tranny demon hacker

https://kolektiva.social/@beka_valentine

trans dyke hacking on @QueeriousLabs@twitter.com and DDoSecrets she/her (xey/xem or ey/em if you're bored with binaries); no child zone, 18+ only 

current twitter handle: @defnotbeka@twitter.com 

former twitter handles: [@beka_valentine](https://kolektiva.social/@beka_valentine ) @illegally_beka

location: SF, CA

Public Key

npub132p3v5et2ujg4jpdklvxdge2qlvlzjgqatar7lpmfarx9syl0x5skgwsev

NIP-05 Address

beka_valentine_at_kolektiva.social@momostr.pink

Profile Code

nprofile1qqsg4qck2v44wfy2eqkm0krx5v4q0k03fyqw473l0sa573nzcz0hn2g8zm6j4

Author Public Key

npub132p3v5et2ujg4jpdklvxdge2qlvlzjgqatar7lpmfarx9syl0x5skgwsev

Show more details

Last Notes

2024-09-05 05:27:58

- reply

if they wanted to be frightening, they should've become landlords and managers at low wage jobs

2024-08-19 15:19:51

- reply

but, this is still an assertion
what is the gricean approach for questions?
maybe something dual??
instead of "say all the true relevant things, and nothing more"
its
"ask for only the true relevant things, and nothing more"?

2024-08-19 15:18:12

- reply

the response that's given is entirely easy to understand. "X and Y" is true, but B said only "X", and also Person A would've wanted to know "Y" if Y was true. so the gricean version makes perfect sense

2024-08-18 05:02:53

- reply

i'm not really thinking of language here, but also people have very little access to how their cognition works, and even the parts that are in principle accessible aren't inherently *accessed*

2024-08-18 04:45:32

- reply

I think that human cognition is pretty definitely structured semantic in nature, but that's not to say that all aspects have structure or that the structure is what first seems like structure

2024-08-18 04:38:05

- reply

its really cool. it feels like the JJ Abrams enterprise, almost?

2024-08-18 04:00:11

- reply

something like @npub1s84…t8k4's Idealized Cognitive Model might be a good reference point for what this should look like

2024-08-18 04:00:07

- reply

I think this is a good reference point for human interactions, actually. using cultural archetypes, personal archetypes, etc. as a means of narrowing in on ideas

2024-08-18 03:58:15

- reply

and what's an archetypal intimate situation? one with just one other person in it, of the sort that the judger-of-intimacy finds attractive

2024-08-18 03:56:52

- reply

something else to consider wrt the audience example is the notion of archetypes. what is an archetypal jazz club audience like? well, judging from Google Images, it's a packed house, which is what Riker got the first time around

2024-08-18 02:41:43

- reply

also relevant is that some properties of things are graded quantitative things, whereas others aren't really. the existence of an audience isn't graded but the size of it is
i'm not sure if there are these kinds of interaction cues that don't invoke graded quantitative things

2024-08-18 02:39:19

- reply

something else that seems relevant is the intensity of Riker's responses. he didn't say just "something more intimate", he had a strong negative reaction to the current situation, suggesting that it wasn't merely wrong but VERY wrong

2024-08-18 01:24:25

- reply

what does riker want? who knows. but we can get a good rough sense by exploring the two extremes first, to see how he feels. his responses are probably going to be similar for similar values. if "Full" is bad, then near full is also probably bad too, so we get a kind of fuzzy judgment

2024-08-18 01:22:57

- reply

its not merely that "more intimate" means "less people", it's directed towards, not merely away from. and that feels relevant
but also, the vagueness of audience size also suggests that maybe some general feeling out matters

2024-08-18 01:20:58

- reply

but as a human, that feels.. wrong
riker wanted an _intimate_ audience, not just a smaller audience. intimacy implies certain things, and so maximizing intimacy means minimizing size

2024-08-18 01:19:21

- reply

from a binary search perspective -- which this is -- that's actually the wrong choice. we'd want to start in the middle, halving the space entirely, and thus making a nice binary search down audience sizes

2024-08-18 01:17:52

- reply

there's a continuum of audience sizes, from 1 to full. why is "full" (and i think also "1"), a good starting point to guess? is it because there's only one direction to move in if it's wrong?

2024-08-18 01:14:56

- reply

going back to Riker on the holodeck, what is it about the audience size that makes the two extreme so salient?

