third word of this sentence if you start counting from the second word? Mitsuku chatbot that won the 2013 chatbot competition got this one right, but if I just slightly rephrased it to, "If I were to ask you what the third word of this sentence is if you started counting from the second word, what would you say?", it got it flat wrong despite having just answered it.

Reflective questions like this are good tests of understanding that are trivial for a person to answer, not so trivial for a system that doesn't truly understand what it's reading.

Surely there is a mistake when you say such an android would exceed the size of the universe, right? Have you done any calculation or can you cite any source which remotely claims this? Im interested.

The English language consists of over 1,000,000 words. Let's suppose only 200,000 in common use, 1/7 are verbs, 1/2 are nouns. Even if you consider only simple sentences consisting of 3 words with subject-verb-object, the number of permutations of subject-verb-object sentences that an English-speaking human can utter are 28,000 * 100,000 * 28,000 = ~78.5 trillion = ~10^13. This doesn't include adjectives, adverbs, conjunctions or anything more complicated we see in real grammars.

Now consider that this sentence you're reading is 33 words long, so the number of permutations of such sentences is hundreds of orders of magnitude larger, on the order of 10^10000.

The visible universe is estimated to contain 10^80 particles, so even if you could encode a logic gate that could recognize a word in a single particle, it seems you're still just a tad short if you're looking to hard-code intelligent responses.

True understanding consists of building a model to generate new facts as needed, amounting to an impressive compression of facts you've already seen.

"Blue time smell rainbow sleep die carrot." Like the vast majority of sentences don't parse into anything.

Also, you're assuming zero preprocessing and the vast majority of input meaningful input is reducible due to patterns. "The great big elephant, almost stepped on me." You separate each chunk and deal with them independently. Thus replace "The great big elephant" with Bob. "Bob, almost stepped on me." prevents N! word combinations.

> You separate each chunk and deal with them independently. Thus replace "The great big elephant" with Bob. "Bob, almost stepped on me." prevents N! word combinations.

You're changing the goalposts. The original post claimed to encode all language parsing and intelligent responses as a static set of if-statements; regardless of whether a sentence makes semantic sense, they must still be recognized in any combination in a fixed set of if-branches and responded to intelligently (like the question "why doesn't this sentence make sense?"). That's not feasible for the reasons I describe.

Parameterization like you describe is precisely the sort of model that an evolved intelligence would exhibit.

Parameterization works with if statements. If (like("The great big elephant,%"))

Further, the structure of if statements can encode abstractions not explicitly expressed.

I already responded to this point in my last post. Even if you were to parameterize as a form of compression, which is outside the scope of the original point to which I was responding, it still doesn't meaningfully compress the results because I can just generate a whole new impossibly large set of intelligible questions, like "why doesn't sentence X make sense?". The number of sentences that require intelligent answers is simply impossibly huge for a fixed set of if-statements.

That might be the case, however you did not demonstrate it. My point was if statements can encode abstractions beyond a single huge case statement. You can for example encode a neural net as a set of if up to some granularity.

A much more central issue is context is important. ex: "Yes." is much harder than "why doesn't sentence X make sense?"

Your example makes sense. But your last paragraph is precisely why its grasping with the most simple and largest of solutions. Id expect the consciousness simulating robot to contain such a fact generator, not some brute force word combiner algorithm. In other words, it'd have a knowledge/concept/argument/etc network.

I recognize the grandparent poster mentioned something about having enough if statements, but yea, thats not how its going to work.

The giant lookup table is objectively not conscious. I mean it has no internal state even. That is why giant lookup tables are often used as examples of processes which appear conscious but aren't.

But what happens when you start compressing that lookup table to an algorithm which is much smaller and produces the same output? It will almost certainly gain a complicated internal state. As you make it smaller and smaller to get to a reasonable size, it gets more like an actual human brain.

It will no longer be a static program - it will have to "learn" from experience and data. That would save a lot of hardcoded rules.

Eventually you will have the smallest possible program that produces the same output exactly. And it will probably be a perfect simulation of a human brain (if that is what you are trying to imitate.) It will produce exactly the same output as a human, and run exactly the same internal algorithm as a human. And now, how on Earth can you say it's not conscious?

Because it's a possibility that consciousness is not merely observable actions or even internal thoughts (the "output")... but the feelings which guide/cause those actions (where feelings are not defined as an algorithm or computation)

So even if a robot and human had the same input-outmap mapping, one could be "coded" by "consciousness", the other by whatever humans invent as an algorithm. Maybe they will be identical, maybe they wont be. Im certainly not sure.

>Because it's a possibility that consciousness is not merely observable actions or even internal thoughts (the "output")... but the feelings which guide/cause those actions...

If it's actually "caused" by something, then that something must be a physical process which could be simulated. E.g. imagine there is some law of physics that says brains in a certain configuration will say they "feel consciousness" or whatever. That would be a real physical law we could simulate in a computer, and create a simulated brain which behaves exactly the same way.

If our behavior isn't caused by anything, if "consciousness" is casually disconnected from the world we live in, then it's irrelevant. It doesn't explain our behavior in any way, or have anything to do with us.

Reply on news.ycombinator.com