Post AVU6XUL6ZiRdXdKiDw by pixelherodev@fosstodon.org
(DIR) More posts by pixelherodev@fosstodon.org
(DIR) Post #AVU6XSfwlFvaNh4KQq by pixelherodev@fosstodon.org
2023-05-09T16:47:07Z
1 likes, 0 repeats
Stochastic parrot paper [1] is interesting, but not for the reasons I expected.I dunno that I fully agree with some of the axioms they express...> Languages are systems ofsigns [ 37 ], i.e. pairings of form and meaning. But the training data for LMs is only form; they do not have access to meaningI'm not sure I buy that it's impossible to figure out meaning just from form; the othello study[2] in particular seems to disagree.[1] https://dl.acm.org/doi/pdf/10.1145/3442188.3445922[2] https://thegradient.pub/othello/
(DIR) Post #AVU6XTKiJgXOQ88syW by icedquinn@blob.cat
2023-05-09T16:52:30.211471Z
1 likes, 0 repeats
@pixelherodev "meaning" doesn't really exist. it's just something that has to be edged around. that's something taoism struggles to express but is hideously apparent when you start looking at the encoding vectors (google models) and SDRs (numenta models)
(DIR) Post #AVU6XUL6ZiRdXdKiDw by pixelherodev@fosstodon.org
2023-05-09T16:49:26Z
0 likes, 0 repeats
The othello study in particular is fascinating; there's a follow up[1] which presents interesting evidence that a version of GPT trained purely on a list of legal othello moves and the results ends up with a *full model of an othello board* in its weights - to the point where you can identify "this number in the matrix indicates whether or not the piece at (x, y) is my color", change the number, and change what move it suggests.[1] https://www.lesswrong.com/posts/nmxzr2zsjNtjaHh7x/actually-othello-gpt-has-a-linear-emergent-world
(DIR) Post #AVU6XWO0xrxbtEi42K by pixelherodev@fosstodon.org
2023-05-09T16:50:31Z
0 likes, 0 repeats
The othello study has convinced me that the neural networks are *not* just information regurgitators.So, again... research is fascinating, but the actual real-world usage of this stuff is just terrible.
(DIR) Post #AVU8kBSwG3PKg9B8Yi by pixelherodev@fosstodon.org
2023-05-09T16:54:43Z
1 likes, 0 repeats
@icedquinn I don't agree, but I agree with a statement that is semantically identical, so.I think meaning is a "real" thing, in the same sense that numbers are "real" - i.e. they're not tangible, but they're, well, meaningful [pun intended].I think, though, that form and meaning are so intrinsically connected that it's impossible to learn to manipulate form without getting _some_ understanding of meaning.You can't learn to manipulate arithmetic expressions without figuring out numbers.
(DIR) Post #AVU8kGQHpLG03MyacS by icedquinn@blob.cat
2023-05-09T17:17:11.213862Z
1 likes, 0 repeats
@pixelherodev i think the philosophers already covered this with the quips about trying to explain color to a blind man (or something similar to this; flatland approaches the same argument)
(DIR) Post #AVU9MdvksIzNxszoSO by pixelherodev@fosstodon.org
2023-05-09T17:19:39Z
0 likes, 0 repeats
@icedquinn I dunno, the way I see it, color vs monochrome is the same as 3d vs 2d. I think you could do a decent job explaining it to someone who was colorblind.Someone who's *actually* blind though... that's a lot trickier. *If* you could figure out a way to explain the concepts of light and dark, really, I think you could bootstrap it without being able to see?I don't think you'd conceptualize it the same way as someone sighted, but I think you could understand it pretty well anyways.
