iiinova — where your digital self keeps evolving

Large Text Compression Benchmark

seems like most competitive approaches to english compression incorporate neural nets these days (and have for a while) www.mattmahoney.net/dc/text.html

mattmahoney.net

blessed lives, those, in many ways

reading older papers feels like authors had less pressure to find punchy titles and the whole vibe was closer to high effort blog posts, which honestly sounds like the pre impact-metrics-optimized sweet spot

120

*consulting a tarot reading after my zen retreat to update my posteriors* wait a second

040

37/50 oh no have i gone too far

i've been dabbling in this and it's hard :c general "recognize discrete audio events" is still way too unbounded, so people make a lot of reductions to gain traction: monophonic only, fixed timbre, etc. eventually it will fall, but the domain is still not bitter pilled

i found inner parts work can help by giving a safe environment where you own both sides (try to love -> watch yourself recoil and run run run -> love that anyway -> watch the running get slower over months, maybe give way to non-running distrust, etc. etc.)

dear Sir, it is indeed weird but it is vibe code hour and thus fuck it we ball

just a schema thing - the answer for a noul is the confidence of true

seems easy to confuse jev on open ended stuff like this by stuffing irrelevant stuff in the "state" field xD

241

brennan @brennan.computer · 16/09/2026

> i don’t really think about code anymore, it’s more about aligning subsurfaces in high dimensional vector spaces github repo is a todo notes app

000

Reposted by skelly tayne

there's a museum here that's all overhead belt systems they all still run, too

2302

maintaining remappings between backwards compatible schemas

tsfms dont solve for that search, but in best case can blackbox all the feature tuning once relevant leading variables are known

fair point - the search for exogenous variables for time series is both potentially extremely high alpha and also 99.999% of all outside information is completely irrelevant. so in a sense there's an undounded search problem baked in.

251

the thing that surprises me is unlike cat photos, you can generate tons of synthetic time series easily and have very granular controls on how to corrupt them with noise

150

though this reminds me I should get back to trying TimeFM-3, initial results were promising

TSFMs are in such a weird place - they're weirdly both high dimensional (lots of time windowing groups + exogenous features) and low dimensional (vision models deal with way more). weird they didn't get the bitter lesson yet. id guess it's a lack of annotated data sets?

150

Lies, Damned Lies, and Benchmarks - TypeSafe AI Blog

this antibench philosophy post is a bit buried, but assuming it's directionally correct, im keeping an open mind. at the model's cost level, there's a lot of room to play and experiment and find where to slot it in. it requires new eyes "where in your app can you use a fast snap judgment?"

typesafe.ai

TypeSafe AI is an AI lab building machine-native intelligence infrastructure for automation, designed to make decisions within software. Try our first System One Model, Jev, in early access.

230

austin (eeek!/ackkk! 👻) @thebadcode.com · 12/09/2026

cool that their takeaway from (a) claiming first credit on NS while (b) forcing someone else to do the low status but hard exposition work, which people like TTao explicitly warned not to do, is “everybody betray me i’m fed up with this worl”

000

Reposted by skelly tayne

so anyway i just started hackin’

71266

skelly tayne @ohnoether.bsky.social · 09/09/2026

i want maintaining a pytorch installation to be my full time job

000

skelly tayne @ohnoether.bsky.social · 09/09/2026

nix is for people who looked at python venv management prior to uv and said “this should be harder. much harder.”

130

skelly tayne @ohnoether.bsky.social · 09/09/2026

navier stokes is useless (crying in the shower)

060

i'm like an even split between awe, future shock, and disgusted by the credit-chasing to the detriment of all else if the capabilities stopped here, now, we could spend decades re-organizing around this tech. but it keeps growing

040

hm oof, that's a dimension of deepfake i hadn't considered. private llms dont help the defender substantially here. hard to see a way for the anonymous internet and these tools to coexist.

i'm mostly not thinking about cybersec here, but about weaponized purposes like oppo research and reputation attacks. having your own llm doesn't help mitigate the damage.

Ah you seem to have cybersec in mind, I think we're in agreement there. I was thinking of social media "oppo research"

still don’t know how people can consider these scenarios and not hesitate to boost pushing frontier level capabilities to consumer-grade open-weights models. if current social media warfare is using swords, it’s like introducing killer drones with machine guns.

230

fixes for jank have been available, just not evenly distributed. maybe this tech can help

010

a coordinated mobo and gpu firmware update and all problems gone. two years ago id either just accept it or stumble on an obscure forum post from someone with the same issue and, if lucky, copy their solution, or, if unlucky, struggle to reverse engineer their fix from their cryptic follow-up

100

not seeing enough discussion of LLMs for debugging weirdo hardware issues. did i know my video card had an outdated VBIOS? no. was i just going to live with occasional video glitches and hiccups? yea probably. but claude dug into the crash logs and determined the exact firmware fix

100

Making me think about those 39 bits saying a lot more between experts (i.e. people with shared codebooks). And what that means when the two communicating parties can assume a shared codebook of effectively ~all of the internet.

all good, was just responding to the plausibility of going much faster today. not sure youd actually want to spend 10s of thousands to beat a game from 2008, but seems to me like it could be done. which is quite a step change from my priors seeing claude plays pokemon from what like 3 years ago

think i still secretly hope for a plateau, just to give us time to catch up

000

maybe? idk enough about the details of the portal run. not trying to do a gotcha point here, just sharing a different point of view on this rapidly evolving and unevenly distributed tech

skelly tayne @ohnoether.bsky.social · 07/09/2026

my read is it's all supply - i doubt any Pro/Max subscription-level consumer is getting priority on those chips, if at all. the typical estimated output from consumer-level models is ~50 tps (not accounting for startup latency), so assuming that, 20x with today's hardware is possible with more money

100

checkout the 1k tps cerebras chips openai is using to serve <undisclosed 1T parameter model that likely includes astra>

111

skelly tayne @ohnoether.bsky.social · 07/09/2026

no way no way this is incredible