2024-08-18 01:11:25

- reply

it might be possible to do something like make each _atomic_ proposition equally weighted, so we can just get on calculating, and then derive the quantities of the compound propositions, but who knows if that's correct

2024-08-18 01:10:04

- reply

a proposition A is informative to the extent that roughly half the worlds make it true and half make it false, but we don't have access to that knowledge. that's a gods-eye-view of things that we lack

2024-08-18 01:09:20

- reply

I'm not sure there's a principled way of doing this in the absence of probabilities, etc.
informativity is fundamentally about the cutting down of alternatives to what the world could be like, and that depends on how many alternatives there are

2024-08-18 00:56:14

- reply

see downthread

2024-08-18 00:53:50

- reply

a hypothesis is informative when the worlds in which its true, and the worlds in which its false, are roughly balanced. finding a yes or no answer cuts down possibilities in half no matter what answer you find

2024-08-18 00:51:35

- reply

so the act falsification doesn't actually narrow down the hypothesis space
so in some sense, i guess stronger hypotheses aren't better. instead you want maximally informative hypotheses

2024-08-18 00:48:05

- reply

but there are a LOT of combinators like `map` which are definable in terms of foldr, lots of strongly claims to make. in that setting, we generally would be better off hypothesizing foldr

2024-08-18 00:45:08

- reply

i'm not entirely sure how to translate these phenomena into a formal, mechnanizable system for exploring hypothesis space, but it does feel quite similar to what Lakatos and Popper would aim for

2024-08-18 00:42:27

- reply

but if `John saw Mary` is true, and `John saw Mary and Susan` is false, then it must be because `John saw Susan` is false. that's a deductive fact

2024-08-18 00:41:52

- reply

so since the speaker said `John saw Mary` and not `John saw Mary and Susan`, we conclude that `J saw M and S` must be false

2024-08-18 00:41:11

- reply

if the speaker believed that to be false, however, then the speaker is encouraged to not say that at all, and say something else instead

2024-08-18 00:40:08

- reply

obviously, since it's possible that John saw Mary and also Susan, that's a possible thing that could be said. if it were true and known to be true by the speaker, the maxims would encourage the speaker to say that instead

2024-08-18 00:36:50

- reply

for example, if a person asks "Did John see Mary or Susan?" and the response is "He saw Mary", it's often understood as also meaning "He didn't see Susan"

2024-08-18 00:33:32

- reply

something else that feels relevant is Grice's Conversational Maxims from linguistics. Grice observed that people often say things that seem to.. mean more than they say on the surface, and part of how this is achieved is by some principles he hypothesized

2024-08-18 00:30:50

- reply

hypothesis here in the sense of "possible explanation you thought up", as part of an abductive process, not in the sense of the hypotheticals in a logical deduction

2024-08-18 00:29:45

- reply

one thing that seems relevant to me is the idea, which feels very Popperian or Lakatosian, that hypotheses should be _strong_, because stronger hypotheses are more easily refuted

2024-08-18 00:27:21

ok so I want to talk about this some more, how we might actually go about building such a system
https://kolektiva.social/@beka_valentine/112980070125642455

2024-08-18 00:12:18

- reply

=== END THREAD RECREATION ===
WHEW. that was a doozy ^^;

2024-08-18 00:12:06

- reply

its like test driven development taken to its extreme, but its richer than just that because the system could also in theory generate *new tests* that the user could validate or reject

2024-08-18 00:11:59

- reply

and then when the user says "no this is wrong", you can relate that back to the hypotheses and also to the program structure, and use that to construct alternatives
in some sense this would be programming and debugging in one, programming-by-debugging sort of

2024-08-18 00:11:53

- reply

you could say the "world" is some structured object, a program perhaps, and the theory would be about how some function's Input-Output relation looks maybe, or who knows what, and that would let you generate hypotheses for what behavior to expect from previously unseen inputs

2024-08-18 00:11:41

- reply

oh yes and this general framework might work quite well for building smart debuggers for software, too. and also smart automatic programming tools, that actually analyze problems to develop novel solutions rather than just generate boilerplate based on prompts

2024-08-18 00:11:30

- reply

i think that this is actually a major component of how people, maybe even some animals, learn things about the world. i think this is _the_ mechanism that underlies belief revision broadly speaking, not just Science, but all our beliefs

2024-08-18 00:11:24

- reply

this would also, conveniently, let us make double use of the holodeck program, to not only generate stories and worlds, but also to do things like programmatically do science
in fact, i think this Lakatosian ranked theorizing thing is probably a core part of cognition broadly

