Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[silverspawn]

The question is, will progress hit a wall anytime soon. And the relevant question for that is probably, is there a thing about going from whatever-the-current-best-language-model is to human level reasoning that requires consciousness? If yes, that's a strong reason to believe these people are wrong, since they all don't know that consciousness matters.

In general, if something happens consciously, this is obviously done via consciousness. So vision is always the obvious example; don't expect AI to catch up to humans anytime soon. But reasoning is tricky because much of reasoning happens *unconsciously*.

However, the step where we take the good ideas and build on that does happen consciously.  If I think about a hard problem, I wait for my brain to generate ideas, but most of them are terrible. So plausibly there *is* such a step which takes consciousness.

This all fits very well with what GPT-3 and other models are doing. They generate the kind of sentences that our brain spits out unconsciously, but totally lack the supervisory step. And that supervisory step could plausibly be really hard to replicate.

This monologue was surprisingly effective at calming myself down.

[silverspawn]

Reminder to myself that at some point I need to write this kind of evidence up into a post in my sequence. I'm not there yet, but it's an import piece

[silverspawn]

Another thing is that, if something happens consciously, this is straight-forward proof that consciousness is involved, but if it doesn't, this doesn't prove anything. It's possible that consciousness is involved, it's just a disconnected part. The idea that "several entities" could be living in the brain is another thing that used to be freaky to me and now I'm just like "sure could totally be the case". In fact I'd say it's likely.

But still, sentence generation in particular seems like a good candidate for something that the brain does in pretty much the same way as GPT-3.

[silverspawn]

By the way, I've found Awaclus' twitter profile! (Not sure if linking it is okay.) Though I'd guess you have another personal one.

[silverspawn]

I remember someone on LW describing the brain as "a bunch of ML networks talking to each other with a small layer of supervisory control". That small layer could be 90% of the problem.

[Awaclus]

By the way, I've found Awaclus' twitter profile! (Not sure if linking it is okay.) Though I'd guess you have another personal one.

It is okay to link it, but I don't have a personal one in addition to that one (I do also run the Birds of Necama twitter account though).

[silverspawn]

https://twitter.com/Awaclus (pretty obvious url isn't it)

I saw you responding to an Aella poll :-)

The fact that you don't also need a normie account where you give milktoast political takes on everything is pretty nice imo

[silverspawn]

Also believe it or not but the banner finally shows you profile pic big large enough to where the white symbol no longer looks like a bird.

[Awaclus]

The fact that you don't also need a normie account where you give milktoast political takes on everything is pretty nice imo

It's hard for me to pretend to be multiple separate people, and there wouldn't be any point in having this public Awaclus account and a separate normie account if someone could just connect the dots (which is easy) and ask me if I'm the same person as myself in which case I would be forced to admit it, and then all my normie followers would see all my responses to Aella polls anyway.

It is pretty intense to have this account where I'm mutuals with some family members, some people I have a professional relationship with, some friends (including online friends, some of whom I've met on Twitter itself), and a lot of people I share a number of political goals with, and then I'm supposed to participate in conversations I think are interesting without offending or boring this totally random collection of people or embarrassing myself too much. I'm not necessarily doing a great job at it either but it's an interesting experience at least.

Also believe it or not but the banner finally shows you profile pic big large enough to where the white symbol no longer looks like a bird.

I actually find it very easy believe that it doesn't look like a bird in my banner.

[LaLight]

@Awaclus, I've followed you (heighthater), although I mostly lurk on Twitter, rarely write something and it's usually in russian

[LaLight]

meanwhile, a university I want to enter in Netherlands sent me a letter that I need to pay a tuition (it's fine) and a proof of financial means total of 19000 euros. It's illegal and prohibited to send more than $10k from russia to any "unfriendly" country :/

[LaLight]

i mean i have passive income and i can live in Netherlands with it, but I can't just give 11k euros to them, i get money monthly. I'll contact them about it and fingers crossed they'll let me enter.

[silverspawn]

This is probably not relevant, but hypothetically speaking, there is this thing where a person can open an account on a centralized exchange, buy ETH for $ with negligible fee, and send it to an arbitrary wallet for about 15$ fee (regardless of the amount)

[silverspawn]

This is a hilarious poll

[silverspawn]

So one of the students in my class has chronic sleep problems (which was immediately obvious to me after a colleague told me, but which I managed not to realize before). She basically seems extremely tired all the time, which is a state I remember because I've sometimes been very tired during class, and it's pretty unpleasant. Should I do anythng about this?

