And Latency can be optimized even more running prefill on the image data while the user is typing the prompt, but even without that it's 3-5s TTFT

Latency is much better for me than any other UI, it's really fast. Running Kimi K2.5, gets me 100tok/s on a bigly loud GPU server.

The newer Qwen models though should be viable for a more reasonable high-end home GPU setup. Probably in 40t/s on 2x Sparks. Or you can plug in an API provider.

Another one of small 15 minute hacks to improve my workflow now that local multi-modal AI is good enough.

github.com/magik6k/scro...

Trigger a hotkey, select screenshot area, ask about a thing, fully private and super snappy

GitHub - magik6k/gscrape: A toolkit for scraping, indexing, and analyzing Go pprof goroutine dumps over time across services. A toolkit for scraping, indexing, and analyzing Go pprof goroutine dumps over time across services. - magik6k/gscrape

So easy to fix things with the recent wave of agentic coding tools like OpenCode. I was debugging a fairly complex issue involving 10s of goroutine trees and over multiple hours

Previously this would be a few jank bash scripts, now it's a nice tool I could opensource - github.com/magik6k/gscr...

The brew guide from GLITCH comes with a grind size reference in a small bag (left). Seemingly wants me to set the grind setting on my Niche Zero to “a different grinder”

AVR-ISP Pogo Pin Adapters It's a small Pogo pin ISP programmer adapter. It allows you to program your AVR project without to solder ISP header and reduce space on PCB.

You can shorten the programming step by 10x with some fancy Pogo Pin adapter, makes it super easy to also automate basic small production run self-tests

hackaday.io/project/1650...

More ATTiny is probably cheaper than a dedicated gpio extender. Actually you can probably also replace the buffer with another ATTiny.

I basically use it as a LUT in some of my designs, it's cheaper than most other components you can put on a PCB, and can replace all non-high-speed logic

Like, looking at the supported device list even 100MiB of repos would seem extremely overkill - handbook.buildwithmatter.com/howitworks/d...

git clone --recursive github.com/espressif/es...

Cloned 18GiB of repos, failed because it needs git-lfs.

WHY IS THIS MUCH CODE NEEDED TO COMMUNICATE THAT A LIGHT SHOULD TURN ON??

Just -fprofile-generate, then run in the benchmark mode, which generated mine_nogmp.gcda, then rebuilt with -fprofile-use

Actually now with coffee in front of me I've realized that I wasn't running your thing 🫣

7995x (192t):
nogmp: 398M/s
nogmp+PGO: 503M/s

Ran on a 7995x: 214M / sec.

Maybe instead of fighting with RDMA and the annoyance that is closed source nvidia drivers on Arch, I can just pipe the 4 Display Port data lanes straight into a qsfp28 port?

Both have 4 differential pairs, with 100ohm impedance, with roooughly similar signal levels, so maybe just maybe 🤔

ham meme (original) YouTube video by burialgoods

www.youtube.com/watch?v=cYxI...

Flashing random firmware and casually passing the "yeah don't worry the hardware id is different" flag will never feel nice

Current boss:

Convincing libfabric to see my NICs on Arch Linux

Current boss:

Convincing libfabric to see my NICs on Arch Linux

Bluesky’s science takeover: 70% of Nature poll respondents use platform Roughly 6,000 readers answered our poll, with many declaring that Bluesky was nicer, kinder and less antagonistic to science than X.

Roughly 6,000 readers answered our poll, with many declaring that Bluesky was nicer, kinder and less antagonistic to science than X

https://go.nature.com/42tH8Ai

Very clearly this shows that the past is now all in the past, and the future is now in the near future.

chatgpt ui with https://neal.fun/stimulation-clicker/ added to the waiting box

I really need a browser extension which adds neal.fun/stimulation-... to ChatGPT Pro progress bar. Otherwise I fall into other time-sinks waay too easily

We’re building the atproto TikTok and IT WORKS!!

Spent the last 48 hours coding!

👇🏾check it out

Follow along
@skylight.social
@reedharmeyer.bsky.social

#atproto #atprotocol #tiktokban #skylight

Absolutely incoherent code from libtrf which is a wrapper for libfabric. Code is good, but I have no idea what it does

I have borderline zero idea what all those random names mean/do, but the thing compiled now. No idea if it works tho.

Pretty fun to enter a new technology and feel absolutely clueless.

Next step is to get LookingGlass setup running, shouldn't take long, but I'm also tired enough today.

Absolutely incoherent code from libtrf which is a wrapper for libfabric. Code is good, but I have no idea what it does

I have borderline zero idea what all those random names mean/do, but the thing compiled now. No idea if it works tho.

Pretty fun to enter a new technology and feel absolutely clueless.

Next step is to get LookingGlass setup running, shouldn't take long, but I'm also tired enough today.

Oh yeah, I run Arch (btw), and of course nothing can compile with the most latestest libraries.

I guess we contribute some to open source tomorrow ¯\_(ツ)_/¯

Added a cable between computers, setup OpenSM and now it links up!

Benchmarked at ~60Gbps with ib_read_bw. The bottleneck is explainable by.... PCIe on the non-threadripper end of this setup

LnkSta: Speed 8GT/s (downgraded), Width x8 (downgraded)

It's ALIVE!

Now just need to put a super jank fan in the other computer and do the same setup there.

It's ALIVE!

Now just need to put a super jank fan in the other computer and do the same setup there.

Bios picture with “network link type” pop up, InfiniBand selected, Ethernet is the other option that used to be selected

Also this says InfiniBand now, no weird packages needed!

Deeeefinitely needed a fan, was ~80-90c to the touch, maybe more.

Now taken care of