2024-08-18 00:11:17

- reply

so we structure hypotheses by ranking them with priorities, but we ALSO structure them by relating them to how we structure the world, by adding a sort of additional orthogonal ranking of hypotheses by how they're relevant to the critiques being offered

2024-08-18 00:11:05

- reply

and then we can use the structure of the world again to select which hypotheses matter: the further in world structure something gets, the further that it's hypotheses should get from relevance. we might pick multiple hypotheses to reject because they _all_ reference hair color

2024-08-18 00:10:59

- reply

but, because the hypothesis makes reference to parts of the world, we can incorporate that into the selection of which hypotheses are relevant for the critique, and say that, for the sake of this falsification, some not-minimal hypotheses are up for debate

2024-08-18 00:10:52

- reply

the critique of The Woman's hair was structurally about some particular part of the world, at some *small* scale, with some proximal things
but that doesn't mean that the hypothesis related to that is especially lowly ranked

2024-08-18 00:10:41

- reply

..we start backing up to a higher-ranked hypothesis, and throw it out, because now its becoming increasingly likely that the problem wasnt the low ranked hypothesis at all
but of course this is all in relation to the critique and how it structures the world, see

2024-08-18 00:10:35

- reply

instead we're throwing away the parts of the theory that are close to the targeted hypothesis in terms of ranking etc. and only after we're forced to go through a bunch of different alternatives, after we've narrowed down the theory space so tightly that we cant proceed, THEN..

2024-08-18 00:10:29

- reply

and when that part is falsified, the alternative hypotheses are closely related, in that, we're not throwing away large chunks of the theory just on the basis of one falsification, on the basis of one rejection from Riker

2024-08-18 00:10:21

- reply

the computer has a theory about what constraints Riker has on the world, with rankings, etc. and it uses those constraints to generate structured candidate worlds, but it also does so in a way that the candidates -- if rejected -- falsify part of the theory

2024-08-18 00:10:13

- reply

..b) the constraints should be related to one another via some kind of relative prioritization
this comes from the fact that the interactions that the holodeck computer is performing look an awful lot like a Lakatos-style scientific research program and theory testing

2024-08-18 00:10:06

- reply

i INCLUDE that, but i also mean that the theory of what Riker wants should be something more than just "Riker was a world that looks like this: <the world, with its structure>"
rather, the theory should be a) separate, as a bunch of constraints which the world satisfies, and..

2024-08-18 00:09:55

- reply

oh and when i say that things should be semantically structured, i don't mean just the objects in the world
like, you could imagine that what's going on is that the world itself is richly structured and all we're doing is mutating the world's structure
but I mean something else

2024-08-18 00:09:47

- reply

THIS is what i want out of assistive AI for art :)

2024-08-18 00:09:31

- reply

..the evolutionary fitness function can be applied not just to the whole scene but to semantic components, triggering localized mutations, and where the CHOICE of the mutation is the one that is maximally informative wrt the semantic content of the fitness critique

2024-08-18 00:09:21

- reply

so what we see in this scene from the Star Trek: TNG episode 11001001 is an evolutionary process, exploring mutations in a space of possible candidate holodeck scenes, where the content of the scene is richly structured by its meaning/semantics, and where...

2024-08-18 00:09:15

- reply

and so with that constraint, there is no most informative choice amongst the non-blond hair colors, so it picks red, which is a fine choice, but which riker then modifies more by adding more cultural cues: "sultry" conveys dark, mysterious, which suggests brown or black hair

2024-08-18 00:09:08

- reply

but the computer also knows that riker didn't simply say LESS blond, he said NOT blond -- he's clearly not thinking of hair color as a spectrum but rather a Blond-NotBlond dichotomy, using a different cultural frame of reference (perhaps the Dumb Blonde trope)

2024-08-18 00:09:01

- reply

and notice that for the hair color, there is really no inherent informational difference between "red" and "brown" with respect to "not blond"
there might be some sort of cultural gradation that we have of hair color, some spectrum from blond to brunette with red in the middle..