(DIR) Post #AVU9MeknoVp2WCsa1Y by icedquinn@blob.cat
2023-05-09T17:24:00.616075Z
0 likes, 0 repeats
@pixelherodev if you see word2vec and semantic folding, they end up deriving meaning vectors purely by edging around the words. same as the philosophers seem to do.semantic folding is a bit more of a deliberate construct though. they do steps to relate sentences based on shared word count, spread them out in a graph, and then discretize that so that a "word" becomes a position in space for every sentence it appeared in.iirc some brain observations find similar behavior. an "apple" has spike train patterns for the appearance, texture, taste etc, which end up clustered together in some way, but something only means apple within the context of all those other experiences happening simultaneously (and, also only mean something by its distance to everything else.)they do a better job citing the names but a long standing theory is the brain has a universal storage format, which means everything is comparabe even if it makes no sense, and is also why sometimes it does make no sense, it's not exactly a vector database but is quite similar.:comfyshrug: :blobcatsip:
(DIR) Post #AVU9Ut3uV48jD9k2c4 by icedquinn@blob.cat
2023-05-09T17:25:39.023337Z
0 likes, 0 repeats
@pixelherodev the engrams "exist" but are themselves meaningless, since everything is just defined by its pressense relative to everything else in the space.
(DIR) Post #AVU9iqllowBw9fumZM by whiteline@shitposter.club
2023-05-09T17:28:11.796344Z
0 likes, 0 repeats
@icedquinn @pixelherodev"Shall I compare thee to a summer's day? Thou art more lovely and more temperate."
(DIR) Post #AVUAaES2VF9Wx08lMG by pixelherodev@fosstodon.org
2023-05-09T17:27:28Z
0 likes, 0 repeats
@icedquinn :O this is basically the design I've been using for my totally-not-an-ai projectcalled such because the term AI is such bullshit these days >_>Before reading your post, I was defining concepts [read: any "thing" known in memory] as monotonically increasing integers. So, for instance, the first time a given word is used, it claims an integer. References to other concepts would be based on integers.Types of relations would just, themselves, be concepts - i.e. integers.
(DIR) Post #AVUAaFnhUC4d8THZlA by pixelherodev@fosstodon.org
2023-05-09T17:27:51Z
0 likes, 0 repeats
@icedquinn after reading your post, now I'm thinking about a way to use n-dimensional vectors instead. Could probably make a lot of things a lot better...
(DIR) Post #AVUAaGUEw26LGPBY48 by icedquinn@blob.cat
2023-05-09T17:37:48.443396Z
0 likes, 0 repeats
@pixelherodev https://iceworks.cc/z/ckjto5dyl00009lis2exvyjxe.htmlhttps://iceworks.cc/z/ckjxo2z5r00001dis1dsfftl2.htmlhttps://iceworks.cc/z/cl7shtl6u0000p3uqm2hdgy4x.htmlyou're close. its a sparse bitmap.n-dimensional vectors are wrong but they are useful.
(DIR) Post #AVUAaKHynYff3XToVk by pixelherodev@fosstodon.org
2023-05-09T17:31:32Z
0 likes, 0 repeats
@icedquinn except it makes figuring out how to decide on their positions a lot harder
(DIR) Post #AVUAjMLuUoYc3tcrNw by icedquinn@blob.cat
2023-05-09T17:39:27.696331Z
0 likes, 0 repeats
@pixelherodev the benefit of n-dims is they are easy to hook up to gradient descent solvers and the neurons are proven to be universal calculators. so you can more easily throw a huge amount of compute to get something that calculates one without really understanding how it all works.that's been a huge problem for Numenta; they figured out how the learning system works but not how encoding is formed, so they're limited to whatever people can figure out how to encode.
(DIR) Post #AVUBNnyZDC5vJlzdCK by icedquinn@blob.cat
2023-05-09T17:46:46.049733Z
0 likes, 0 repeats
@pixelherodev the hippocampus does that. we don't know how.numenta has some made up base level encoders for data. cortical has an analytical method for doing it to semantic meaning of words.i have some bizzare 4am visions about ways to make it work, but they're not tested. mostly involving experience replay stores and fine tuning based based on that, with an added level of sparsity in that i think the whole cortical "casette" is itself sparsely updated.there are known issues in human cognition where the hippocampus can't understand hierarchical concepts. it can encode a layer and it can only understand complexity that was previously burned in to memory, and working memory similarly can't understand it, so there is some function of working memory (the phenological loop) and some read-writable casette (the hippocampus) and then blocks of modules can be elevated to the read-write system (why memory changes when accessed, i think, something causes those sections to get lifted up at some point because that casette was contributing too much error during experience replay)untested fever dream bullshit though, so, ymmv.