iiinova — where your digital self keeps evolving

daniel @daniel-j-h.bsky.social · 15h

Reposted by daniel

ponder @ponder.ooo · 22h

the future of gaming is that your friend yaps at you about their favorite new game, the one their agent made just for them, tailored to their preferences, and you zone out because why would you give a shit, all you want to talk about is your favorite new game, the one your agent just made for you,

713410

all layers on gpu? speculative drafters set? flash attwntion on? compiled for nvidia backend? there are sooo many things that can go wrong, it takes quite a while to tune this; could try auto-research an optimal setup overnight

130

daniel @daniel-j-h.bsky.social · 16h

awesome work, would the same work by fine-tuning small projection heads on top of the frozen encoders? what are the pros and cons vs this approach here?

000

daniel @daniel-j-h.bsky.social · 17h

TIL how to pronounce dinosaur

000

daniel @daniel-j-h.bsky.social · 17h

Have you seen the related work in ghostty itself gist.github.com/jake-stewart... hachyderm.io/@mitchellh/1...

daniel @daniel-j-h.bsky.social · 17h

the hardest part of AGI will be getting audio capture to work reliably on linux

daniel @daniel-j-h.bsky.social · 09/10/2026

y no vulnerability-fixing tho

daniel @daniel-j-h.bsky.social · 09/10/2026

let's grep and find through code bases like it's 1999

000

daniel @daniel-j-h.bsky.social · 08/10/2026

they just opened up how they did it github.com/goodreads/math

github.com

080

daniel @daniel-j-h.bsky.social · 08/10/2026

but can opus 5.5 call my parents saying that I love them

120

daniel @daniel-j-h.bsky.social · 08/10/2026

How much faster is it? How much better is it? How much money does it save? How should we accept this without any serious evaluation?

000

wonderful! what's latest in terms of approx. nearest neighbor searches on MaxSim / late-interaction re-ranker outputs? regular ann libs like faiss won't do, no?

100

z-lab/Qwen3.6-35B-A3B-DFlash · Hugging Face

have you tried the dflash speculative drafter huggingface.co/z-lab/Qwen3.... depending on your workload that should speed things up a bit more 😍

huggingface.co

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

oof, in Germany I see it starting at ~4k EUR without any disks or extras how's the thermal regulation? On a bosgame m5 I have seen it getting limited to 80W after a few minutes as it's getting too hot that and the slowdown from large context makes ds v4 (not even 4.1) barely usable there

350

Also re-rankers and especially late-interaction re-rankers e.g. of the colbert kind. That's on the way from embeddings to llms but seems like not super well known?

daniel @daniel-j-h.bsky.social · 06/10/2026

Does that mean passing a watermarked text through another LLM and having it rephrase things a little gets rid of watermarking; while for images this will not work by design?

100

woah can it choice on morse code 🦧💡

Where's GLM 5.3, MiMo v2.6, DeepSeek v4.1 in those charts tomcompare. They're at roughly ~300b params. There's a table that shows some of them in comparison if you scroll all the way to the right. Why try to bend reality. This comes out sooner or later anyways.

~500b is also quite the chonker the competitors (deepseek v4 flash, glm 5.3 flash, mimo v2.6 flash) are all around ~300b moes they're amazing and wonderful for self-hosting on datacenter grade gpus (single b300 for example) or even unified memory devices

and for extra tinfoil hat points you can self-host both clef models, too! the flash one fits on a laptop or mini-pc; no dedicated gpu required

daniel @daniel-j-h.bsky.social · 04/10/2026

Same for vision machine learning. Everyone trained and benchmarked on ImageNet but no one asked about the licensing aspect of those images. Maaybe fine for research but at some point this stuff turned out to be actually useful but we still kept training on these images (and later also videos).

0110

16 GB is limiting; recent open weights models are really good but they're not that small try the gemma4 e2b, e4b, or 12b versions and see how things go the problem is the good ones like deepseek v4 flash, glm 5.3 flash, mimo v2.6 flash all require 128 GB or more while being slow to run

020

daniel @daniel-j-h.bsky.social · 04/10/2026

it makes me less scared starting a project connecting to IOT devices via BLE on Linux for sure

100

daniel @daniel-j-h.bsky.social · 04/10/2026

Fascinating to see it pulling OpenStreetMap data through Overpass. It's meant for simple queries. Interesting to see if the OpenStreetMap people see more load due to agents.

this, and also qwen3.5 is a prev. prev. generation model omg what are they doing over there

i'm just sad it's yet another polarizing topic as if we don't have enough of that already; we need more nuanced takes and less us vs them

what's the economic incentive to do any of that?

qwen3.5, gpt oss, .. guys we're in late 2026 what is this

140

what pareto frontier are they talking about

120

you're already ahead of most people out there using jev if you do this 😛

000

#jev is fun! you don't have to care about evaluating it on your domain's test set, understand how and when it fails, bias, or explainability. We all decided to yolo it this time.

calculators they're putting them in there

Will Oremus @willoremus.com · 02/10/2026

Reposted by daniel

feels like there is a group of smart and well-intentioned people who long ago made up their minds that AI Is Bad, swore off using it, and now get all their info about AI from each other. not saying they're wrong that ai is bad! but i think they're gradually losing touch with what it is and does

235966105

anyone tried distilling Jev into a modernbert based classifier??

000

👸✨🧎🏻🧎🏻🧎🏻

looks like clef flash is qwen35 9b from 6+ months earlier this year and I guess there's just hasn't been an upgrade in that size category unfortunately

000

told you they're putting calculators in there

wariziles @riziles.bsky.social · 02/10/2026

Reposted by daniel

Please stop anthropomorphizing computers. I cannot "turn on" a computer because I just don't have that dog in me.

413018

pi is strategically important for open source and the open llm ecosystem let's see how things are progressing especially with commercial models getting rl-trained on their harnesses