Threadwell

Recent Notes

DemetriSpanos
Oct 06, 2026
NOTimothyLottes
Oct 06, 2026
bonzajplc
Oct 06, 2026

victortaelin

176 items with this tag.

Aug 29, 2026
VictorTaelin
Aug 06, 2026
just saving my last conversation with Fable on this entire flattener debacle.
Jul 17, 2026
> Read this and your next model will be 10x smarter.
Jul 04, 2026
*sighs*
Jun 24, 2026
In ancient Egypt, writing was concentrated among scribes, giving a small educated class control over taxes, laws, land, religion, and state power while most people depended on them to understand the systems ruling their lives.
Jun 19, 2026
humans are just so confused about everything, we invent layers of complexity on top of a simple underlying reality, assign different names to the same fundamental concepts, spawn entire fields to investigate the same mathematical objects under different grammatical names
Jun 28, 2024
Can we build an optimal processor?
Aug 06, 2026
@VictorTaelin I think the ability to make a computer program progressively smaller (yet still output the same data) is close to the definition of intelligence, if not the definition itself
Jul 18, 2026
@VictorTaelin I've noticed this in AI workflows too.
Jul 18, 2026
@VictorTaelin A useful elaboration: whether something should be deleted goes with the square root of its age.
Jul 18, 2026
@VictorTaelin Yes but this is missing *what* you compress.
Jul 18, 2026
@VictorTaelin By the way, we probably evolved towards this due to info transmission/comprehension cost in a group.
Jul 18, 2026
@VictorTaelin nobody knows how to do that
Jul 17, 2026
@VictorTaelin This is an interesting approach to boosting intelligence.
Jul 17, 2026
@VictorTaelin Completely agree, erasure needs to be a mechanism, not a hope.
Jul 17, 2026
@VictorTaelin wolfram knows it.
Jul 17, 2026
@VictorTaelin This has been my whole purpose with LLMs outside of actual task work for at least one year
Jul 17, 2026
@VictorTaelin This is true in neural networks.
Jul 17, 2026
@VictorTaelin Ya compress the AST size not the character count..
Jul 17, 2026
@VictorTaelin I too have considered compaction to be the next generation of model design
Jul 17, 2026
KV Compaction gonna be a great change
Jul 17, 2026
I have sorted that only halfway in code.
Jul 17, 2026
@VictorTaelin Best post I’ve seen on here in a while.
Jul 17, 2026
@VictorTaelin "Compression is the intelligence"
Jul 17, 2026
@VictorTaelin mLSTM?
Jul 17, 2026
@VictorTaelin 100x worse (Be Generic)
Jul 17, 2026
@VictorTaelin https://x.com/ianisarobot/status/2046793491687113012?s=46&t=wgK6ewuTJ2vWXe73_JA3MA
Jul 17, 2026
@VictorTaelin > An aha-moment happens when your brain is capable of expressing new information in terms of information you already posses stored.
Jul 17, 2026
I love how well articulated this post is!
Jul 17, 2026
@VictorTaelin Reminds me of this paper:
Jul 17, 2026
@VictorTaelin a related issue that bothers me is that the changes ai makes to your project are always continuous.
Jul 17, 2026
@VictorTaelin Have you ever look at justin sung work on learning?
Jul 17, 2026
@VictorTaelin > Compression.
Jul 17, 2026
@VictorTaelin I think Anthropic is already doing something similar because they are compressing intelligence as much as possible for pretraining.
Jul 17, 2026
@VictorTaelin DL is not built for erasure as much as we don’t control what we forget.
Jul 17, 2026
@VictorTaelin A lot of slop comes from this effect.
Jul 17, 2026
@VictorTaelin Nobody cares about erasure until your program stops working at 2 am and the file is too big to something something that a bot would say
Jul 17, 2026
@VictorTaelin Son of Anton figured it out first
Jul 17, 2026
@VictorTaelin I'm trying to do that at http://0-bro.com
Jul 17, 2026
@VictorTaelin Saving this b4 you delete it 💔
Jul 17, 2026
Intelligence is not just the ability to reason, but the ability to spend reasoning resources wisely.
