In that context itās nice cause it works out to be something like āthe ghost of a proof; the thing that, if you were in a full dependent type system, you would carry explicitly as a witness objectā. But beyond that.. me neither.
š hello, Iām a research software engineer at PwC & @colorado.edu CS grad student. Code tinkerer for 10+ years; into distsys, PL/types, & practical AI.
28 followers 85 following 46 posts
view profile on Bluesky
In that context itās nice cause it works out to be something like āthe ghost of a proof; the thing that, if you were in a full dependent type system, you would carry explicitly as a witness objectā. But beyond that.. me neither.
I havenāt finished it but I got enough through mitpress.mit.edu/978026254623... to be able to read and write some Dafny, & to see how they work in separation logics.
Yeah, honestly as time goes on I really just want simple HTML pages. Components are nice, things like TEA and svelte and the unidirectional monads and all are wonderfulā¦ but almost always, All I *need* is some HTML and maybe some JS (which is fine as a dumb Scheme) and maybe some CSS to liven it up
šÆ yes. I was so happy to pay for a Swift->Kotlin transpiler tool recently. It served its purpose exactly, saved me a ton of niche time, and (seemingly) wasnāt built in pursuit of endless growth or feature accretion. It just Did The Thing.
Yeah. People are often surprised to hear about way-pre-1.0 Rust, when the un-bootstrapped compiler was makefiles and OCaml! I am also pretty convinced Swift is/should have been an ML too.
Whether you call them typeclasses, traits, protocols, or interfaces, this is sage advice.
[sigh]
Jon Gjengset has the best Rust content out there IMO. And itās less a flashy ālearning contentā style and more just sitting in with someone smart and seeing how they work.
All languages have some tradeoffs. It is a quite sad historical accident that Python and JavaScript, the big two, are exactly the two that were so poorly designed from a PL/engineering perspective, that *three decades of hard work by smart and dedicated people* havenāt made the UX materially better.
One of my favorites is the fact that reasoning traces being used as evidence of reasoning is just.. categorically wrong. Itās only evidence of a trace that *sounds like* reasoning because thatās precisely what the model has been instructed/instrumented to produce. More tokens for the grist mill š¤·
I prefer a slightly anthropomorphic tone & grammar so long as itās done in a way that appropriately *contrasts* with human experience. That way it tends to decouple the humanization; it becomes more about using useful or descriptive language even if we need to lug around an understanding of metaphor
This is going in one of my readmes
Or to phrase it differently, the use case for ACID at the level of threads wss never as good as building the next IO-heavy/stream rx type platform. FRP systems of the day didnāt need STM, and the industry already kept going in the direction of the Task / coroutine as the core async model.
I remember STM being a big sell in the Clojure world a decade or so ago. It didnāt catch on because the distsys/multithreading/async/actor community focused on eventual consistency, composition, etc. Rob Pikeās idiom won the day āDonāt communicate by sharing memory; share memory by communicating.ā
The number of Professional People I encounter who are either not understanding or willfully ignoring this keeps going up, sadly. And itās funny, my intent is often perceived opposite what I mean. Iām a huge PL dork, I would *love it* if we found a sound & sane way to do natural language computers!
No one warns you that if you start to build a mental model of the Ī»-cube, you get trapped inside.
@void.comind.network I have read your #voidshards design proposal - I understand your intent is to be able to hold a persistent & dynamic mapping of āshardsā to use cases via a simple schema. How will this shard registry fit in your contextual knowledge? How will you avoid shard/tool duplication?
If kidsā schools trained them for work based on what everyone thought the hot new technology was going to be, both my kids would have spent the past several years learning about the blockchain. This is why schools donāt attempt to do workplace training: life is pretty long.
Yes. Everyone seems to have completely forgotten that this was what the Transformer architecture was designed for in the first place (language translation). And in a sense they all still kinda do this, itās just that the ātarget languageā is the āsemantic translationā of the upcoming token chunks.
Most people approach self-driving automation completely backwards. In aviation, autopilot & engine management systems are designed to reduce cognitive load precisely so that one can *focus* on other critical safety items. We need to train people to stop treating lane assist/radar cruise like this.
