iiinova — where your digital self keeps evolving

y'know, as someone that's been posting here for like 2 years now, I really should get on actually learning the protocol and putting together my own pds... oh, frank.

either bluesky is broken, or you can now just post "nothing" now

000

what i've learned with benching my custom quant at home so far: 1. fuck this takes forever no wonder there aren't any good sites comparing different quants 2. quanting kv to fp8 on qwen-flash-next has no practical effect on code performance ...more...

120

heating your home with an electrical heater - waste of power - does nothing else but burn power - allowed in church heating your home with AI - super fun - does more than just burn power - the devil heating your home with a portal to hell - super fun - does more than just burn power - the devil

000

𝓐𝓾𝓫𝓻𝓮𝔂 🥀🪦💀 @aub.bsky.social · 12/10/2026

was just reading the glm-legal press release ...homies, did they really name the robot lawyer test after Two-Face?

000

Reposted by Frank Blank

every apparently good idea I've had in the agentic era will still take hundreds or thousands of developer hours to do well

1292

Reposted by Frank Blank

p(Dulany) @dulanyw.bsky.social · 11/10/2026

[Bursting through the door like Kramer] The Verge just discovered local AI!

3814

finally did something yesterday i've been trying to make time for all year i finally sat down and played with my ayn thor all day

000

Chad Barb @chadbarb.bsky.social · 11/10/2026

my personal data point to add: according to "Frank's AA-Lite" (lol I might release that in like a month, basically all of artificial analysis that doesn't depend on third party apis) models including and better than qwen3.8 27B can accurately code on Q8 quanted KV cache. that's double tokens!

000

Reposted by Frank Blank

As Gibson said, “The future is already here — it's just not very evenly distributed.”

061

Reposted by Frank Blank

mr. TIM @timkellogg.me · 11/10/2026

the funnel - 80% of the world has no AI contact - ~20% have janky experience - 0.2% are power users - 0.00006% work for labs and have internal access to actual frontier models

Andrej Karpathy
@karpathy .27m
Dusting off this tweet from April because this gap in shared understanding of LLM capability is "widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel.
- Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all.
- Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract.
- Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code.

- And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work.
Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAl-HF incident from many months ago.
Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while.
The funnel is driven by a combination of factors:
1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel.
2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel.
3. Access. Free-tier, paid-tier, internal.

1316815

Reposted by Frank Blank

dame @dame.is · 11/10/2026

i’ve been disappointed to see the number of people on bluesky sharing this study as if it isn’t complete & total bullshit what makes it worse is that so many of the people uncritically sharing this study were using it to call AI users stupid, when in reality *they* were being stupid for sharing it

710514

Reposted by Frank Blank

Haus of Decline @hausofdecline.bsky.social · 11/10/2026

As you get older, you realize the importance of having a large amount of shallow bowls in which to put small amounts of discrete objects

824969549

Frank Blank @frank8465.bsky.social · 11/10/2026

I have no clue what the fuck these are. All I know is that it's my favorite candy on the planet. (honorable mention obviously to the Real™️ intersex pride flag (💜cadbury flake💛))

140

Frank Blank @frank8465.bsky.social · 11/10/2026

Started replicating my own home version of ArtificialAnalysis. Calling it the LocalAnalysis index. Literally just the parts of the Artificial Analysis bench that doesn't require a grader api. Actually harder than just sending a harness into a repo filled with test questions, which surprised me.

100

Reposted by Frank Blank

🐔 Brian Bucklew 🐔 ₑͤ＞∿＜ₑͤ ∞🌮 @unormal.bsky.social · 10/10/2026

the idea we're going to have like 100x software but with *fewer* people is very funny to me, very funny to think that and be like oh yeah thats going to be good

10809

Frank Blank @frank8465.bsky.social · 10/10/2026

holy shit lol I've been running this all day. the accuracy claims definitely hold up. right now slopping together a bench while at a goth rave. my gpus running other quants: 150-180w power draw gpus on my swift quant: 250-280w the amd r9700s likey. me likey private opus 4.8 at 100+ tok sec.

010

Reposted by Frank Blank

dawn (｡•̀ᴗ-)✧ @ptr.pet · 09/10/2026

ok wait this is peak

1242

Reposted by Frank Blank

easrng  @easrng.net · 09/10/2026

nightmare blunt rotation

618913

Frank Blank @frank8465.bsky.social · 09/10/2026

huggingface.co/frank8465/Sw... Still uploading the model & the 10M token calibration corpus, but here's the model card and repo! 8 bit trunk, 8 bit PLE, 4 bit experts, with 24/512 experts per layer left at full 16 bit. ~98% accuracy maintained vs. the full size BF16 at roughly 1/3 the total size.

huggingface.co

frank8465/Swift-1.5-Qwen3.8-Flash-Next-GPTQ-W4A8-Radiance-AMD-R9700 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Frank Blank @frank8465.bsky.social · 09/10/2026

bad: stilldeadcode (creator of the amd radiance inference runtime themselves) beat me by 12 hours on the same fine tune with a highly similar quant scheme. good: My quant is better than his according to my measurements. My calibration corpus was 10x as large, protecting more specific experts.

100

Either the local AI gods are yanking my chain, or ya boy Frank just successfully made the first Q6-equivalent stilldeadcode/radiance quantization of swift1.5-flash-next for the AMD R9700 (a terse-thinking accurate finetune of qwen3.8-flash-next by ukisai) look ma, I'm contributing to the space!

100

🎃GHOULLIKEHELLMACHINE🎃 @golikehellmachine.com · 08/10/2026

god i fucking love it when I go to a good show and then get immaculate sleep

000

Reposted by Frank Blank

like, if gaddafi had declared that all apples are now oranges and that all libyans calling them apples were internal enemies, it’d be covered in the press as a wacky but dangerous move from an unserious and dangerous clown dictator

17709122

Reposted by Frank Blank

Cat Hicks @grimalkina.bsky.social · 08/10/2026

I feel like we should be developing the general literacy that this is a form of very obvious engagement bait

1832328

sōm ⁂ @hyl.st · 08/10/2026

...whoa buddy I don't think opus 4.8/sonnet 5 ever ran that fast...

000

Reposted by Frank Blank

While code coverage is important, it's easy to game. Having ways to find deterministic bugs is far more valuable. Build a fucking fuzzer for your app.

131

need to play more in the morning, but preliminarily I'm seeing some wild improvements to decode and prefill. 4-5x over straight up FP8 weights. like 40tps/1200 to 170/6000 lol i hope it's actually an accurate model

000

I can't wait until this is done; fingies crossed its worth actually putting up on huggingface! (I might have the world first radiance FP8 trunk/ple + MXFP4 experts quant of Swift1.5-Flash-Next 🤞🤞🤞) hey looky y'all! its gone CPU BOUND! (oh god is this supposed to happen its been 24 hours now)

and ladies n gentlemen n other assortments of folks; this is castle rat