2024-08-18 00:08:51

- reply

so the computer, being smart, really ought to be exploring the mutation space in such a way as to maximize the information gained by subsequent critiques of the choice being made

2024-08-18 00:08:44

- reply

by picking the extreme case of intimate crowd -- a crowd of one -- the computer can get a nice sense of what is "too intimate", whereas a Packed Crowd - 1 is "more intimate" but not in a way that's information dense

2024-08-18 00:08:36

- reply

rather than going from Packed Crowd to Packed Crowd - 1, the computer tried the opposite extreme
partially this is because riker said "something more intimate", but also it's the most informative choice the computer could make

2024-08-18 00:08:28

- reply

also, if the computer were SMART, what it would be doing is not just exploring these mutations but also trying to do some kind of theory building to model Riker's desires, which it plausibly seems to be doing

2024-08-18 00:08:20

- reply

..the decor or the band either. tho it did also change where she was seated, what she was drinking, etc.
so mutations to generate new candidates are more likely when they're closer to the reason for mutation, but they CAN happen at larger scales than just that reason

2024-08-18 00:08:13

- reply

..just the size of the crowd in some sort of minimal mutation way. but it did stay NEAR the source of the mutation, by not straying far into things like the decor or the band
similarly for the woman, it changed the whole woman not just her hair color but it didn't change..

2024-08-18 00:08:06

- reply

the crowd was very close to the cause of the mutation, which was precisely the size of that crowd, so that's all that changed
the computer didn't merely _reduce_ the crowd size, by subtracting some people, it generated a new crowd with different people, so it didn't mutate..

2024-08-18 00:07:58

- reply

the lady's hair was not the right color, so the computer regenerated the hair color, not everything AND the hair
the scope of the mutation, in a generic algorithm sense, is confined to things nearby the cause of the _demand_ for mutation

2024-08-18 00:07:42

- reply

the computer, being smart, should be trying to tailor the scale of the changes to the scale of the rejection
the crowd was "too large" so the computer re-generated the _crowd_, not everything AND the crowd

2024-08-18 00:07:34

- reply

constraints are updated and a new, more constrained choice is made subject to those new, tighter constraints

2024-08-18 00:07:27

- reply

but then the woman isn't quite what he wanted so he rejects some aspect and the computer randomly picks another, and so on
its this interaction, this conversation with the world, where some constraints are added, a random choice is made, some critiques are offered, and the..

2024-08-18 00:07:21

- reply

he wants an audience to it randomly picks a large one, so he has to explicitly rank that as a bad choice, so the computer re-generates and instead picks the smallest possible audience fitting the criteria, one person who right might find "intimate": a beautiful woman

2024-08-18 00:07:14

- reply

==END==
what we see here is that Riker is narrowing in on different scales and different aspects of the scene and specifying them semantically, but also the computer is having to specify some aspects randomly

2024-08-18 00:07:08

- reply

Riker, to the Binars: Great job boys..
Riker: But computer, blonds and jazz seldom go together.
the woman disappears, a redhead appears at another table.
Riker: Now that.. is truly exceptional. But more sultry.
the redhead disappears and a brunette appears at the bar

2024-08-18 00:06:39

- reply

Riker: Now an audience.
the room is now full of bar patrons, bar staff
Riker: Whoa.. too many.
the patrons and staff vanish
Riker: I was thinking of something.. a little more intimate
a blond in a red dress appears seated at one of the tables, smiling at Riker

2024-08-18 00:06:32

- reply

and he enters the bar and there's a jazz band's equipment on stage on the left, a bunch of tables, the bar in the back. no people
Riker: Now I'll need someone to play with. A trio. Piano, bass, and drums. And a [trom]bone for me.
the instruments and players appear

2024-08-18 00:04:48

- reply

Riker: Computer, I'd like some place to play some music, a little atmosphere
Computer: Specify.
R: Jazz.
C: Era?
R: Circa 1958.
C: Location?
R: Kansas City... no, New Orleans. Bourbon Street Bar, New Orleans. Around 2am.
C: Program complete. Enter when ready.

2024-08-18 00:04:34

- reply

i imagine as the sort of pinnacle kind of situation something like you find in Star Trek, where Riker walks onto a holodeck and just describes a scene and gets it, but then iteratively refines it with fuzzy specifications
gimme a moment, i'll transcribe the scene i have in mind

2024-08-18 00:04:24

- reply

and the semantic dimension lets us really narrow down the space of possible structures to operate on, so that our GAs only have to explore a very constrained region of the genomic space

2024-08-18 00:04:16

- reply

this is especially true of depictive things, depictive art, stories, etc. rather than more abstract qualitative things like abstract art or music without lyrics
the depictive dimension is the semantic dimension of it, there's a thing we're trying to depict that we can describe