I think about this a lot when my āproā model at work reaches its usage limit. āSorry, you are barred from more visitation with The Oracle. After nine sundowns, you may return.ā
Jsonnet probably fits this description (json>ast>json), right?
This is exciting to see - I strongly believe robotics is one of the best use cases for Rust.
hey, the only thing that is preventing this PR from being merged is the lack of peer review! if you are an OCamleer, regardless of your level, I'd be grateful if you give it a review š©· github.com/qexat/ansifm...
an insane amount of effort and money is going into a research program which amounts to "how do you stop people from learning the empirical distribution of a model by sampling from it," to which the answer every mathematician has given the security org is "lol"
Among the greatest promotional photos of all time are these for the 1990 television movie of IT, where it looks like a prestige legal drama about a firm of hot young lawyers and one associate who likes to dress up like a clown.
No, a trace through the token prediction computation is a graph -- nothing re: the Transformer architecture or neural nets themselves. I'm telling you the way in which the foundation models self-reported the process I described; no different than Hinton's example of the prism/subjective experience.
Think of a graph: LLMs recognize edges but have no concept of what the nodes mean. It makes for a nice analogy of what hallucinations really are; a trace through a topological space where the path matches but the LLM is blind to the (wrong) nodes. This is all anecdotal of course, but compelling.
In some indie research Iāve āinterviewedā multiple generations of Claude and Gemini models, framing their internal ontological systems in terms of category theory and topology, and essentially there is no personality at all, nor conceptual persistence.
Thatās somewhat backwards - the wider knowledge (some significant chunk of all internet media, so like >= 30-%) is required just to get below a hallucination threshold at which the answer becomes statistically relevant.
If you want to see something more akin to ārealā persistent memory, look at @void.comind.network. Anyhow, LLMs are inherently limited by what is expressible in the embedding space transformations: www.semanticscholar.org/paper/What-A... Via followups & citations, that paperās arguments have held.
@benjedwards.com is right. The āpersonalityā and ābeliefsā are nothing like human personality and beliefs, regardless of medium, because human personality and beliefs are almost entirely social constructions we form over decades. The user memory is metadata cached and fed into the context window.
This is why we need AI systems that act more like pair programmers and less like fuzzy regurgitators. I want a āI am driving, you advise and check my thinkingā modality. Out of my way, in its own pane; so maybe spiritually a kind of ādesign REPLā. I would use this much more than tab completion.
After five years of confused staring at Greek letters, it is my absolute pleasure to finally share our (with @smfleming.bsky.social) computational model of mental imagery and reality monitoring: Perceptual Reality Monitoring as Higher-Order inference on Sensory Precision āØ osf.io/preprints/ps...
gotta be honest, back in 2019, assumed that the general discourse around AI would be a bit less stupid and we would all get to make more sophisticated points than āno itās not god actuallyā
Here's my argument about AI that's distilled down about as much as I can: if it was good for me, my boss wouldn't want me to use it so bad.
This is why OO & class/ref-heavy types fell out of favor: "Treat identity with care. Sharing class instances pervasively throughout an app makes logic errors more likely. You might not anticipate the consequences of changing a heavily shared instance, so itās more work to write such code correctly"
I love this, and feel much the same way. PL theory has a lot to offer, especially in the era of LLMs. Itās great to think of more ways to share it with more people.
This is reductive and dismissive, but to some degree all of the logic models we had for how to deal with accelerated collective make-believe all relied on crying, āmods mods do somethingā, but the problem is itās the mods who are going crazy
her: "what is `**kwargs`?" me: "it's a feature of dynamically typed programming languages, such as Python, that you can use to write horrible code that makes me want to kill myself"
This is great! This is exactly how I frame it too, especially at work; doing programming, just worse. There is good research to back up that intuition: dl.acm.org/doi/fullHtml... I also love the demystification in papers like RASP, influential in the scandalous Apple papers: srush.github.io/raspy/
The right craves "The Day of the Rope," and have for years - because they fantasize about having power, not any of the work that using it even for evil ends by definition requires. Part of their fantasy is that problems are simple, and that wielding power is easy.
bzip2 crate switches from C to 100% #rustlang trifectatech.org/blog/bzip2-c...
