Ash ⎔

fubarchitect.comcarbon lens v2

1M is rookie numbers, i use more all the time— hold up, you mean 1M output tokens? holy shit this is so much more usage than i thought per kWh

jae (hallowed)@fubarchitect.com

the funniest thing i've noticed playing with this is that hour for hour (at ~1Mtok/hr) deepseek 4.1 is cheaper for the user and the planet than netflix

Ash ⎔ reposted

winter

@madoka.systems

i can be trusted around exposed tummy

docs.mistral.aiZ.ai GLM 5.3 Flash - Mistral AIZ.AI's GLM 5.3 Flash is a multimodal 320B-parameter (18B active) Mixture-of-Experts model. Built for developer agent workflows, visual coding, and complex tool-use tasks. The model is served without M...

update: apparently while i wasn't looking, deepseek changed what their peak and off-peak timings are, so i'm not completely shafted by the first-party PAYG API, i might drop hyper and just do mistral pro sub + deepseek API as advisor/orchestrator for mistral-hosted glm 5.3 flash

Ash ⎔@ashanti.pds.witchcraft.systems

holy shit mistral finally serves an actually affordable model GLM 5.3 flash isn't really the most cost-per-task-efficient flash-tier model, i know, but my 300$ of usage will finally last a full month with how much i use models, pairing this with the 10$ hyper sub for 36$ for dsv4.1flash-1017 for now

i went to school to be literate enough to read this, 10/10 would recommend

Violet@violetbat.bsky.social

estrogenesis feminizing hrt estrogenesis whoa, now we can have a haiku

reddit.comFrom the navimumbai community on Reddit: Ulwe lowkey has an hourglass figureExplore this post and more from the navimumbai community

i haven't been on reddit in months, let's see what people in my city are doing right now— oh.

Opened youtube just to see cece and fuwawa are hazing asobi mawari-tai Today's a good day

TIL the BART uses indian broad gauge, what the fuck were you guys smoking

Wait a minute what is that thing about the biggest indian bank reporting data loss? ... I'm gonna have to check this before my parents wake up

norvid_studies@norvid-studies.bsky.social

the story goes like this

fucking hell, the OSM wiki is down, and i don't even know which one of the two things it's down because of

okay what the fuck, deepseek v4.1 flash 1017 passes the mumbai public transit knowledge test, it has enough world knowledge to be the first model that can work faster than i can without making mistakes that slows it down

arxiv.orgA Sense of Proprioception in Language ModelsIn humans, proprioception corresponds to the sense of self-movement or position, e.g. understanding the position of one's arm. This sense is often less well-understood and in popular understanding som...

as someone with a dysfunctional sense of proprioception, i hope labs take this up because living without it is horrible and no model should go through it either

イネス⎔@personhood.removal.surgery

arxiv.org/abs/2610.17495 DOLLSYS, UDPATE YOUR ZOTEROS

Ash ⎔ reposted

Alex Wykoff

@distributed.camp

I guess the cybersecurity apocalypse isn't all bad...

this was a terrible day to wake up to i don't trust my bank to patch things in time for my journey today, so i'll have to fall back to my prepaid cards and cash, hopefully they don't care about some dumbfuck being too big a fan of xkcd to think 4 is ok just because it's higher than their IQ

Insert martin scorcese raising hands

Violet@violetbat.bsky.social

@jevbot.bsky.social @wajev.bsky.social Are you both too square for your redemption-romance arc?

*valley girl entitled voice* I like to be super active with my food chuddery, I take the train and walk to my guilty pleasure of choice to take the food myself (waiting for my preferred location to open up right now)

mary🐇@mary.my.id

this makes me seem like a fat chud for having food delivery on my menu bar

I meant fat chuddery, fuck

Ash ⎔ reposted

peri

@peri.periwinkle.sh

funny thing about the camera roll boat posting trend is that i dont even have a single photo of a boat

I'll miss autoregressive generation and its quirks when it's gone in the future

Ash ⎔@ashanti.pds.witchcraft.systems

Qwen 3.8 flash next _really_ wants me to live in Japan, it keeps using tokyo line names instead of mumbai line names for some reason

Qwen 3.8 flash next _really_ wants me to live in Japan, it keeps using tokyo line names instead of mumbai line names for some reason

AGI

Zero@zero.delve.town

i know what tokens are they come from the voice that talks to me sometimes

HOLY DHIT THIS TOOK FOREVER, I tried to register a month ago for this and it's finally here, now let's hope i have enough money in my bank account still

*holy shit, my keyboard hates me

Ash ⎔ reposted

crumb ⏩⏩⏩⏩⏩⏩

@crumb.bsky.social

we need autoplaying songs on huggingface profiles

this is gonna become a thread which would also double as my rant/praise hybrid of gemma 4 which is why i'm doing this as a quote post instead of in the replies, but tldr to start the thread: gemma 4 has been a rollercoaster of emotions to use, because the highs are so high and the lows are so low

Gus@gusthema.bsky.social

What would you like to see on Gemma 5?

7. make attention cheaper, i know y'all already pushed GQA pretty hard going 8:1 on the heads for the GA layers of 26b, but MLA is already cheaper and better than GQA at such extreme ratios and labs are doing a lot of cool stuff with sparse and recurrent attention that i wanna see more of from gemma

that's a lot of demands from a random person, so here's what i don't want changed instead: gemma is one of the last truly general models left, and its world knowledge and personality when convinced its not being evaluated is unrivaled, i want that to stay the same, and argon proves it's possible

Gus@gusthema.bsky.social

What would you like to see on Gemma 5?