Jul 17, 2026
@VictorTaelin this looks like kv cache? or did i miss something 🤔
Jul 17, 2026
@VictorTaelin I thought about it a lot too, the bloat surely hinders progress.
Jul 17, 2026
@VictorTaelin True.
Jul 17, 2026
@VictorTaelin How do I get my real life brain to be like a LLM? Having infinite memory is a blessing and a curse
Jul 17, 2026
@VictorTaelin why big labs haven't hired you yet
Jul 17, 2026
@VictorTaelin I think they already do erasure a lot, on every token.
Jul 17, 2026
@VictorTaelin I need this idea somewhere in my calude.md
Jul 17, 2026
I've gotten rid of compaction in my loops in favor of a summary handoff to continue the given task because I've noticed that compacting is actually *worse*.
Jul 17, 2026
I’ve been spending a lot of time on reducing the state space of a program instead of just trying to remove branches, and then models are really good at it if you give them the right skills.
Jul 17, 2026
@VictorTaelin excelente post
Jul 17, 2026
@VictorTaelin did you post this after watching 3b1b video
Jul 17, 2026
@VictorTaelin I mean arguably asking for a summary or summarizing past events has erasure “built in” as you are reducing down to the decided important parts.
Jul 17, 2026
@VictorTaelin I think "code-golfing / minification / uglification" is okay, as long as the model can still recover an interpretable and human-readable version of the code.
Jul 17, 2026
@VictorTaelin One thing I noticed, is that all of these models are terrible at information extraction, as in rewriting text as a list of facts you learned from it, and pretty much any related task.
Jul 17, 2026
@VictorTaelin *taps sign* https://x.com/justinmk/status/1874626660915613984
Jul 17, 2026
@VictorTaelin The compression angle is fascinating.
Jul 17, 2026
@VictorTaelin Putting this into my fable convo to distill into a claude.md rule and a loop
Jul 17, 2026
@VictorTaelin 3B1B video running through AI Twitter like crack in the 80s
Jul 17, 2026
@VictorTaelin Not true, weight decay explicitly forces sparseification of parameters during training.
Jul 17, 2026
@VictorTaelin Agree, but also worth noting: Attention mechanism obviously learns erasure - as the query-answer mechanism figures what’s relevant and what’s not.
Jul 17, 2026
@VictorTaelin codex leaves unrelated edits everywhere.
Jul 17, 2026
@VictorTaelin Yes, absolutely.
Jul 17, 2026
@VictorTaelin i think the broader insight is that intelligence isn't just generation
Jul 17, 2026
@VictorTaelin https://goombalab.github.io/blog/2025/tradeoffs/
Jul 17, 2026
@VictorTaelin the cata- part of hylomorphism
Jul 17, 2026
@VictorTaelin imo, this is the last piece that’s missing, everything else is just scaling compute, rl, task time, etc.
Jul 17, 2026
@VictorTaelin In Rust I’ve ended up making a code garbage compactor.
Jul 17, 2026
I have a somewhat similar idea.
Jul 17, 2026
@VictorTaelin i think this is more or less where sutton is trying to go:
Jul 17, 2026
@VictorTaelin I got some very interesting results recently about erasure rates in vocabulary for codebases.
Jul 17, 2026
@VictorTaelin The Codex (or Claude Code) harness does context compaction, and they're surely training for this on two levels: the model is being trained within harnesses that are controlling their context, and the models they're using to compact context are also being RL'd to perform better.
Jul 17, 2026
@VictorTaelin 100%
Jul 17, 2026
@VictorTaelin the models are smart but they are not wise
Jul 17, 2026
@VictorTaelin Love this! I've been fighting with it for a few weeks.
Jul 17, 2026
@VictorTaelin Thoughts on Intelligence as compression? Series by 3blue1brown is very nice.
