LR

15 May

new LR's Devlog¹ entry for 15.5.2026 git.sr.ht/~duangle/tex... ¹ formerly known as the "ad-hoc devlog" #devlog #sxpp #nudl #tukan #nowhere #duangle

problem: your voters learn from AI what a fraud you are solution: 1. change term from AI to SI. 2. roll out "SI" trained in your favor. 3. get your lemmings to use SI before they find out what a fraud you are. 4. oh shit too slow, midterms are here

LR reposted

Тsфdiиg

@tsoding.bsky.social

17h

Тsфdiиg@tsoding.bsky.social

Rederedereference

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

but sorry, even though i know corporations are made of people, i just can't find it in my heart to like them and their gilded cages, these dependency structures they're trying to lure us into. i'll take their birdseed though if it is placed far enough from a cage.

so yeah, things got quite complex since i was 12 years old. i'm not sure i like it. is it experience or trauma? i'm still figuring that out.

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

i am set in my ways, stubborn to change, and it takes a quiet kind of peer pressure for me to become restless and try to make my way, but i have to do it in my own way. i have particulars now! i have distinguished taste! i don't stay up until 6 am, i do not sleep under tables at conventions anymore.

one thing hasn't changed though, and that's my almost pathological self-sabotaging stance towards commercial exploitation, even though that exploitation keeps bringing me gifts.

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

and i suppose you can try to find your own threshold on that curve - "when is the point that i switch allegiance?" - when i was young, that point was close to the foot of the S-curve. i was going to places before anyone else did.

i knew nothing and so i was fearless and full of curiosity, and also, my elders encouraged me. but now i'm in my 40s. i'm about approaching the middle of the S-curve. now i am part of the majority of people...

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

and so you just get jerked back and forth between these two things: stay here for a while, no we all go there now, well i'm not going, shit everybody is going, well guess i'm going too, but now we'll stay here for a while.

i guess that's a pretty good informal description of an S-curve; and plotted as movement, a bell curve distribution along the time axis.

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

B) but conversely, if you want the hottest newest thing you're going to run into diminishing returns invariably, because new insights arrive in quants. it's not a steady tap of new inventions, it's small breakthrough events whose outcomes have ramifications of scalability.

and so you can only finetune and scale for a whole and thus fill out the local optimum completely, until it gets stale - and then you're at A until the next bombshell drops.

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

and i guess the urge of trying to keep it all together is only human, but so is the excitement of discovery. and given the scale at which humanity operates now, i don't think either of these two groups can win, plus people "switch allegiance" all the time.

A) you try to pin down your culture to a certain expected state and invariably it gets corrupted because it is a fixed target. it is thoroughly known and thus exploitable.

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

the past 20 years we were constantly cautioned against being too hyped about the future, what on account of all the people that progress overrolls - they see progress rollin, they hatin

and this is also not the first time this happened though more in the other direction. the same thing happened in the 70s, and it was the conservatives who stepped on the brakes and canceled a dozen research projects, some of which created things like... let me remember... ah yes, the internet

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

effectively a loose generalization of moore's law to technological progress as a whole, and it turns out... he was probably right. we also just got gene therapy.

"welcome to the ai psychosis party" i wish it was a psychosis. 🫤 that would be curable. i did have an actual psychosis in 2020 so i can directly compare.

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

and suddenly it's pandemonium. the monkeys are going crazy. the historical lineage should be: nuclear bomb -> LSD (which timothy leary famously described "as significant as the nuclear bomb") -> transformers

leary also had a fascinating talk about how progress grows exponentionally, he used the unit "jesus" to measure it, using sentences such as "by this time we've gone from 2 jesuses to 10 jesuses"

who would have thought that mathematicians and information theoreticists make the fiercest weapons. everyone. everyone would have thought that.

we just can't help it. you dabble a bit and whoops you accidentally created *checks notes* an append-only machine with an appending tape head following a (2^18)(2^18) -> 2^18 probabilistically compressed state table, plus a content-addressed oracle that retrieves from its own past tape.

23h

@mariozechner.at i'm trying to make my TTS extension stop playing as soon as i start typing but there seems to be no event for this? we looked at an existing extension that claimed to have this, pi.dev/packages/@be... but it turns out, it's an exaggeration - it only stops when pressing return

lrlrlr.neocities.orgT-Minus: The AI Bubble Pops

@edzitron.com i made something to cheer myself up and maybe you like it too? bsky.app/profile/lr.d...

LR@lr.duangle.com

lrlrlr.neocities.org/pop/

lrlrlr.neocities.org/pop/

lrlrlr.neocities.orgT-Minus: The AI Bubble Pops

@mariozechner.at @me.snemarch.dk meanwhile i figured out the correct framing for my context compactor so the model uses it extensively: subagent launching! we pretend we're the subagent when the checkpoint is set, and when subagent_result is called, we compact, and the model buys the deception :)

@me.snemarch.dk you mentioned beads and i thought that was a way to speed up context rollback, but it's a ticket system?

this anti-AI post's coding experience with claude and subsequent distrust is precisely why i'm not worried about them taking our jobs: mastodon.gamedev.place/@lritter/117...

LR reposted

celestialmaze

@cmzw.bsky.social

ziptie nodes v0

merman - the executable mermaid github.com/paniq/merman

bsky.app/profile/lr.d...

David Bushell 🪿@dbushell.com

here we are

trickle down economy is not self-forming. you need cunning highwaymen in green tights drilling holes into their coffers from below.

canis lupus wallstreetii: sell me this pen me: *runs away with the pen, uses pen to write award winning trilogy* thanks for the pen, sucker canis lupus wallstreetii: well done

the pen is mightier than the short

@fasterthanli.me did you see this? github.com/slee-persis/...

@mariozechner.at feature request for pi, since i struggle with short contexts here a lot. no idea if this will work.

LR reposted

innes mck

@innes.bsky.social

if you're into some of the worst people you know having a terrible time: airbnb has changed it's rules so owners can no longer require guests to clean up, do dishes, take out bins etc, and as you can expect they are all being super normal about it

block.xyzDessa - the Conversation Model trained only on speechYou learned to talk by listening. So did Dessa. A research preview of a full-duplex speech model.

@fasterthanli.me aww you locked the thread... i was gonna add he's gonna be real pissed when dessa leaves him for alan watts... again!

LR reposted

Amos Wenger

@fasterthanli.me

block.xyz/dessa Yeah this is gonna do more damage than GPT-4o 🫠 It's truly impressive stuff, don't get me wrong, but Jesus Christ, if you think people had AI psychosis before now...