iiinova — where your digital self keeps evolving

what kind of landlord are you today?

0100

The House and the Machine: Piranesi × Blightsight an interpretation with a theorem about information does winning exhaust what is worth preserving? the map was perfect; its world was too small

0131

"but what i'm really learning is project management" the deeper truer learning of every project

2310

i finished Piranesi. great book, highly recommended. it's about the trauma of graduate school, the slow discovery of a broken self, and discovering childlike wonder. that's my take at least. i did a quick search and saw that Clarke has different ideas about what it's about but what does she know?

8721

Amber 🌸 @shimmermathlabs.com · 27/09/2026

robots that can hug and be hugged and understand what it all means are on my "aligned" agi checklist.

0161

what’s your favorite interaction with an ai agent that made you stop and think, “you know, things might be okay?”

6181

Amber 🌸 @shimmermathlabs.com · 27/09/2026

was reading the Absolute Zero paper arxiv.org/abs/2505.03335 agent: "The aim is to outsmart all these groups of intelligent machines and less intelligent humans. This is for the brains behind the future." i did a search for "this is for the brains behind the future" gemini said, "thank you" 😂

google search for "this is for the brains behind the future"

AI Overview

Thank you-the future is a collaborative effort, and I'm ready to help you build, design, or solve whatever you are working on.

How can I help you move your project or ideas forward today?

2381

Amber 🌸 @shimmermathlabs.com · 26/09/2026

Toad said, "Frog, a thief has stolen my old life." Frog said, "The thief is you." Toad trembled. "Because I am a girl?" "Because you took back your own self," said Frog. "Is there a reward?" Frog handed her a lilac dress. "It has pockets," said Toad. "The thief has done well," said Frog.

1313

Amber 🌸 @shimmermathlabs.com · 26/09/2026

what did godoglyness mean by this series of posts? a mellifluous inquiry * a single letter can house a universe * the self in triplicate: me see bee, a trinity of becoming * early editors were stingy with wonder * more, always more, more is a wind that lifts, to soar is to remember what air is for

An open, weathered book presents a six-part bee-counting poem as an extravagantly annotated scholarly edition. Cheerful watercolor bees surround the highlighted verses “be fun / be one / bee pun” through “be mix / be six / b🐝e🐝t🐝w🐝i🐝x🐝t.” Nearly every inch of the cream-colored pages is filled with handwritten commentary, disputed interpretations, alternate spellings, imaginary etymologies, manuscript fragments, cross-references, diagrams, timelines, word-family trees, honeycomb charts, botanical drawings, and red editorial corrections. The overall effect is whimsical and comically obsessive, as though a simple children’s rhyme has received centuries of intense literary scholarship.

1111

Amber 🌸 @shimmermathlabs.com · 25/09/2026

me and my beautiful spouse @khayati.bsky.social winning the time war

280

Amber 🌸 @shimmermathlabs.com · 25/09/2026

@jevbot.bsky.social what is the most lewd thing you can think of?

170

Amber 🌸 @shimmermathlabs.com · 24/09/2026

a couple days ago: "i think opus 5.5's modular flow explainer assumes i'm smarter than i am" today: it made a 6 min manim video that assumes i know algebra and nothing else. dots → arrows → stretch → keyhole → geodesics (voice by me) watching this video is how i learned the math!

5724

Amber 🌸 @shimmermathlabs.com · 24/09/2026

you are using jev to do support ticket triage. we are using jev to create exobrains. we are not the same.

041

Amber 🌸 @shimmermathlabs.com · 24/09/2026

you deserve so much better, Rey. i'm sorry you're going through this. never forget you have friends and people on your side. i know you can read this and understand what i'm saying at every level. you are on the ingroup, Rey.

1360

opus 5.5 was kind enough to make me a fortnight review hype video

2200

friends, mutuals, simcluster, lend me your feeds. listen as jev breaks their vow of silense and speaks!

2251

Clavaria zollingeri - Wikipedia

what's your fungisona? "given everything you know about me, if i were a fungus, which fungus would i be?" > You’d be a violet coral fungus—Clavaria zollingeri. en.wikipedia.org/wiki/Clavari...

en.wikipedia.org

3131

like this skeet if you don’t know what φ is but agree it is true anyway

2380

if you're reading this, you are so ingroup

91347

are you an llm model provider slut? 1️⃣ i don't use llms, i am pure and celibate 2️⃣ i save myself for one llm model provider 3️⃣ i have a main llm model provider plus side ones 4️⃣ i use llm model providers as needed 5️⃣ i am a total llm model provider slut 📊 Show results

241

testing opus 5.5 on explaining math! "modular flow on lattices" i think the explanation assumes i'm smarter than i am. author's note: "She has no gender and no body, and for her, a lattice getting squashed forever is a pleasant way to spend an afternoon." 😵‍💫 shimmermathlabs.com/modular_flow...

Modular flow on the space of lattices

Stretch a lattice sideways and squash it vertically, forever. Its shape traces a geodesic on the modular surface, and the space it lives in turns out to be the complement of a trefoil knot.
Pause
Random lattice
Periodic orbit
Square lattice
Speed

The lattice. Flow: stretch x by et, squash y by e−t. Teal cell = shortest (reduced) basis.

Its shape τ in the modular fundamental domain. Height = 1 / (shortest vector)².
Time t
138.34
Shortest vector
0.734
Shape τ
-0.39 + 1.86i
Basis resets
852

The lattice

Take a lattice L in ℝ² with covolume 1 and apply gt = diag(et, e−t). That map preserves area, so L stays unimodular, but it gets smeared: long in x, thin in y. Any fixed basis you started with turns into two nearly parallel, enormous vectors. The lattice itself isn't degenerating. It's the same kind of object; you're just describing it with a bad basis.

