iiinova — where your digital self keeps evolving

it bangs btw less verbose reasoning (which was the primary annoying thing about DS to me) thus less time and tokens to accomplish the same task, also just better performance

benchmark comparison between new and previous deepseek flash where new wins substantially or at least minorly in all benchmarks

2580 58

stellz @piss.beauty · 21h

happy update day to all who celebrate

banner on deepseek api saying deepseek flash 4.1 has been updated today

1391 39

stellz @piss.beauty · 16/10/2026

there's even a reviewer calibration mode where you launch a panel of models to review the same diff, & then the main agent validates their findings before doing a quality comparison between the reviews it got here i'm comparing gemini-medium, luna-medium, luna-high, & ling-high

and this one. make it very concise

Pi harness screenshot: three parallel subagents (two luna-reviewers, one ling-reviewer) reviewing commit 2f358e5dd0 in openmw_alt’s water-aware transparent bin code. They’re reading source files and git history

Pi harness screenshot: a review-panel comparison on commit 2f358e5dd0 where all four arms (ling-high, luna-medium, luna-high, gemini-medium) found no defects. ling-high was the cheapest ($0.00093) and fastest (34s) at score 9, while luna-high cost about 12× more for no extra insight.

140 4

stellz @piss.beauty · 16/10/2026

the 2nd round of review for each step uses a weaker review panel than the 1st for cost reasons, but can get away with it because we ask them to focus on verifying that fixes applied after the previous round of review are good includes using a intermediate commit to supply a diff since last review

A dark green terminal screenshot from the Pi harness, showing a “ling-reviewer” subagent running a follow-up review of fixes in an openmw-benchmark repo. The task text lists six prior-round fixes to a suite load-mode change, followed by a condensed Rust diff touching app.rs, report.rs, and suite.rs

130 3

stellz @piss.beauty · 16/10/2026

granted, i have flash model madness - i use a custom panel of different flash models that review each other's work in a structured way to do %95+ of work including tracking quality vs cost of review panel members over time

table showing lots of data about the performance of each model+effort level combination when used as a code reviewer

it shows number of runs, avg review score, latency, and cost information for a number of flash models

3191 19

stellz @piss.beauty · 16/10/2026

that's just you knowing that coffee tastes good. this is caffeine addiction speaking

Stacked cases of Celsius sparkling drinks in several flavors, including Peach Vibe, Fuji Apple Pear, Mango Passionfruit, Strawberry Passionfruit, and Kiwi Guava, piled in a corner next to a few Snapple bottles.

6570 57

stellz @piss.beauty · 15/10/2026

thank you for the free tokens Liang Wenfeng, very cool

DeepSeek Usage page showing a blue notice about an API outage on October 15, 2026 (8:11 to 13:04 Beijing Time), with compensation credit for usage from 10/8 to 10/14. Balances are $9.32 topped-up and $3.62 granted, with $8.93 total cost, $6.34 cost over the last 7 days, 3,584 API requests, and 648,880,061 tokens.

DeepSeek billing page on the “Granted” tab, showing one active $4.88 credit created 2026-10-15 04:24:41 with no expiration, labeled “Service Compensation.”

3220 22

stellz @piss.beauty · 14/10/2026

it (experimentally) works with pdfs too it decomposes the pdf into text & images (including rendering diagrams to images) using poppler tools (pdftotext etc.) the qwen model is multimodal so you can input images & get those images considered in the summary

description of some work done to implement understanding pdfs in a multimodal model that doesn't natively support them

280 8

stellz @piss.beauty · 14/10/2026

dedicated summarizer / compactor is the first useful thing i've found to do with this spare 16gb m1pro machine. with a bit of tuning qwen-3.5-9b quality on this can be decent if not amazing on longer contexts

1331 33

stellz @piss.beauty · 14/10/2026

gemini-reviewer: a way to make use of free gemini subscription access, from the same harness i use other models in, pi (which you can't normally do) it shells out to the gemini harness "agy" with a code review task, streams back responses & tool calls so you can see progress. video is sped up 3x

