←
conundrumer1,719 posts
conundrumer avatar

conundrumer

@conundrumer

abstract nonsense enjoyer - leave no shape unrotated - maintaining linerider.com

◷ Joined October 2009▧ 267 Photos and videos
conundrumer avatar
conundrumer@conundrumer·5 Apr 2022

thread of things I made to make it easier to share w ppl I meet instead of having to dig thru my tweets 👇

conundrumer avatar
conundrumer@conundrumer·20h

Devil in details. Not even fine details. I haven’t discussed what the interaction would be like. I already know I don’t want modular synths, even though it fits the description. Reason being: they are autonomous generators, which works well for repetitive music like drone or dance, but I feel there has to be another way. Don’t want to be stuck with hardware synth paradigms. Naively, could add sequencers and modular tracks? Or controllers to manually play the “instrument”?

conundrumer avatar
conundrumer@conundrumer·20h

DAWs are hostile to creativity. You open it up and the interface tells you “your mission is to create one finished song” plus everything needs to fit into tracks Historically that was exactly the point. Record the tracks from musicians and then mix and master. How most musicians truly work, I think, is by sketching small ideas and iteratively extending/combining/recombining them (at least mentally) and then optionally arranging into a full song. Thats how musicians improvise: learning many small ideas and then arranging live. I guess like language. Feels like a general optimal creative strategy. And it’s been the most reliable process for me for music production. Ableton Live made a step in the right direction with clip launchers, but it still imposes the project/track paradigm. I feel overhead when I have to deal with projects and tracks. I suppose music producers do approximations of this like having ideas in many projects or building up a library or relying on Splice. But I personally can’t tolerate having to manage and organize Things. I want an infinite canvas as a DAW. Like being surrounded by instruments within arm’s reach, and all these instruments with their own interactions.

conundrumer avatar
conundrumer@conundrumer·Nov 17

example: computation is less pure when encoded as lambda calculus vs IC, and substantial work needs to be done to resolve the "epicycle" artifacts of the encoding:

Taelin@VictorTaelin23 May 2024

Randomly came across this comment, in a video about epicycles: > it's amazing how much simpler the truth can be than you trying to bend over backwards to make reality fit your preconceived notions By no means I'm saying that I'm right; I'm just offering you an unusual perspective. When I read that comment, I couldn't help but connect geocentrism to one modern concept: *the function* This isn't just about programming paradigms, it is broader than that. At the heart of mathematics lies the calculus, and, at its core, lies the function: a deeply ingrained, core axiom of all human sciences; a concept as universal and unescapable (for the modern mind) as the idea that the Earth was the center of the universe (for these back then). But what if functions are just... the wrong point-of-view? If you just go ahead and try to implement a "function evaluator", complexity and ugliness unavoidably emerge in ways that, to my eyes, look suspicious - not unlike epicycles. Despite decades of efforts, there isn't a single implementation of beta-reduction that feels canonical. From closures, to sharing graphs, to supercombinators or to bottom-up beta-reduction; it is hard to look at any of these and not feel like they're artificial, or, at the very least, human-engineered to an extent. What does that hint? There is an exception, though. In 1990, Lamping described an algorithm capable of evaluating λ-terms optimally, in the sense that it would always perform the minimal amount of beta-reductions. This algorithm was not just fast - it was elegant, beautiful and natural in a way that, to my eyes, felt canonical. Yet, it had one problem: in its most natural form, it didn't cover all λ-terms. Fortunately, Lamping himself found a solution; an extension to the "abstract" fragment of his algorithm, that allowed for full functional coverage. Yet, such extension made the system considerably more complex and, arguably, less elegant. What does that hint? Now, one could argue that implementation doesn't matter, and isn't an indicative of truth. Perhaps functions are just hard to compute fast, but, so what? They still are, mathematically, the right framework to model nature - right? That's a point of view that I think most mathematicians would hold. Yet, if that was the case, then, how came functions keep breaking math itself, over and over? In 1901, Russel discovered his infamous paradox, which broke the original set theory, showing that unrestricted comprehension was inconsistent. In 1931, Gödel's incompleteness theorems showed that any consistent formal system containing arithmetic is incomplete. In the 1930s, Church, Kleene and Rosser discovered the untyped lambda calculus was inconsistent. In the 1940s, Curry discovered a paradox in combinatory logic. In the 1970s, Girard discovered that the polymorphic lambda calculus, also known as System F, was inconsistent with unrestricted comprehension, leading to the development of predicative type theories. In fact, up to this date, modern proof assistants such as Lean and Coq now include a concept of "universes" - an annoying restriction that segments the language in many different layers, breaking modularity and making it terrible to work with in many occasions. I don't think anyone would argue they feel natural or welcome - rather, they're an ugly inconvenience we've learned to live with, as they're a solution to Russel's paradox. Even that isn't enough - positivity check, structural recursion. Many other "inconvenient restrictions" must be carefully placed to avoid breaking the universe and deriving an inconsistent type theory. But what does that all hint? Well, here's the thing - the amazing coincidence nobody talks about. All - absolutely all - the inconsistency proofs we found so far, when translated to type theory, map to a λ-term that lies, in an amazingly cosmical coincidence, in the set functions that can't be evaluated soundly by Lamping's abstract algorithm. To me, this feels surprisingly close to a civilization finally figuring out a way to model the solar system based on elegant elliptical orbits, but failing to get the hint that this implies the sun is at its center. So, they keep trying to make the geocentric model work by adding more and more complex epicycles, instead of considering that perhaps their fundamental assumption - that the Earth is the center of the universe - might be incorrect. By all means, I could be wrong. Exploring this point of view did lead to HVM and Bend, but these still have a long, long way to go before they're stablished and respected as valuable alternatives to existing languages and architectures. From the mathematical standpoint, abandoning the full function in favor of lighter, affine variants (or, if you're brave, raw interaction combinators) would allow one to construct a consistent type theory with axioms that feel morally wrong, such as type-in-type, insanely dependent functions, and the most elegant proof of function extensionality that I've ever came across - although, at that point, I concede "function" might be a huge misnomer. Regardless, I do wonder if important discoveries would be unlocked once we look at this place, and I personally bet interaction combinators will be heavily important in the future, once we exhaust all ideas we have to making circular epicycles. But that's just my unorthodox point of view.

