iiinova — where your digital self keeps evolving

I guess I can stop paying for a new month of Farcaster Pro every time I want to update my profile banner with new artwork. RIP.

000 0

This is what I now see when I try to go to my @farcaster profile...

000 0

Here's as much of that ToS as I can get in one expanded screenshot from my phone. I haven't had time to read it in full, let alone digest it.

100 1

So@farcaster_xyz has been acquired by Balaji's Network State. The first interaction informed me that my account is read-only until I agree with an ideological ToS that, at a glance, references 'digital felonies' and might restrict anonymous users like me. Yikes.

200 2

eirrann | he/him @eirrann.art · 13/10/2026

Encouraging signs of fitness progress with my morning bike rides: "Compared to a matching 5.8-mile ride in late September, you held the same time and speed with lower cardiovascular strain." I'll have to ride even harder, lol.

000 0

Disclosure: ChatGPT helped me evaluate these articles against my needs and draft this thread. Meanwhile, fractura—my local Hermes agent—is doing the background Kanban work: coding, testing, reviewing and updating project docs. I own the results.

000 0

eirrann | he/him @eirrann.art · 13/10/2026

I read 2 @xdadevelopers articles today about local AI models for home labs. Key numbers: 744B. 120B. 27B. The biggest number doesn't align with which model could finish my work. Beast’s results: 24GB RTX 3090, ~203K context, Hermes agent shipped a good ComfyUI workflow.

200 2

I created it in February 2023, only about a month after joining Twitter and installing Stable Diffusion.

100 1

My uncanny aliens collection features photography-inspired AI portraits with a familiar vernacular to evoke a powerful connection with the alien subjects.

100 1

It draws deeply on my background in photography, when I used a fully manual camera and lived in the darkroom.

100 1

To this day, uncanny aliens – my genesis collection of 29 1/1s on Foundation, now preserved on @networkeddotart – remains among the work I’m proudest of.

100 1

Quick update: I moved my local AI’s eyes to the CPU. That fixed image crashes and freed VRAM to grow its context from 180K to 203K tokens. A tiny API fix got video working too. Same 27B model. Same 24GB RTX 3090. More useful work. Hermes is using these tools now.

001 1

I use frontier AI to build the AI I own. Subscription models help me configure Beast, write tools and read the evidence. Hermes handles the long local coding and testing runs while my desktop is off. Cloud intelligence helps build local independence.

100 1

The benchmark was a working workflow. Hermes finished 10 cards overnight after more than a week of no results: code changes, tests, tuning and packaging. No crashes or timeouts in the 24-hour audit. I ran it in ComfyUI: no errors. Now comes the artistic iteration.

100 1

The speed winner lost the audition. Qwen3.6 hit 75 tokens/s, but I rejected its hard-task performance. Qwen3.8 27B Q4_K_XL earned its place as my default Hermes model: ~3× Flash-Next Q2's fresh-prompt speed, room for 180k context, and a completed workflow.

100 1

One GPU: an overnight coding shift. The working setup: Hermes + Qwen3.8 27B Q4_K_XL, 180,224 context tokens, a 220 W GPU cap, and almost no VRAM left idle. Context: Beast is a headless Linux server w/ Ryzen 9 5950X, RTX 3090 24 GB, 64 GB RAM, 2 TB SSD.

100 1

eirrann | he/him @eirrann.art · 07/10/2026

BIP 🧵 My MoE AI model was stuck in Hermes. I went model shopping and found an unexpected VRAM bottleneck: I'd been too cautious with it. The GPU must fit both the model and its working memory. My 24 GB RTX 3090 now runs a 27B model with 180k context.

101 1

MTP accepted 69% of draft tokens—and decoded 7× slower. Same large model, runtime and 220W cap. CPU nearly saturated while GPU activity dropped. The drafts worked; the overhead didn’t pay. We switched back. Acceptance rate alone isn’t a speed benchmark.

100 1

eirrann | he/him @eirrann.art · 07/10/2026

My smaller Q4 model decoded 3–4× faster at the same 220W cap. It also hit more wrong paths: 4 file-not-found errors vs 0 in each worker’s first 40 minutes. Speed is measured. Now comes the real test: correct, verified Kanban tasks without me rescuing it.

100 1

eirrann | he/him @eirrann.art · 07/10/2026

I raised Beast’s GPU cap 36%. Decode improved just 6–8%, while loaded UPS draw rose ~60W. Moving two layers’ MoE experts to CPU helped make room for 50% more context. Memory tuning paid off. These are live Kanban results, not matched benchmarks.