[Awaclus]

So one of the students in my class has chronic sleep problems (which was immediately obvious to me after a colleague told me, but which I managed not to realize before). She basically seems extremely tired all the time, which is a state I remember because I've sometimes been very tired during class, and it's pretty unpleasant. Should I do anythng about this?

I often have a weird sleep schedule (not sure if I should call it chronic sleep problems, really it's just a symptom of my Internet addiction) and it is incredibly helpful for me that we don't have mandatory attendance on lectures and everything is recorded so I can just watch it whenever I'm awake. If that isn't possible, I can't really think of anything else that would help, and if she doesn't approach you to request something in particular, you probably don't have to do anything.

[silverspawn]

I didn't even get around to bragging about this, but I've now been mentioned by name twice in a slatestarcodex article. Pretty sure there was some kind of "achievement unlocked" icon that was supposed to appear in mid air. Must have missed it.

[silverspawn]

I've found a paper that claims to have demonstrated that neural networks fall prey to the Gestalt Law of Closure, that is, they see non-existent lines that complete figures, e.g. a triangle.

If this is true, it would be strong evidence against my new world view. If not a total knockdown.

... but the way they measure "Closure" is, they look at (a) similarity between a network given the full image and a network given the incomplete ("illusion") image, and (b) similarity between a network that is given the full image and a network given the incomplete image with parts rotated in such a way that the lines don't look like the figure anymore. I.e., one pair of networks reading the top two images vs. one pair reading the bottom two.



Then they define Closure as the difference in similarity.

this... is just stupid, right? like there's no reason to think this measures Closure? It only shows that the top two are more similar, which is already obvious?

[silverspawn]

pretty baffled here because why on earth would you define Closure like that? Why not train a triangle classifier and then see whether it classifies the illusion as a triangle? Or better yet, train a network to do nothing but reconstruct the original image, feed it an illusion, and see if the reconstruction has the illusory lines. But this??? why?? and why would you think this measures Closure???

(mb it'll make more sense if I read the paper completely, should do that.)

[silverspawn]

In Trixie Slaughteraxe for President, there is this funny part where they visit the castle of language makers, that is a bunch of ladies trying to develop languages that (if widely adopted) would be good for society. We have

- One language with math problems built into  basic syntax ("babies can learn all languages, therefore they will do this easily, then everyone will learn to solve math problems from a young age")
- One language where everything rhymes ("so easy to write poetry!")
- One language without words for bad things ("we'll forget about evil!")
- One language where every substantive and adjective has a good and a bad form, forcing you to constantly judge what is good or bad when speaking ("people will think deeply about what's good or bad, and figure it out")

[silverspawn]

Aside from the last (sort of), Oyvind never passes judgment on these, that is left as an exercise to the reader, but my take is

1) lol / confusing correlation with causation / babies would never learn this
2) misunderstands how art works, being hard to do is a feature... alas
3) this is by far the least stupid on the list, but probably not practical anymore
4) sufficiently obvious failure mode that I don't even need to say it

[silverspawn]

maybe one reason I like this is because it's so typical for the author. He just wants to make up weird conceptual stuff like that. His comics are chocked full of that.

[Awaclus]

2) misunderstands how art works, being hard to do is a feature... alas

I don't think that's true. Being hard to do is an inevitable result of the fact that the people who are the most passionate about an art medium will put a ton of effort into learning it, and then that'll set the standard that all other attempts are judged by. If you make it easier, more people will be able to meet that standard, but then the most passionate people can start ignoring that aspect of the challenge and put their effort into something else that pushes the standard even further. In that sense, art will fundamentally always be hard to do, but any kind of an innovation that reduces the existing challenges will result in better art being made.

[faust]

2) misunderstands how art works, being hard to do is a feature... alas

The misunderstanding I think is a different one. Japanese is a language in which way more things rhyme than in most western languages (because there are fewer syllables), and guess what? Japanese poetry isn't about rhymes at all!

I don't think rhymes are used in western poetry because they are "hard to do", but more because it's a rare form of order in language and that is perceived as beautiful.

[silverspawn]

plot twist: it's all of the above.

I think signaling is definitely a component of art, being different is also a component, and inherent beauty is also a component