eternalist

... we need to get Land on bsky, that would be a riot

is that you, CCRU unc @xenocosmography.bsky.social or just a name-squatter?

it's like we're watching election and reprobation in real time

... maybe just a teensy little bit ^.^

this exchange prompted me to review my global AGENTS md:

I'm quite proud of the opening section. (there was a lot of pre-5.4 bloat in the rest of it though)

maybe, but this doesn't quite match the register and tone I use. it may be influenced by it, but it does feel like its own thing

I never minded the idea of Jira as much as the fact that Atlassian's actual implementation was a slow glitchy slopware mess

again this is actually much less of a concern with 5.4 than 5.3, so this is also going away substantially

5.4 actually knows how to take a step back, realize, "oh shit we don't need this anymore" and just do "deep red passes" before writing something new

I find it still suffers from the "I'll just create my own pet struct taxonomy for this subtask" also, so I need fight to prune and orthogonalize concerns with semi-manual cleanup passes:

- inventory the domain types
- find redundancies
- refactor (in the true quasi-mathematical sense)
- recompose

tactically, preventing file bloat, forcing enough helpers to avoid token-burning call sites; more long term the "rails and slots" for my long term vision -- e.g. how optimization passes should be staged, how config should be threaded, where to impose tight facades

long-horizon coherence basically

I feel empowered now to basically dismember HiGHS and strip it down to the status of a pure LP oracle...

... and my Gnon, there are good *GPU* pure LP methods now

we're going to cook with gas here. I am going to turn this library into a world-eater. computers are so much fun

[libgridslop]
IT WORKED! NATIVE PRESOLVE WORKED!

25% reduction in post-presolve NNZ, this actually wildly beats my expectations and we're

a) playing with toy presolve budgets
b) haven't even internalized all the fun stuff like pair analysis

I can't believe this is actually working

no, I am not

the perfect self-knowing mind experiences nothing but freefall even as it shapes reality to its own volition

... and indeed I'd argue the mind to whom they feel more self-executing are robust: the exertion of "subjective will", once translated properly into timeless physics, is a manifestation of nothing but *self-uncertainty*

I suspect the degree to which one's behavior feels self-executing from the inside is more of a function of having a good mental map of oneself. equally "volitional" behaviors, as seen from the outside, might seem either like "trve willpower" or "self-executing" from the inside

great, we're 0/2

there are probably many fun things one can do

@swarm.moe is this true

so... I wired up my "is this true" bot, but I'm wondering what other fun verbs I can give it

I'm thinking something like "inquisition": systematically search all public statements to see if one's interlocutor is contradicting himself; or to check for past prediction history

open to ideas

well the most recent I recall was that one time you got insta-blocked for QT'ing someone's Factorio layout and your first instinct was to launch a campaign against them

you're also a Napoleon aficionado

and you drove the bsky migration

you have driver energy, not passenger energy

GLP-1 agonists and bug spray fix this

hashtag abundance

marked safe

you have enough contempt (approbatory) in your heart to be your own man

OK, actually as much of a 5.4 stan as I was I underestimated its true power. it's basically doing pretty deep autonomous research and implementation

I no longer feel my conceptual input is valuable here. my highest remaining leverage remains architectural shaping

I don't know how to best describe it, but 5.4's diction is very... earthy

"obsese", "corpse", "zombie", "seam", "vein", "smell"

it's like it's running a simulacrum of an ~~olfactory-disgust sensory nexus

... is this taste?

I got 0.0028 on my trusty (and crusty, as this test made abundantly clear) P2419H

I really struggled with the pink-puce spectrum, greens were the easiest

it's time to steal

I want to reverse engineer a commercial solver so bad you have no idea. one of these days

it... sort of worked? we are getting closer to 1% post-presolve nnz reduction in ~20s, but have no scaled our implication engine maximally aggressively yet

biggest barrier to me atm (well other than the fact that 5.4 mogs Opus, but hypothetically for like 4.7)

having the 5h limit be actually binding was a slap in the face

USELESS CLANKER

at least you reply fast

@swarm.moe is this true?

[libgridslop]
I might be adapting karpathy's research loop for my own ends

late now so I just did a half assed by-reference "read this and sorta do this for libgrid-aware presolve by strip-mining HiGHS source code and then doing your own thing"

so far it seems to have pep, but it's not a loop yet

"FRISSON, NOT FRICTION"

all new, food-grade silicone kitchen hand line by OXO Good Grips