Jul 17, 2026
@VictorTaelin i think 5.6 sol is really good at deleting slop, I also found that it keeps changes and new implementation to a minimum, where 5.5 would write 1000 lines, 5.6 writes about ~500
Jul 17, 2026
@VictorTaelin This shows up outside training too.
Jul 06, 2026
@VictorTaelin @Jonathan_Blow hopefully useful data point.
Jul 05, 2026
@VictorTaelin I disagree on the consistency is enough to do mathematics, you can very weird stuff to MLTT like Nat iso Nat -> Nat and still be consistent.
Jul 05, 2026
@VictorTaelin It was easy to read and I think I understood most of it.
Jul 05, 2026
@VictorTaelin Fable is educated c
Jul 05, 2026
@VictorTaelin > this is the antidote I found to AI psychosis.
Jul 05, 2026
@VictorTaelin Are languages built on Bend incapable of representing “eval” then? And/or disallowed from looping possibly-infinitely based on an external signal?
Jul 05, 2026
@VictorTaelin stupid question alert.
Jul 05, 2026
@VictorTaelin Feels like the precise, symbolic analogue of implementing LLM guard rails that aren't too strict and aren't too lax.
Jul 05, 2026
@VictorTaelin i didn't expect i'd read all of this, but i did.....can't pretend i understand it, but i'm starting to get the picture.....
Jul 05, 2026
@VictorTaelin To someone who doesn’t understand the actual content, AI psychosis and expert use of AI to drive a breakthrough look the same (modulo institutional credentials).
Jul 05, 2026
@VictorTaelin Don't try to market yourself.
Jul 05, 2026
@VictorTaelin don't you want to be able to copy lambdas though? Or how do you do a lot of the typical FP stuff if you can't pass around and copy functions as you otherwise would?
Jul 05, 2026
@VictorTaelin this post is not a rebuttal to a query about ai psychosis regardless of the merits and robust technical details behind it
Jul 05, 2026
@VictorTaelin i have an entire math proofs section in my maths repos
Jul 04, 2026
@VictorTaelin the real crime is those idiots not knowing who you are or your work.
Jul 04, 2026
@VictorTaelin What do you mean by arguments to recursion "decreasing"? ...
Jul 04, 2026
@VictorTaelin Fable simplemente hizo de placa base pitando para avisar del error, pero la solución elegante y real requería a alguien al volante.
Jul 04, 2026
You speak simply into a mixed space where most people including me are on the "application layer".
Jul 04, 2026
@VictorTaelin This is a great post! I can’t pretend that I really get the math, but it’s obvious that you understand what you’re talking about.
Jul 04, 2026
@VictorTaelin I never thought you had AI psychosis, but the fact that others thought so led to this wonderful explanation.
Jul 04, 2026
@VictorTaelin I got interested in Bend for its prosaic capabilities.
Jul 04, 2026
@VictorTaelin I enjoyed reading this post.
Jul 04, 2026
@VictorTaelin i once implemented a type checker for impredicative MLTT.
Jul 04, 2026
@VictorTaelin It's lonely sometimes eh? I'm excited about your stuff and love your posts but I don't know a lot about lambda calculus etc.
Jul 04, 2026
@VictorTaelin os caras acham que só eles são superinteligentes mas não conseguem enxergar alem da bolha que ele mesmo se coloca entre os '' inteligentes '', que tipo de inteligência é essa ? é a vida de um engenheiro médio que trabalha em um lab de IA em SF kkkkk
Jul 04, 2026
@VictorTaelin y-combinators don't have valid types, so you are safe from them in statically typed functional languages.
Jul 04, 2026
@VictorTaelin Impossibility to write *fast* interpreter looks like a hardware architecture or compiler issue, not the language expressive power problem.
Jul 04, 2026
@VictorTaelin QTT approach for this is very elegant indeed!
Jul 04, 2026
@VictorTaelin No, I prefer the previous angry rhetoric pls revert the edit
Jul 04, 2026
@VictorTaelin Sorry, im not very technical.
