Remote
Kevin Graaf
@kevin@fedi.kbg.sh
mastodon 4.7.3Computer scientist and journeyman grapholinguist. Visiting Fellow with the Visual Interactions in Early Writing Systems (VIEWS) Project at the University of Cambridge. New Yorker in exile.
Views are my own.
269 Followers
705 Following
30 Posts
Joined October 20, 2025
I think a lot of people want to draw a distinction between ML and LLMs or “GenAI” the latter of which I don’t think is a scientifically useful category. I must admit I am basically a total skeptic of ML for nearly all applications, because in my mind, such methods mean largely giving up on theory building. Yes these can be useful, certainly as a crutch, but never as much as deterministic theories or at least direct theories of object phenomena because ML models never explain anything
The worst argument from the pro-LLM crowd has to be the accusation of “shifting goalposts.” It is so obvious that new technologies can clarify positions, reveal deeper commitments, shift perspectives, raise new questions, and answer others. Before LLMs, many uses of (for instance) computer code were coterminous in motivation. Now they are not.
Replying to
And I’m sorry to say but Linux itself has been a disaster for OS research and development. It’s 2026, and all we have to show for it is a clone of Unix, an utterly unremarkable early minicomputer OS, a design old enough to be my parent. If LLMs are so great, where’s the new interesting work? Where’s the groundbreaking new kernels? This field is an absolute joke sometimes.
I have NEVER wanted an object more
Replying to
Hey everyone, just wanted to say I’m very appreciative of the positive responses! I’m currently finished up a conference trip/sort of am on vacation, so I’ll try to respond to everyone tomorrow
Doing some deep dives on structured text editing. Is anyone aware of any work on *data structures* for editing tree-like data? I’m specifically interested in stuff like multiple cursors, efficient undo/redo, etc. This is directly relevant to my research re: Telecode (which is a tree, or hedge rather), but I’ve not been able to find any papers covering data structures. Sort of unclear to me whether ropes and other usual suspects generalize well to a structured domain
Confession: I mostly hate command lines and I think the obsession with them in the tech world is basically a cultural signifier and little else. I constantly see obtuse command line interfaces that are worse in every way than simple GUI would be, but are preferred, I guess, on the grounds that that’s what the epic bigshot “coders” use.
My recent talk on computerizing hieroglyphic scripts just got posted online! There’s a demo of my Maya font in the presentation (with of course, a few typos I realized I made—but this is interesting in itself, the world’s first Classical Maya typo?) I hope you find it interesting and please feel free to ask me any questions you might have.
Kevin Graaf: Computerising Hieroglyphic Scripts
Replying to
@pigworker every single success of “AI” is a failure to abstract. LLMs are merely decompression algorithms, if they work it is only because what they are being tasked with producing is not itself adequately compressed
Replying to
@jonmsterling@mathstodon.xyz @mevenlennonbertrand@lipn.info at the risk of sounding conspiratorial, it’s hard not to see Google and Microsoft attempting to de-federate email under a paper thin pretext of “security”: I have my personal email under my own domain and it’s 50/50 whether it goes to spam, and I do have all the DNS security stuff set up right.
Replying to
@whitequark oh, very cool. I wonder if now would be a good time to revive my IRCaudio extension I was working on for Mumble-esque audio channels
Replying to
@liamoc @jonmsterling I am a daily emacs user, but I really cannot stand this hair shirt culture (in fact I just posted about this) where you have to use the most obtuse, arcane tools out of some kind of mass hazing ritual.
Replying to
@mathias@rhizospherelabs.com if I’m understanding you right it seems like you want a SPICE simulator. LTspice is an OSS implementation
Replying to
@erincandescent@akko.erincandescent.net side note but I really liked cohost a lot, it was by far the coolest and most interesting of the Twitter replacements in my opinion
Replying to
@liamoc @jonmsterling yeah I use Unix (NetBSD) because there’s no other option, but I can’t say I particularly like it. I think the Unix Hater’s Handbook was essentially right then and now that the Unix culture more or less rewards obscurity and poor design. Generations of programmers now seem unaware you can actually just design good interfaces.
Replying to
@mnl perfect! Yeah, I was thinking the same thing, CRDTs are probably the place to start looking. I’m still sort of unclear on how CRDTs interact with more elaborate tracking of undo/redo like emacs has. And even more unclear when we start talking about editing history (eg if I accidentally ctrl-V my social security number, I do want that scrubbed)
Replying to
@hipsterelectron@circumstances.run I’m actually very curious how that would have manifested with respect to the structure of the script
Replying to
@adele this seems like a fantastic project! Do you have any screenshots or a publicly accessible instance? I’d love to see what it looks like in action. Great work!
Does anyone have experience migrating fediverse instances from one server software to another? I’d like to move mine from mastodon itself to snac2, but I want to keep my followers and so forth intact. Is that actually possible?
@whitequark this is awesome! Which daemon are you using?