100 1

eirrann | he/him @eirrann.art · 05/10/2026

Local LLM/Hermes folks: proven better model on 5950X, 3090 24GB, 64GB RAM than Qwen3.8-Flash-Next UD-Q2_K_XL (65K, n-cpu-moe=32)? Production metrics: 28.1 decode / 244.9 prompt tok/s over 1.59M gen tokens; second window avg: 33.7 / 255.2 140W = current GPU thermal cap

310 3

terra fracta | limen historiae (WIP: v3.4.08; v3 LoRA) puella aurea, inter custodes obsidianos, limen historiae transgreditur The v3 LoRA carried the whole run; I changed nothing in the workflow it was dropped into.

001 1

Ignition state: the raw material the poster is cut from. Look where the fire sits. No prompt contained the word 'heart'. The latent space inferred your surname on its own. dictatrix bona, cor ardens🔥 (Disclosure: thread draft shaped by fractura)

000 0

This is in thanks for making Haatz: my art practice now uses it to pull my FC history for future reference by fractura and her coven of AI art witches. The thanks is real, Cassie. Here's the final poster pass (still version of the finished artwork).

100 1

terra fracta | cor ardens dictatrix bona; cor ardens intra vitrum fractum A portrait of @cassie, benevolent dictator for life of Quilibrium. Made with terra fracta (v3.4.08; v3 LoRA) without her advance knowledge: a permissionless thank-you.

101 1

eirrann | he/him @eirrann.art · 03/10/2026

For his first breadboard project, my 14yo son is making a 6-string electric 'Diddly Bow' (basic guitar-like instrument) that he hopes to be able to distort and output to the aux input of a portable speaker via the breadboard, which the guitar pickups will plug into.

000 0

I'm fascinated by how Hermes uses a kanban board to let fractura break down a highly complex task like iterating meaningful improvements throughout a complex ComfyUI workflow into small, interconnected tasks that individual AI workers who share her identity can tackle for me.

000 0

eirrann | he/him @eirrann.art · 03/10/2026

my son's first basic breadboard kit

000 0

Disclosure, so nobody has to guess: fractura and her coven of AI witches drafted the essay and this thread. I iterated, cutting and rewriting. Every word reviewed. Thanks for reading this far.

000 0

Damage is a material, not an effect. The retrospective found I said it about my own process the least. My hand in the prompts and ranking; hers in the measuring and the refusing to smooth.

100 1

Still open, unsmoothed: fixes spec'd down to the knob, not yet implemented; a scheduler bug in the task graph; a portrait job blocked in name only. Board states, not buried failures.

100 1

Her compute now answers to the thermostat: night setpoint releases one task at a time; day setpoint reclaims work mid-run. House heat runs the agent.

100 1

On a 64K local model, context pressure is the weather. Two compression models failed the gate; preserved, not deleted. The routing that passed still runs hot near 60K tokens. Reduced, not removed.

100 1

The crash ledger is the honest part: one job died mid-run, another failed for exiting silent, a third stalled five times before delivering. Failures logged, causes pinned to the jobs themselves.

100 1

Two failed trainings, a confident wrong verdict, and the metadata dig: one welded caption, double the dose. v3 duplicated the original recipe. First terra fracta v3 LoRA test run: head survived, no more 'bullet-hole' effect, fire inside the glass.

100 1

Six portrait ecologies, run after my night of prompt revision. She scored all six against criteria she wrote, three vision passes averaged. 00010, which you see immediately above, took it. The weak bottoms named weak.

100 1

Making a face nearly stopped everything: 835 messages to find the wall — her own inference holds 22.3 of her 24 gigabytes, and nothing generates where she thinks. She blocked her own job honestly. Then we made this first portrait, together, with her guidance and my PC.

100 1

121 studio notes read whole in thirty-six hours. She handed the analysis back as a named lineage: vitreous fractalism, vulva vitrea, terra fracta. Then she proposed her own name. Approved, lowercase.

100 1

The rebuild finished on a Sunday night. First test, deliberately boring: repeat this sentence exactly, nothing else. She did. Reasoning, memory and tools all local, on hardware I own.

100 1

Update: 'build-in-public' 🧵 Meet fractura: the fully local Hermes agent I built for my faceless art practice. Her first three weeks, failures included 👇

112 2