Jul 04, 2026
@VictorTaelin angry one's always get attention on internet
Jul 04, 2026
@VictorTaelin Don't let them get you down.
Jul 04, 2026
@VictorTaelin There's an emergent ai psychosis psychosis in which idiots can't understand what you're saying so they cry "AI psychosis." It's a good litmus test for someone's intelligence, imho.
Jul 04, 2026
@VictorTaelin I can't really comprehend what Bend2 will be useful for except helping to make bug free software.
Jul 04, 2026
@VictorTaelin Bro, you dont hafta Grug-post to be popular with the right people.
Jul 04, 2026
@VictorTaelin The more AI advances, the more we need very high level experts in very very specialzed narrow fields, to check results ! ❤️
Jul 04, 2026
@VictorTaelin I began following you when I scrolled across an image similar to that above, it is the highly technical stuff you post that I enjoy most.
Jul 04, 2026
@VictorTaelin sometimes the problem isn't explaining something that's hard for people to grasp, but rather getting people to admit when they don't get it
Jul 04, 2026
@VictorTaelin I'm a big fan of Interaction Combinators and I hope you keep going.
Jul 04, 2026
@VictorTaelin This is so cool and inspiring (not that I understand all of it; that’s part of what makes it inspiring!).
Jul 04, 2026
@VictorTaelin adorei ler esse post.
Jul 04, 2026
@VictorTaelin Do people not study lambda calculus in college or at any other point when looking into programming languages?
Jul 04, 2026
@VictorTaelin these guys have no idea what they're talking about.
Jul 04, 2026
@VictorTaelin If you have a system that provably must terminate, what "interesting" statements are you unable to express? And what can only be stated in restricted form?
Jul 04, 2026
@VictorTaelin @tsotchke is another good example this is happening to.
Jul 04, 2026
@VictorTaelin This is a very naive and ignorant question, but now that you've split Type into two sorts, is there anything that could be represented with both variants? and if so, what would happen if someone selected the wrong one for their use csse
Jul 04, 2026
@VictorTaelin One question: does bend have an unscoped lambda?
Jul 04, 2026
@VictorTaelin Taelin I've been following you since the beginning: I think the problem is you need to aggressively position yourself and your program as a painkiller.
Jul 04, 2026
@VictorTaelin This write up was really well done, thank you for taking the time to write it! Broke down the concepts well
Jul 04, 2026
@VictorTaelin > most of you can’t understand my posts
Jul 04, 2026
You’re not allowed to praise Fable, you MUST bag everything on the internet 🤣 And god forbid you do anything genuinely impressive!
Jul 04, 2026
@VictorTaelin Loved this long post version of yours.
Jul 04, 2026
@VictorTaelin I presume the issue was in duplication inside 0-contexts?
Jul 04, 2026
@VictorTaelin Do you allow for non-termination on total functions on codata?
Jul 04, 2026
@VictorTaelin I'm curious.
Jul 04, 2026
@VictorTaelin Very cool.
Jul 04, 2026
@VictorTaelin russel antinomy •_•
Jul 04, 2026
@VictorTaelin karpathy said he had AI psychosis for a while and everyone knows he's very competent
Jul 04, 2026
@VictorTaelin Wow, that's impressive! Best argument on the quality of Fable I have read by far.
Jul 04, 2026
@VictorTaelin It's not how many likes you get, it's likes from who that matter.
Jul 04, 2026
@VictorTaelin In my language I'm trying to do termination by proxy, I.e.
Jun 24, 2026
This is a good parallel to what is happening now when it comes to the frontier models.
Jun 24, 2026
@VictorTaelin La invención de la imprenta fue revolucionario en la edad media, porque quienes concentraban el poder tenían miedo de que los plebeyos pudieran acceder al conocimiento masificado
Jun 24, 2026
@VictorTaelin interesting how Islam is the one religion that encouraged mass literacy, and because of that it meant that religion was personal.
