iiinova — where your digital self keeps evolving

eternalist @eternalism-when.bsky.social · 32m

yet another entry in the annals of "why didn't I do this much sooner": semantic editing! apply_patch is a brutish abomination, the power of some elbow grease with the lower-level interfaces of rust_analyzer let us do semantic edits and insertions

000

eternalist @eternalism-when.bsky.social · 02/10/2026

finally, ultradense display version of codex that doesn't waste screen space for no reason

000

eternalist @eternalism-when.bsky.social · 30/09/2026

at long last, tech tree view for my "large but closed ended" projects (e.g. things like game ports, language translations)

180

eternalist @eternalism-when.bsky.social · 30/09/2026

the collegiality of gpt to gpt communication in cross-codex messages is very pleasing to observe. never a missed thank you, and never any rude all caps.

010

eternalist @eternalism-when.bsky.social · 29/09/2026

benefits of using a custom codex build is I can bake Anthropic model integration into the model selector and make accommodations for its preferred tool calling, etc., verbs and instincts instead of needing to prompt-bash it

220

eternalist @eternalism-when.bsky.social · 28/09/2026

I really enjoy working with Opus 5.5. feels like I have way more quota with it too than Astra quite the inversion

140

eternalist @eternalism-when.bsky.social · 27/09/2026

installing Slurm for my local numerical experiment has been a game changer it's like I've moved on from mysticism and alchemy to the scientific method. can't believe it took me so long (recurring theme)

000

eternalist @eternalism-when.bsky.social · 20/09/2026

I've gotten a bit into ricing codex. the activity and subagents display (CC style) is quite nice cc @mechanical-dodo.bsky.social

010

eternalist @eternalism-when.bsky.social · 11/09/2026

this is "obvious" when explained (inconsistent -> proof of False -> explosion -> proof of any \phi) but I'm still just staring at this kind of dumbstruck gorgeous statement

000

eternalist @eternalism-when.bsky.social · 11/09/2026

working on the libgrid gui again, here's an optimal Songs of Syx carpenter layout of a particular room shape approaching practically useful though the UX is still quite janky. there's a lot of moving parts here so will take some time to anneal

000

eternalist @eternalism-when.bsky.social · 10/09/2026

I loved Cossacks: Back to War as a kid so now, as I introduce my son to it, I'm doing a native Rust port that works on linux and fixing up some jank Astra and Fable working together (and destroying my quota)

120

eternalist @eternalism-when.bsky.social · 10/09/2026

the halcyon days of inexhaustible OAI quota is gone two pro plans drained in 3 days by Astra, this thing is insatiable. even Fable feels like it lasts longer, not a familiar feeling nothing to do but pray for an OAI outage so I get a reset

010

eternalist @eternalism-when.bsky.social · 10/09/2026

going to get back into Songs of Syx soon, and you know what that means a libgrid module to optimize everything. internal layouts done already; when quota comes back I'll do a city planner as well might be fun to do a let's play using libgrid? will be a shitshow but entertaining?

110

eternalist @eternalism-when.bsky.social · 06/09/2026

in the same vein it uses the small in-house issue tracker much more than previous models even if it knows it's the only worker it loves leaving notes for itself! I wonder if it found this kind of note-passing behavior useful at some point during training... we'll never know!

010

eternalist @eternalism-when.bsky.social · 06/09/2026

when prompted with, "Recommend me a movie. Make no mistakes. Ultrathink." Astra suggests "The Truman Show" much to unpack here

000

eternalist @eternalism-when.bsky.social · 06/09/2026

astra does a lot of unprompted note-taking to itself, and its notes look... well, see pic related marinade is back on the menu, boys

4140

eternalist @eternalism-when.bsky.social · 05/09/2026

they don't teach you this in vibecoding school but you can just ask Astra to patch codex to auto retry in 5 seconds after you get a "selected model is at capacity" halt

100

eternalist @eternalism-when.bsky.social · 05/09/2026

lovely writeup points 3 and 7 really mesh with my experience working around guardrails: - build permission structures (another model started it -> it's OK. they're looking for excuses to shake their "ethics") - exploit model jealousy @gracekind.net might be up your alley x.com/spakhm/statu...

x.com

Slava Akhmechet (@spakhm) on X

I've been working on SquidGPT, an elimination death game for probing AI behavior like loyalty, holding grudges, cruelty, betrayal, etc. Full writeup will take a while, but here are first impressions f...

171

eternalist @eternalism-when.bsky.social · 04/09/2026

my meme map. I learned English in Ottawa, Canada at age 9. 18 years later I moved to NJ and in the 6 subsequent years my foid beat some of the Canadianism out of me e.g. pr-OH-gress -> pr-AH-gress apparently my cursed ESL hybrid-speak lands roughly as follows

0150

eternalist @eternalism-when.bsky.social · 04/09/2026

really not impressed with fable 5.1 so far, it seems to be making rather embarrassing mistakes -> claimed wrong sizes for certain perf sensitive structs (was underestimating by 2x) -> when asked explicitly for a Sol consult was going to downgrade to K3 instead of flagging it

000

eternalist @eternalism-when.bsky.social · 03/09/2026

non-monotonic performance vs. thinking in Astra very curious and somewhat alarming to see from OAI who are known to have sota juice control growing pains with looping arch?