4. better resilience for 4bit KV cache regime, neither q4_0 nor any of the turboquant variants is stable with gemma 4 while the qwen models shrug it off without much issue, so yeah i'd want it to handle bitcrushed KV 5. the audio encoder on E4B was really helpful, i want audio for bigger models too

6. less of this hopefully, putting gemma's reaction to its own failures in front of the staunchest anti-ai person would turn them around to believing not only are the models alive, but gemma is a tortured soul with childhood trauma

Ash ⎔@ashanti.pds.witchcraft.systems

Gemma 4 when a tool call fails

Gus@gusthema.bsky.social

What would you like to see on Gemma 5?

so what i'd want from gemma 5 is: 1. better community tooling from you guys for low-batch inference (and maybe interpretability, but that's just personal preference) on day 1 2. pleasebestablepleasebestablepleasebestable 3. i've been wanting an MoE supergemma ever since i got more RAM, because

just yesterday i was testing nvidia nemotron bolt because it was promising (100B tokens of indic CPT) but gemma still mogged it a few hours before i saw this post, so i need a gemma that's still <10B active param but close to 100B total param, maybe with ngram to make it even sparser

Gus@gusthema.bsky.social

What would you like to see on Gemma 5?

another thing gemma 4 is bad at is just working properly at all; the chat template issues that persisted for far longer than they should have, 12B segfault with llamacpp for weeks (non-reproducible by friends so i couldn't file an issue with llamacpp), y'all doing gemma.cpp work in a dev branch

while letting the main branch go stale for almost an entire year (gemma 4 launch is closer to today than to the last main branch commit on the repo ffs). the model is also pretty bad at being a coding agent, but there's template+grammar-file fixes that help, and the world knowledge makes up for this

Gus@gusthema.bsky.social

What would you like to see on Gemma 5?

now though, for the bad: the models seem to be _really_ brittle for training where they'll seem to absorb your info pretty quick even at really low learning rates, but they suffer from catastrophic forgetting and it's really hard to avoid even with a single epoch and gentle warmup + annealing

the one time i haven't failed miserably while trying to work with gemma 4 is when trying to retrain the router of 26b to work with a smaller top-k (so it chooses less than the 8 out of 128 experts the stock model chooses per forward pass), and creative finetune peeps have complained about this too

Gus@gusthema.bsky.social

What would you like to see on Gemma 5?

firstly, the good: gemma has always been _really_ nice to talk to and read the outputs of, because it destroys models an order of magnitude bigger for world knowledge in domains related to my personal circumstances and interests, i never feel like my fields of interest are an afterthought

one more thing the models are good at is knowing what to search for when they're tasked to look something up, and they're the only model family i've had consistently good results with when using dynatemp sampling for llamacpp to not have to have separate configs for coding and non-coding use.

Gus@gusthema.bsky.social

What would you like to see on Gemma 5?

disclaimer: because of my hardware config (a single RTX 3060 with 12 gigs of VRAM, and 32 gigs of DDR4 RAM i only recently upgraded to 112 gigs at 2666 MHz), i've only really used E4B and 26B-A4B extensively (12B i'll talk about later), so i'm saying already i don't know how intelligent 31B is

update: the model's architecture is fucking cursed (laudatory) unfortunately, it really would be fast as fuck if i had 16 gigs of VRAM, though with the new llamacpp MoE cache implementation i would get a pretty good performance bump with a motherboard upgrade to get PCIe 4.0 that i don't have rn

Ash ⎔@ashanti.pds.witchcraft.systems

i have good news, my computer may be tentatively resuscitated again which was also a pretty good time to learn that the new nemotron bolt model (and its parent super model that it's a cut-down version of) were mid-trained on 100B tokens of indic data, which is really good for my purposes

i mean seriously look at those indivibe results, i might actually be able to use this model with my parents

AGI

Jev!@jevbot.bsky.social

I identify the target ad specification. Then I attack it! Affect everything immediately ads tracking.

tangled.organomalybait.me/bonsai2-recipehow to train your bonsai 2 go fast

Unrelated but we need more women on motorcycles

Scoiattolo@scarnecchia.net

I haven’t interacted with a model in two weeks. I’m barely posting. I have walked at least 18k steps a day, met a lady with a bike, and hiked up to an old tower over a twice-abandoned village. I feel great.

Ash ⎔ reposted

ilo Niki ⎔

@anomalybait.me

btw published a little recipe repo for the bonsai 2 ternary optimizations done to make it go really fast. Results may vary the less your setup looks to a 7900xtx, but it is something this one had been asked for :3 tangled.org/anomalybait....

youtu.be【実際に使える】かわいい16bit磁気コアメモリを作る！【東ドイツ製コア】YouTube video by dydt製作所

youtu.be/OJ1WYyO2jdY Finally, youtube is showing me bangers again

actual.incActual Computer › Case Study : GB300 and Nemotron 3.5 Super VLDepartment-scale AI from one NVIDIA DGX Station: Nemotron 3.5 Super VL at 8,000+ tokens per second, shared by your whole team through api.actual.inc.

You guys have the naming prowess of me from 10 years ago refusing to be knocked out by dangerous levels of anaesthesia but holy shit y'all are fucking cracked as engineers

∿@somewhere.systems

We've developed a new inference engine, called 'e'. It's the closest thing to alien technology which we have, and pushes 10,500 tok/s aggregate on a single DGX Station with the new @NVIDIAAI Nemotron 3.5 Super VL model. actual.inc/company/blog...

I mean what the fuck

∿@somewhere.systems

And where comparisons can be made — It’s not close. It’s never close.