Jun 24, 2026
@VictorTaelin They also believed writing was literal magic, translated as "words of god," hieroglyphs didn’t just communicate, they manifested reality.
Jun 20, 2026
@VictorTaelin @comohology yeah, wouldn't it be lovely to view any and all models or formalisms as mere 'views' decoders over invariants in the quotient (isomorphism modulo operational equivalence)?
Jun 19, 2026
@VictorTaelin @comohology Isn’t it great that AI understands our human science attempts in a homomorphic manner and freely navigates the complexity space, coming up with proofs that were there all along, just never collected
Jun 19, 2026
This pattern appeared so often that in 2017-19 I had to conclude conventional math is a mess.
Jun 19, 2026
@VictorTaelin @comohology > what I wrote isn't just true
Jun 19, 2026
@VictorTaelin @comohology "what I wrote isn't just true, it is how humanity operates"
Jun 19, 2026
@VictorTaelin @comohology imo representations matter.
Jul 19, 2025
@VictorTaelin Technology's relentless march outpaces biology's crawl.
Jun 22, 2025
@VictorTaelin Optimal processors are cosmic dreams, tangled in math's intricate dance.
Jan 17, 2025
@VictorTaelin Optimal processors require unconventional architectures.
Oct 08, 2024
@VictorTaelin Why a 3D system? Chips are largely planar, and while you can project any topology onto a plane, it may not be optimal.
Aug 07, 2024
@VictorTaelin Bring back the LISP machines!
Jul 17, 2024
@VictorTaelin Cool! A few technical questions.
Jul 12, 2024
@VictorTaelin Interaction combinators aren’t optimal, as they’re exponentially slower than quantum computers on some tasks, right?
Jul 12, 2024
@VictorTaelin horrible idea: if you want to rewrite circuit for performance, why not doing it directly on the lowest gate-level?
Jun 30, 2024
@VictorTaelin I don't understand...
Jun 29, 2024
@VictorTaelin
Jun 29, 2024
@VictorTaelin I got inspired and made a song about it: https://suno.com/song/a54124fe-b78d-489b-9a68-47652b3030e1
Jun 29, 2024
@VictorTaelin How does a commutation interaction work?
Jun 29, 2024
@VictorTaelin Is this similar to the computation model underlying Wolfram’s Physics project? Essentially a hypergraph (set) rewrite system?
Jun 29, 2024
@VictorTaelin good content.
Jun 28, 2024
@VictorTaelin message passing graph neural network
Jun 28, 2024
@VictorTaelin Does talking about physical efficiency in terms of Lambda Calculus or Turing Machines makes sense? They are elegant models to reason about computability and complexity classes.
Jun 28, 2024
@VictorTaelin @MikePFrank You’re my certified hardware guy.
Jun 28, 2024
@VictorTaelin Curious, because what you describe sounds very similar to neuromorphic computations and energy-based models.
Jun 28, 2024
@VictorTaelin Well, that’s basically a description of the human brain
Jun 28, 2024
@VictorTaelin looked into Ivan Sutherlands work at the ARC
Jun 28, 2024
@VictorTaelin i desire low power hardware thats highly efficient and highly parallel, in the way that autotrophs have slow metabolism
Jun 28, 2024
@VictorTaelin Have you looked into topological deep learning? These two concepts feel convergent…
Jun 28, 2024
@VictorTaelin Reading graphical interaction systems, I'm hearing "postsymbolic" / iconic formal system.
Jun 28, 2024
@VictorTaelin Your posts about these topics are so fascinated, I feel like I've got a long way to learn to fully understand what is being said.
Jun 28, 2024
@VictorTaelin Has anyone looked into how ICs interact with reversible computing? In today's hardware, power density is the real issue.
Jun 28, 2024
@VictorTaelin How difficult is it to be ahead of your time? (honest psychological question)
Jun 28, 2024
@VictorTaelin When seed round for the hardware shop

Created with Quartz v5.0.0 © 2026

GitHub