2182 18

stellz @piss.beauty · 14/10/2026

my 3 different flash models reviewing each others' work

2823 82

stellz @piss.beauty · 08/10/2026

testing the OpenMW optimizations with less wasteful shadow settings has yielded smaller gains than originally predicted but still significant: %12-24 gains in fps the level of reduction in pure gpu draw time in some scenes is hilarious to me (see "GPU Timer" rows in the tables)

Benchmark tables comparing OpenMW engine performance across three game scenes (Balmora, Ebonheart, and Pelagiad) under default shadow settings. Across all scenes, the optimized builds reduce draw traversal times and GPU frame times compared to upstream base, yielding net framerate increases of roughly 23% in Balmora, 15% in Ebonheart, and 1% in Pelagiad, with the majority of the performance gains coming from OpenMW engine optimizations rather than OSG build differences.

Benchmark tables comparing OpenMW engine performance under high-quality shadow settings (4 cascades at 2048 resolution) across Balmora, Ebonheart, and Pelagiad. The fully optimized build delivers net framerate improvements over upstream base of 24.0% in Balmora, 11.8% in Ebonheart, and 3.7% in Pelagiad, driven by substantial reductions in CPU draw traversal and GPU frame times that stem primarily from OpenMW optimizations, with modest supplemental gains from the optimized OSG build.

2181 18

stellz @piss.beauty · 06/10/2026

lol i based my branch off of upstream master, the unreleased dev work for 0.52 ...which has ~%13 lower framerate than the current 0.51 release! so perf gain vs the current release is smaller, especially if your shadow quality settings are lower. i guess i need to figure out their regressions

terminal showing agy discussing benchmark results, 0.52 base being slower than 0.51 and my branch

050 5

stellz @piss.beauty · 06/10/2026

southbound platform at wilson ave gotta be one of the prettiest stops to wait at in the city

southbound L train platform looking out onto a cemetary

0532 53

stellz @piss.beauty · 05/10/2026

openMW runs on basically whatever OS + arch, it compiles to arm64, it shows "kind: apple" here which confirms it

040 4

stellz @piss.beauty · 05/10/2026

been cooking something for morrowind (via openMW) enjoyers not done, but current results exceed my expectations: ~%35 framerate improvement in the benchmark outdoor town scene, running on an m1pro machine (thus, running OpenGL via the Metal translation layer)

Performance comparison table titled "Balmora Performance: Before Any Optimizations (master) vs. Now (gpu-measure)" tested at 3456x2160 native Retina resolution. It shows average framerate increasing from 44.1 FPS to 58.7 FPS (+33.1% to +37.0%), frame time decreasing from 22.68 ms to 17.03 ms, CPU draw traversal dropping from ~19.50 ms to 15.35 ms (-21.3%), GPU draw time dropping from ~30.40 ms to 21.72 ms (-28.6%), and draw calls reducing from ~3,326 to roughly 2,980 (-10.4%)

3492 49

stellz @piss.beauty · 30/09/2026

obviously writing code like this is very cool, fun, makes u feel smart, but what if we didn't have to use unsafe

rust impl of interleaving planar stereo buffers together

1150 15

stellz @piss.beauty · 30/09/2026

fascinating bit in the argon news for rust simd enjoyers the dj software project i stalled out on earlier this year has plenty of explicit unsafe simd code in it because previous experience with autovectorization made me not want to rely on it for really hot paths, it felt brittle previously

3531 53

stellz @piss.beauty · 27/09/2026

she's really on top of it

me messaging Yuri to ask to put a package away but she already did it

2390 39

stellz @piss.beauty · 25/09/2026

what if I got 3 beers deep at the airport and played morrowind for a 4 hour flight

stellz with 20oz beer in laguardia airport

4290 29

stellz @piss.beauty · 25/09/2026

3 trains and a bus to the airport 😮‍💨

stellz waiting for the q70 bus to LaGuardia

5800 80

stellz @piss.beauty · 24/09/2026

my mod manager, g3.8-flash, fixed a rendering issue causing character creation & inventory models to not be lit by figuring out it was a known issue & changing the right openMW config settings, pretty quickly

screenshot of the conclusion of gemini's "agy" agent harness fixing a openMW rendering issue