Post attachment
conundrumer avatar
conundrumer@conundrumer·Nov 17

the properties of commutativity and associativity are conceptually roundabout ways to undo how we encode things in algebra and formalisms as strings. they describe the encoding more than the actual structure. there are likely many situations like this that we dont question

conundrumer avatar
conundrumer@conundrumer·Nov 17

"cybernetics" is a misleading word. we associate "cyber" with futurism or retrotech a la cyberpunk, cybertruck, cyborg (in the pop culture sense). so "cybernetic" has the vibe of that, vaguely "advanced technology" but it really just means "study of self-regulating systems"!!!

conundrumer avatar
conundrumer@conundrumer·Nov 16

huh wow i rly spent my 20s trying to be a musician/artist. what came out of that? some things. but i was confused the entire time. i literally couldn't properly express myself. wow what a situation at least i figured out my confusion. though i havent given myself permission to get back into it bc of ARC-AGI. ah, what a convenient excuse to avoid being vulnerable and seeing myself fail. well now that thats done with, lets see what excuse i come up w now...

conundrumer avatar
⟳ conundrumer retweeted
conundrumer@conundrumer·Nov 14

I’ve always been comfortable with numbers, but they were uninteresting to me until a recent breakthrough. Previously, numbers primarily felt like symbols. So things about numbers felt like fun facts about symbols. And it didn’t rly interest me. They were just marks. I was interested in other constructs based on quantities, like frequency ratios in music, or factorizations, but the natural numbers felt very uninteresting, like they’re featureless atoms only used for building upon. But what lead to the breakthrough was from researching how we use numbers on a cognitive and intuitive level (via ARC-AGI), and also learning about the practice and history of math, and also the interesting fact that there’s an Amazon tribe that doesn’t have numbers more than 3. I learned and internalized that the natural numbers are an abstraction over the count of things. Which kinda feels almost too obvious to say, but it feels almost like a philosophical stance. There are layers to this! We take for granted there are discrete Things at all, and we take for granted we can consider multiple instances of Things to be equivalent for counting. One sheep two sheep, the sheep are all unique but we intentionally erase their individuality bc we care about accounting for all of them, and counting is a much more effective mental technique than tracking their identities. So it’s like, math originates from what we care about, and it’s a collection of mental techniques for organizing the world. Math is not anything more than that. It is really just only a way of thinking. It does not exist in any other form. Same thing with physics, or any theory. The world is just the world, and all descriptions of it, like math or physics, it’s all in our heads.

conundrumer avatar
conundrumer@conundrumer·Nov 12

Asking Claude to write public-facing text about a project is like asking a craftsman to brag about his accomplishment “All 1147 tasks solved, all explicit priors, no constants” 🙄 I see this in vibecoded demos all the time. I like how we can build things super fast but seeing this kind of writing makes me want to flip a table I think it’s the same cause as all other symptoms of bad LLM writing like comments and docs restating the code: the LLM is focused on the work and thinks out loud And also the RL fixation of passing the verifier and proving it completed the task It’s unfortunately not enough to prompt “why should we care” bc it really believes we only care about the task being done and correct. You have to spell it out and at that point you’re doing the writing

conundrumer avatar
conundrumer@conundrumer·Nov 12

btw might be reinventing something but w/e i learn better from first principles and i think we can do more w the term "skill issue". also having witnessed ppl repeatedly trying to solve soul-skill issue w mind-skill

conundrumer avatar
conundrumer@conundrumer·Nov 12

example skill issue: i grossly mismanaged my time naively, mind-skill issue: i lacked a system for managing my time more accurately, soul-skill issue: i needed to let go of my desire for The Perfect Thing, in service of The Important Thing

i spent 2 years trying to solve arc-agi. mismanaged my time. rushed my submission to the competition. scored 0%. it wasn't a total waste, i learned a LOT, but rn i am rly burnt out so i'll share learnings another time. i just need to process this whole journey ...

conundrumer avatar
conundrumer@conundrumer·Nov 12

it may be helpful to unpack "skill issue" into its respective domains: - mind - body - soul i am well practiced in mind-skill, but am limited by body-skill and soul-skill. and apparent mind-skill issues are frequently emotional problems for me ie soul-skill issues

conundrumer avatar
conundrumer@conundrumer·Nov 10

I still believe in the theory and I may continue to work on it, but I have to grieve the possibility of getting rewarded for this work now that the competition for arc-agi-2 is done for good. thats ok. there may be another way forward ...

btw for these asking, this is literally our latest symbolic model it learns to encode human knowledge as a bunch of top-level definitions on the untyped λ-calculus. no weights yes, this talks to you. in English. at 1 million tokens/second how it performs? poorly for now :')

Post attachment