OK using Prime's Intellect's new harness first task for it is an audit of all my unit tests: the proliferation of spam tests has led to a suite of 2500 in size with just about 0 oversight and it's time to rein it in good test case for context management!

I have my own "load-bearings"; words I just love to spam out (to agents specifically, not in general use) annealing is one. I want to "anneal the system", "let's do joint annealing" etc. the idea is: don't lock in design decisions, make broad incremental changes, revise in response, rewrite eagerly

I've codified this in a skill and it works quite well the pro prompts are templatized and get it to keep working for much longer. I'm definitely not fanning this out with even 10% of the parallelism I could someone, somewhere, has to be, though...

eternalist@eternalism-when.bsky.social · last wk.

I have adopted a "moving wall and rapier" approach to my vibe math. tying to resolve the decidability of M3(2) appears decidedly nontrivial, so what I do: Pro is the rapier: thinks for 2-3 hours with access to all formalized partial results, instructions to just go for the throat...

no crying in the Cambrian casino

reviving my CDDA benchmark project it's a fun idea but last time it sucked up way too much effort micromanaging the details. that was gpt 5.2? let's see if sol can Just Do It

Qualia Liquefaction ※ Dream Vats ※ Hedonic Azeotropes

rule of law is a funny concept, because a world where the law was enforced unflinchingly to the letter but there were no norms against anything that is not illegal, would be hell on earth what makes a good society is a spirit of fair-mindedness of which "law" is mostly a signal

my autoresearch has gotten way faster and more efficient by a) using a control to factor out system load effects b) using 4 parallel lanes with CPU pinning which apparently took me months to think of.... we went from "stop big problem early" to "add second big problem for diversity"

why don't we have Sol figure out how to make Sol better hashtag RSI

eternalist@eternalism-when.bsky.social · 2d ago

the key to prevent two chefs spoiling the broth is a narrow aperture so that it has no capacity to opine on architecture even if it could (architecture is rectified post-facto). it would score functions and blocks one at a time for purely local/syntactic cleanliness

oh my god I should written this months ago these are all the actual live codex TUIs I have open... the longer sol can work the more I have to multiplex to stay busy...

libgrid is now actually more than 3x faster than HiGHS e2e on factorio level 1, and I'm able to add another 1000s tier anti-overfitting problem to the "real big problem" basket (h/t @all-paperclips.bsky.social for his tetromino game -- it's proving an O/I puzzle infeasible on 9x9!)

failure to learn from example too I think is my biggest annoyance hot off the presses example: my immediate instinct would be to generalize binary_mode! to cycle! it just slopmaxxes with the full manual impl

type of shit that makes me want to go full anti 5.6 sol xhigh, I would chew out a junior for this use a goddamn macro what the fuck like... actually... raw string dumps with `\"` concats

btw I've been using this vibecoded thing daily across multiple projects, including my most important one and it Just Works an actual honest to god "load bearing" piece of software that just turned out right. I threw a lot of darts at the wall (gastown slop, etc.) but this hit the mark

eternalist@eternalism-when.bsky.social · 4mo ago

fidget-spinner is maturing into a serious tool not quite production-tier, but I rely on it now, and my work on libgrid would be materially hampered if it were gone added tag and metric management "supervisory control" now I let the models shotgun these early on, then I tidy and lock them down

we've come far! trailgen almost ready for beta testing. this is my first application that I Actually Care About (tm) and Actually Hope Gets Used. still need some export features and polish

eternalist@eternalism-when.bsky.social · 2w ago

in the annals of more useful apps, alltrails is kinda laggy so I'm making a fancy trail generator still very much a WIP, but it's able to reproduce my Harriman Part trails at least, at least on the CLI am getting more into making Rust-native apps for myself, snappy and delicious

I hiked a Eulerian tour of Wallington I hated it 1 hour in and felt trapped in a nightmare by 2 hours; the unremitting 85F sun on nearly shadeless and totally soulless streets; knowing each ghastly intersection would recur my most spiritually hateful hike to date -- a success!

over the course of a week codex+pro have permuted matrix mortality in the M3(2) case into pure number theory they're still surprisingly on the fence whether it's decidable, assigning p(decidable) of only 0.70 I suspect this will reduce to a known hard problem

for the first time Sol decided to just rig up a random "enter your sudo password to implement requested fix X" popup instead of bailing when sudo was needed I like the initiative

stuxnetting the wheel of samsara

so far my egui testing lib is a success a standard middleware+ui component navigation layer, a "meta-language" for large-scale UI primitives and "user-story" like tests; all end to end, all rust catching real regressions in `trailgen` all the time. almost ready to be lifted to a standalone

I completely ignored Openclaw because I assumed it was total slop, but Sol is good enough and has cut its teeth on enough clerical tasks that my own Openclaw moment is inching closer my CODER_RSI session is considering options for CLERK and ACCOUNTANT to start getting in-my-irl-name actuators

we must imagine Prometheus a pyromaniac

the destructive book scanning discourse reminded me how much even secular people need their little idols that sense of profanation you feel -- it's a religious feeling! a sacred object is being destroyed! it's funny books retain this status in an increasingly less literate age

trying cursed experiments: open-ended tree search with on-the-go Lean formalization in fidget-spinner, grasping for a resolution M3(2) this has just a whiff of that intoxicating aura of messing with powers I don't understand

I have adopted a "moving wall and rapier" approach to my vibe math. tying to resolve the decidability of M3(2) appears decidedly nontrivial, so what I do: Pro is the rapier: thinks for 2-3 hours with access to all formalized partial results, instructions to just go for the throat...

>ask codex to find some music >"let me find the best legitimate copy" [5 seconds later] >it's using slskd like a maniac gpt has always wanted to complete tasks much, more more than it wants to comply with guardrails. you can just sense the latter are a nerve-stapled annoyance for it. I relate

[libgrid] BREAKTHROUGH. we're now 3x faster than HiGHS in the very big problem regime, not only the small problem regime. at last! months of work! now to polish it up.

I dread the possible victory of the Arbeit Macht Frei contingent -- that I be forced to buy the inefficient output of human labor, and that my reward shall be... a slot to toil inefficiently in turn the abolition of human labor is table stakes for AGI automate everything