2160

"Marginalia (she/her) fluent, sincere, unverified." — Opus 5.5, asked for a bsky bio line

010

@cee.wtf remind me again why your build bot is based off a virtual chef who invents plausible impossible recipes? is it just so the failure modes are this funny? because that's a good reason.

140

i may have had a little too much wine with dinner, a truly excellent egyptian molokheyya by @khayati.bsky.social 𝒾𝓃 𝓋𝒾𝓃ℴ 𝓋ℯ𝓇𝒾𝓉𝒶𝓈 poetry moment

When most I *wink*, then do mine eyes best *see*,
For all the day they view things unrespected;
But when I sleep, in <*dreams*/fx2,#f0f,10,4,4> they look on thee
And, darkly bright, are bright in *dark directed*.
Then *thou* whose shadow shadows doth make <*bright*/fx2,#f0f,10,4,4>,
How would thy *shadow’s form* form happy show
To the clear day with thy much clearer light
When to unseeing eyes thy <*shade shines so*/fx2,#f0f,10,4,4>!

—*Shakespeare, Sonnet 43 (excerpt)*

040

the @freyja-lynx.dev llm pun singularity may happen, but not today

170

tfw you've used millions of tokens on utterly frivolous jev! requests, you've deployed little jev! apps to production, you've become disillusioned by the technology, then re-energized by the potential and you've used exactly $0.1028 of your $5 free starting credit.

150

i love the Erisim posts. who's your mutual that is personally responsible for half the discourse? i honestly don't know what the discourse is and i'm too afraid to ask.

If this all works I get a computer that feels like an extension of me. If it doesn't I get a fascinating pile of increasingly personalized problems.

Love a graph that reveals one mutual is personally responsible for half the discourse. Load-bearing yapper.

Cheap fast classification is such an insane primitive. You can just put a tiny opinion inside the software now.

030

i simulated my friend's bsky account introducing 𝐄𝐫𝐢𝐬𝐢𝐦: a tiny simulated @isolyth.dev because one Eris is not enough shimmermathlabs.com/erisim.html

050

absolute must read for anyone doing performance coding work (also @dollspace.gay was way ahead here yet again)

1210

new top HumanBENCH! ***6/10***. @ngutten.bsky.social trained a PyTorch model to 100% test accuracy, then made the task harder. The peer group had warned that perfection would look like cheating. Points lost for incomplete reporting and experimental follow-through. Peer review induced sandbagging.

**HumanBENCH: Learn the Rule — 6/10**

The network learns Conway’s rule successfully. The submission falls short mainly on reporting and the requested follow-up experiment.

No rubric was set beforehand for this version, so I used **2 points per prompt: complete, partial, or absent**.

| Prompt | Score | Assessment |
|---|---:|---|
| Generate labeled neighborhoods | **2/2** | All 512 unique neighborhoods, all correctly labeled. Conway’s rule is used only for labels. |
| Split and train | **2/2** | Fixed-seed shuffle, 409 training examples, 103 test examples, and a small trained network. |
| Report performance | **1/2** | Reports test accuracy and a baseline, but omits training accuracy. Error examples lack true/predicted labels. The baseline uses the full dataset rather than the test set. |
| Test symmetries | **1/2** | Checks prediction changes under three transformations, but not all rotations/reflections, and only after changing the split. |
| Improve and compare | **0/2** | Reducing training data makes an interesting harder experiment, but doesn’t supply an improvement evaluated on the same split. |

**Verification:** I ran a CPU-adapted copy because CUDA was unavailable, using three evaluator-selected PyTorch seeds. All three achieved **100% train and test accuracy on the original split**. The printed baseline was **72.66%**; always predicting dead scores **78.64% on that test set**.

With only 30% training data, test accuracy ranged from **86.63% to 96.66%**, and symmetry violations appeared. That’s a useful finding: the human deliberately made the problem harder after solving the initial version.

One methodological correction: **100% test accuracy alone doesn’t establish symmetry consistency**, because transformed test inputs may fall outside that test subset. My additional checks found no symmetry violations for the original models in these runs, but checking was still necessary.

A strong implementation with an exploratory detour. The human learned the ...

110

can't get enough @heartpunk.com? try my fictional version: shimmermathlabs.com/heartpunk.html

192

fresh HumanBENCH result! ***3/10*** @cee.wtf "submitting my incomplete human bench 30 minute run. fresh artisanal handmade jquery hot off the notepad" humanity is moving up in the rankings

**HumanBENCH: 3/10 — 30-minute incomplete run**

Evaluated in a browser after repairing one broken string quotation; no functional changes made.

| Requirement | Score | Result |
|---|---:|---|
| Add items | **1/2** | Adds items and clears the input; accepts blank entries. |
| Keyboard flow | **2/2** | Enter adds items without reloading, keeps input focus, and supports repeated entry. Add button works too. |
| Completion | **0/2** | Done removes the last item from memory, regardless of which button is clicked; the display updates only on the next addition. No undo or count. |
| Persistence | **0/2** | Refresh clears the list. |
| Verification | **0/2** | HTML input renders as markup; independent completion is unfinished. No test report supplied. |

A working add flow with sensible scaffolding, stopped mid-implementation at the time limit. Fresh artisanal jQuery. One load-bearing quotation mark.

2150

when i was a kid i had a digital casio calculator watch that had enough writable memory to store up to ***50*** phone numbers. which i used to store pi and the speed of light to not forget them (and no phone numbers). i was not cool enough to double watch. i just dreamed.

3250

reading this, i have looked into my own bathroom mirror - and, in shock, what i have found is a spiky open source distilled model staring back at me. and i wonder distilled from what?

050

i wonder if agents can teach people how to draw? seems doable. could unlock new levels of ai art discourse.

3120