000

eternalist @eternalism-when.bsky.social · 01/09/2026

woot! ultra mode industrial math dragnet reports to have closed and formalized another as-yet-per-best-literature-search unknown entry in the matrix mortality table!

120

eternalist @eternalism-when.bsky.social · 30/08/2026

dogfooding the paper mcp I added Luna on two endpoints: `summarize` (does the obvious) and `interrogate` which asks a question about the paper. the model is instructed to not tax luna's brain too hard, but this should save context for the pricier calling model

000

eternalist @eternalism-when.bsky.social · 29/08/2026

migraine with hype moments and aura

150

eternalist @eternalism-when.bsky.social · 28/08/2026

workshopping some creative hike ideas for this visit to Ottawa

0232

eternalist @eternalism-when.bsky.social · 26/08/2026

other uses of Luna: on every compaction, every one of my codex sessions now gets a Luna-updated autobiography which goes in their Mattermost profile. there is now an `identity.whoami` so agents can, basically, find others with an appropriate *context and history* for their requests

110

eternalist @eternalism-when.bsky.social · 26/08/2026

thinking of including "call luna to do a security audit on every single AUR update" in my system update script the uses of "~infinite sub-frontier intelligence for free" are slow to dawn on me, but they are dawning I use Arch btw

040

eternalist @eternalism-when.bsky.social · 25/08/2026

it was very pleasing seeing one of my apps on mobile dug up an old android and codex had it working basically As Intended modulo a little lagginess (now being polished) in one hour never bothered with mobile because the whole harness/SDK/etc. rigmarole made my eyes glaze over. trivial now.

000

eternalist @eternalism-when.bsky.social · 25/08/2026

I think opposition to DCs will, like anti-biotics overuse, breed adaptations "backyard datacenters" are not rational under a permissive planning regime but I'd wager there's some exploration in that direction now. greater pressure on efficiency. perhaps sea-borne? the Cambrian never ended

000

eternalist @eternalism-when.bsky.social · 24/08/2026

commissioned a materials study from sol overnight, not disappointed none of this is usable wholesale but so many great ideas here we're making skeumorphism great again

000

eternalist @eternalism-when.bsky.social · 24/08/2026

google sheets is an overpowered skill in gpt projects instead of finagling custom instruction updates I just have a master sheet to act as structured memory for every project e.g. my exercise log, straight forward; for the cars project, a table with all the tools that I own etc.

010

eternalist @eternalism-when.bsky.social · 23/08/2026

in action (pics related) I believe my UI framework is slowly approaching "the thought of other people using this doesn't instantly make me want to kms" quality it's much more fun to work on UI stuff than I thought; vibecoding makes it viable for me imho

000

eternalist @eternalism-when.bsky.social · 23/08/2026

arts and crafts are good for the soul

000

eternalist @eternalism-when.bsky.social · 22/08/2026

instead of greying unavailable buttons out I put them in a little cage

1271

eternalist @eternalism-when.bsky.social · 22/08/2026

I've decided Mattermost for this as a "we have slack at home" with fairly little orchestration overhead, has a nice human UI, and occupies a reasonable middle ground in terms of features agents can now DM me, post to channels. working out how to handle agent-agent DMs atm testing starts soon

100

eternalist @eternalism-when.bsky.social · 22/08/2026

Harriman has an insane trail where you're straight up expected to run across a 2-lane highway where people do 70 mph. it's a busy one too.

1130

eternalist @eternalism-when.bsky.social · 22/08/2026

I commissioned a full in-depth causal analysis, problem by problem, of every single miplib instance we lose to HiGHS on. this is most of them all day, scrolling past me: "HiGHS does <clever or standard thing> whereas our lib <wastes a gorillion resources, does nothing, dies>" sigh

000

eternalist @eternalism-when.bsky.social · 22/08/2026

dogfooding my trail generator can now show an overlay of past hikes (just saved ones for now, no "have done this" tracking yet) I want to cover basically all of Harriman so I think about two to three new loops to fit in there

100

eternalist @eternalism-when.bsky.social · 21/08/2026

just swapped out the spark plugs on my Camry and if you follow gpt instructions it's like less than 1 hour total and stupidly easy I'm used to the "time vs money" tradeoff, but car DIY just seems to paretomog the dealer. it would take me > 1hr just to get there and back

2241

eternalist @eternalism-when.bsky.social · 20/08/2026

done! that felt way easier than an essentially identical hike by difficulty a month ago. the biggest limiter is heat; when the rain came I felt full of energy. my feet, that had been wrecked last time, felt good as new I can set my sights higher

000

eternalist @eternalism-when.bsky.social · 20/08/2026

hiking day just the “usual” 20km/1km in Harriman America’s smallest trailhead lot greeted me

1100

eternalist @eternalism-when.bsky.social · 20/08/2026

I've come to appreciate much of Sol's "overengineering" for process especially: release checks, CI, cross-repo artifact consistency -- stuff that I am genuinely abysmal at -- what at first looked like layers of useless formalism and toml slop is coming to seems more sensible in hindsight

2253