170 7

stellz @piss.beauty · 24/09/2026

openMW has never been better. you should play morrowind

Morrowind in 3rd person near the opening. player is a dunmer, and the character Tarhiel is dead on the ground in the foreground (from jumping way too high)

2141 14

stellz @piss.beauty · 24/09/2026

which one of these would you prioritize. i'm not sure i have the time to do skyrim-in-morrowind but i'm seeing from the mufoverse that it's that time again

morrowind opening scene on the pause menu, running on a mac in openMW

110 1

stellz @piss.beauty · 14/09/2026

got that smoking ferris shit on today

stellz face selfie wearing a black baseball cap with the rust crab smoking a joint

1260 26

stellz @piss.beauty · 06/09/2026

end of an era of freeriding, I knew this would happen eventually. so long and thanks for all the tokens

GitHub copilot free access ending soon notification email

0331 33

stellz @piss.beauty · 04/09/2026

finally I can find this sort of shit again in a usable website

180 8

stellz @piss.beauty · 01/09/2026

oh worm? claude good again maybe?

As our models’ scientific capabilities improve, our investment in scientific progress is also growing. Last week, we previewed the Model Hardware Standard, which allows Claude to directly and safely operate laboratory equipment

030 3

stellz @piss.beauty · 31/08/2026

new place is so cute, we got the keys

3860 86

stellz @piss.beauty · 29/08/2026

if so this is still wild to me

Update 12/18/2025: The author has updated the book based on my two main criticisms in this post. She explains the changes in her blog here. I think the chapter now leaves readers with a better understanding of how water’s being used by AI. I have some remaining background disagreements with some of the presentations of the issue, but the author was super direct here and engaged a lot, and corrected the two major issues I found, so I want to commend her for that. Really great to see.

3240 24

stellz @piss.beauty · 24/08/2026

people are getting too comfy posting random non dance events on RA. don't be posting your tenderqueer card game on there, this is spam

Resident Advisor screenshot showing a "queer UNO tournament"

6291 29

stellz @piss.beauty · 24/08/2026

limited edition can check it's pretty good, does taste like limoncello

"sparkling limoncello twist" Celsius can

1130 13

stellz @piss.beauty · 12/08/2026

I'm pressing x to doubt

000 0

stellz @piss.beauty · 10/08/2026

which way, fair ferengi clothed daughter or faggy son

2364 36

stellz @piss.beauty · 04/08/2026

logging into this account rocks

atmosphere login UI showing "piss" typed with my account as the first typeahead result

3490 49

stellz @piss.beauty · 01/08/2026

doesn't seem true blog.google/innovation-a...

AI use is not limited to white collar workers, it’s also assisting workers in predominantly physical and manual occupations with adjacent tasks: AI use for work is not limited to jobs traditionally seen as knowledge work. While not as prevalent, workers in manual and technical trades (e.g., auto technicians, industrial mechanics) are using conversational AI as a live collaborator for real-time diagnostics, troubleshooting, and on-the-fly learning. When workers in these areas use our AI tools, they’re 2x more likely to use multimodal AI (i.e. using AI to create images or video). For example, automotive technicians and industrial mechanics use AI to interpret complex test results, debug electrical wiring, and inspect machinery for wear.

2102 10

stellz @piss.beauty · 29/07/2026

also idk what rust codebases Fernando was looking at but the described usage pattern is not true at all. we're absolutely stuffing those refs in structs all over the place

Is this doable? Surprisingly, yes. If you grep a large Rust codebase—either libraries or applications—you’ll find that the usage pattern of references is:

The overwhelming majority of references are just arguments to functions.
Sometimes you see a reference being returned from a function.
Rarely, very rarely, are references stored in structs, and often the lifetime is 'static.

This usage pattern is why Rust’s lifetime elision feature works. This lets you avoid writing lifetimes in most code. For example, you can write this:

210 2

stellz @piss.beauty · 27/07/2026

hearing about the whole festival douching station is sending me

Insta post saying
"You know what…. I’m not saying anything regarding the matter of whole festivals douching station, the use of the phrase “main shit of the day” or the concept of waiting in line publicly in the pissing down rain and mud to spray shit from your arse at a festival to be fucked by a meth head who pisses Florescent orange. Im protecting my mental health on this matter"

2141 14