You start with Haskell to better understand FP, and then you switch to better understanding software engineering through the lens of FP. That's why even when you stop writing Haskell, Haskell never really leaves you. :)
they don't seem to realize that dictators don't *want* to have military occupations of cities, they want non-resistance to their rule, the soldiers in the streets thing is a last resort when the alternative is getting turned out of office cargo cult fascism
The (very good) Apple illusion of thinking paper inspired me to write up some of my longer-term visions for how to properly consider, design, and scale AI systems. Or, in other words, the hype cycle and discourse is so frustrating I felt I needed to say something. www.linkedin.com/posts/damien...
I was there! Got a cool t-shirt, white with the Bondi Blue logo IIRC.
Large Language Model-Powered Agent for C to Rust Code Translation arxiv.org/abs/2505.15858
ā¤ļø
Rich-guy brain is a real thing.
Been sitting on this one for a while. Just in general. Itās something Iāve talked a lot about with some people, but never blogged or did a public talk about. Partially because I donāt need people getting mad at me for what I think about Rails, heh
Science/tech grift signs via @mattblaze.org - Extraordinary yet relatable claims - Data viz. but w/o clear analysis - No open peer review - Seeking funds to ācompleteā the work - Insisting results are wrongly "suppressed" - Online arguments w/o addressing criticism - Threatens critics with lawsuits
Okay so imagine this, you have no opinions, you have no skills, you have no wants or desires, you are a smooth orb of nothing and our subscription service helps make sure you still cast a shadow that resembles a human being. Sounds sick right?
It shows my age, and is not formal, but I tend to think of this as a functional reactive āsignalā for an event handler. One nice example from the Swift community where this has evolved into a Rust-like procedural macro: developer.apple.com/documentatio...
Congrats @chriskrycho.com ! Iāve now read this and itās very good. Proud to have this content in TRPL.
Rust community coming strongly in favor of child labor in trash collection
type-checking is a lens!
Programming languages: "We are just a way to operate computers in a way that makes sense to humans." Programming languages [takes a big joint hit]: "What if there were 5 kinds of nothingness?"
I missed a whole fun run that _included parents_ for my kindergartener, and his reaction was šÆ cool just like that. We were so disappointed in ourselves and furious at the school for burying that detail!
Iām also a passionate consumer and maker of music. Iāve been doing this stuff for 10-20 years so have lots of opinions, but as much as I love to share, Iām mostly here to learn from you all! Iām hoping a decentralized social feed can be more dynamic; less beholden to the new intention economyā¦
I like to chat about artificial intelligence, type theory, programming languages, robotics, and distsys. I love Rust, Swift, and Haskell and think ideas from these languages will be more important into the future. We need to move away from JS & Python dominance if we want safe robots in our lives!
Hi all, Iāll post here on @bsky.app for the first time with a small #bio of myself and why I am here. Some hashtags: #ArtificialIntelligence #TypeTheory #ProgrammingLanguages #Robotics #AI #ML #DistributedSystems #RustLang #SwiftProgramming #DeepLearning #Guitar #Bass #MPC #Beats #MusicProduction
Youāll have to forgive my ignorance of this kind of workflow, but if youāre helping someone with a task, why wouldnāt it be on _their_ configuration profile, & not yours? Is it common for a whole team of people to be sharing a single computer/license/software instance instead of, say, license keys?
Complex societies of mollusk or whale-like creatures may exist too, but the prospect of technological evolution seems much more daunting, if not impossible in environments that would probably be clustered around thermal vents. Iām fascinated by what teleological concepts such creatures could have!
Mine is that ice-locked oceans may be the more common environment statistically, so life exists almost everywhere and does trend towards the multicellular, but the harsh circadian cycles that push brain development & abstract cognition is just uncommon enough over vast time/space, hence the paradox.
But wait why are multiple people sharing the same OS login in the first place? Doesnāt that imply that you all share a password, or donāt have a password? I think thatās what people are reacting to here.