2024-08-18 00:04:06

- reply

at some level the semantic dimension is more important because when we're envision some piece of art or some piece of music or whatever, we *think* of it in structured semantic terms

2024-08-18 00:03:32

- reply

the semantic dimension could be aided dramatically by also including linguistic cues not just evolutionary pressure cues. for instance, "darker hair" could be a semantic cue about the subject's hair color, but it's not precise so there's wiggle room to have a GA apply

2024-08-18 00:03:25

- reply

or a picrew creator or whatever. these all fix a particular domain: portrait art, and maybe art style, or whatever, and let you manually select attributes for different features, in a highly structured way, but you could evolve that too, instead of manually selecting it

2024-08-18 00:03:19

- reply

but once we're talking about that, we're going to need to have some semantic level that allows us to select the very fact that there is, indeed, a subject
there are actually assistive art things that do this already for fixed domains, right
this is a character creator

2024-08-18 00:03:13

- reply

im not sure what the "Right" structures to evolve would be for art, but i think it's going to require some exploration. it could also be that they're not obvious ones like spatial things, but rather we might want to use more human oriented semantic ones

2024-08-18 00:02:57

- reply

2024-08-18 00:02:46

- reply

until finally you end up with, among other things, "x1 = x0 + vt + 1/2 at^2"

2024-08-18 00:02:38

- reply

some genetic algorithms techniques exist which do this for mathematical theories. i've seen some GAs for "evolving" a theory of newtonian kinematics by starting from an initial genome of like, the equation "x1 = 0", and the mutating the RHS enriching the mathematical expression

2024-08-18 00:02:24

- reply

the object being manipulated here is no longer a flat genomic sequence but rather a rich structure that *gets richer* as things evolve
the "genome" is not some fixed size array of "genes" but rather the genome is a complex structure

2024-08-18 00:02:11

- reply

and all of this stuff has to be amenable to the genetic algorithm, all of these choices: subdivide a quad, focus on a subregion or superregion, paint a random proportionally scaled spot of some color

2024-08-18 00:02:01

- reply

but another thing you probably could benefit from is the ability to move up and down the scale of where you're changing the imagine -- sometimes you want to change something about a tiny area, brighten it or darken it somehow, but then also pull out and change a containing area

2024-08-18 00:01:55

- reply

you might want to be able to, on the other hand, have overlapping subdivisions, so that you can have two regions with the same size but which aren't strictly adjacent, but rather overlapping, maybe

2024-08-18 00:01:49

- reply

so you might have the spatial layout of the image be treated like some kind of spatial tree that has arbitrary detail, but only if you explore to that level of detail, like a quad tree, where you subdivide into quads only as needed

2024-08-18 00:01:41

- reply

so i think that what you *really* want is to be able to have a more richly structured notion of what you're producing. if its an image, for instance, you might want to have not only the scale you're working at but also the regions you're working within on an image

2024-08-18 00:01:29

- reply

because the images themselves are not structured, the only structure information is the scale that you're working at, the image itself is structured only at the pixel level or whateve

2024-08-18 00:01:19

- reply

but i can also imagine that this is going to have some problems, because if all you're doing is painting onto a canvas, say, then you can't really explore more alternatives that are "like these, at the same scale of detail, but not the same as these", you can only increase detail

2024-08-18 00:01:11

- reply

and from there you can select the "best 3" or something, out of 10 or 20, and then in the next round the blobs that gets paint get smaller, maybe, and this keeps happening over and over until you've sort of fleshed out all the relevant detail and you have a portrait

2024-08-18 00:01:01

- reply

and some of those will be more portrait-like -- blobs toward the middle of the image, or maybe the middle and also the whole bottom if you're including shoulders in the portray, say
and others will be way off: a blob at the top anywhere is weird not portraity

2024-08-18 00:00:49

- reply

that way, if your target, for instance, is a portrait of a woman, you might for instance be able to start with a picture that is a few blobs of color in random places. large blobs rather than small ones, on the scale of maybe a third or a quart of the image. something big

2024-08-18 00:00:38

- reply

what you want instead is some sort of process that "narrows down", that is to say, toward the beginning of the process, the process is very macroscopic, very rough, it's about getting the vague shape of things quickly, and then whittling down to finder details as you progress

2024-08-18 00:00:28

- reply

thats just gonna take a while, and part of the reason why is that there's no granularity there. if your algorithm does nothing but set pixels one at a time, the contribution that any single pixel has to the overall result is marginal