dpc

Funny how no one remembers about about gender equality and other liberal principles when it's time to throw you in a meat grinder. politico.eu/article/strong…

@JaWijata @GOGcom Ciebie nikt nie pytał. 🤣

@dpc_pw @GOGcom As a slav living in slavic country, i do not give that permission.

We are sorry for today’s newsletter. We made a series of mistakes: We created an unfortunate visual association by placing the runes incorrectly, we used the wrong logo for The End of the Sun, we did not check the display of the newsletter on mobile devices - which came out worse - and we didn’t port the feedback from our German QA to other languages. While half our email base received this email, we stopped the send-out as soon as we noticed the error. We are now revising our review process to enforce more checks and catch these mistakes earlier.

Yeah, I'm used to it.

Nice. gemma4:26b-a4b-qat will be my new default, I think.

@grok Oof. Zunlimited Cash.

I don't have time to look into it. Does it mean a clanker found a _new_ exploit in Zcash, or what? @grok

@THDDoric @propagandopolis Envy stays constant.

@propagandopolis Honestly, I just see things are getting better in this image. What’s even the problem?

'Metamorphosis of a class society' — Cartoon by Danish communist artist Herluf Bidstrup (1965) depicting inequality through the ages.

In a last few weeks I've been trying out many of the smaller local LLMs with a hope of finding something fast yet usable as a personal assistant that could check my emails and calendars and do some basic actions and bookkeeping. It might be me and my local problems issue, but anything below `qwen3.6:35b-a3b` is pretty much unusable for me and I don't know what people even do with these. qwen3.6:35b-a3b-apex-i-balanced-mtp gives me ~25 gen_t/s on both 12G VRAM GPU and UMA laptop, and works near perfectly: doesn't make any big mistakes, can even think through stuff like meeting conflicts between calendars. qwen3.6:27b seems even smarter, but on my hw it's ~15get_t/s or less on _heavy_ quantization, at which point it doesn't seem worth it. Looking forward to have hw that run this thing in like >50t/s, that would really open so many possibilties for me. qwen3.5:9b gets ~85gen_t/s and feels promising but makes tons of stupid mistakes with basic tool calls, and is notorious for getting stuck in loops. Should probably add loop detection in Tau, but if it can't edit a file reliably, how is it going to do anything higher-level correctly. And that's also how all weaker models do, just in a bit different proportions and frequency. gemma4:12b-q4 can go 40gen_t/s but feels like 9b. mellum2:thinking-q4 is impressively fast at 150gen_t/s but also 9b behavior. lfm2.5:8b-a1b 200gen_t/s but even more than others just a toy. Donno. Maybe someone will find this info useful for something.

I uploaded my first custom quantized local model yesterday night. I'm getting a praise about how good it is right away. What can I say, I was born on Sunday. #6a20345367f5ea856ae0b62a" target="_blank" rel="nofollow noopener">huggingface.co/dpcpw/Mellum2-…

So Iranians have nukes now, huh.

Making your own coding harness feels very much like playing Factorio. You just can't stop tinkering, experimenting, optimizing and expanding. "The slop factory must grow." Now that basic pieces are in place, the focus is on finding best higher level workflows. I'm testing whole team compositions: agents dedicated to researching, writing software, reviewers messaging each other, locking working directory, reconciling and updating design documents. Custom harness tools, fancy mechanisms that no other harness there has. It's all coming along quite nicely.

Still need to try it out in practice, but that mellum2 get insanely good numbers on my desktop on my 12G VRAM ADM card.

I might have invented a new flavor of goto-oriented programming: godbolt.org/z/7oan669Ws

Irrespective what you think about it, I guess it will happen eventually in one way or another. Maybe we could get Universal Basic Tokens? 🤣 "✨Vibecoders of the world, unite! 🛠️ Seize the means of slopification, comrades! 🚀"

@fu5ha I wasn't even aware, I just happen to be using nightly Helix. Added a mention in the README. Thanks!

@dpc_pw tried patchmark, quite cool, worth noting in the readme that it requires a nightly build of helix for lsp showMessageRequest support

After I opened the laptop in the morning I've noticed that reconnecting works and ssh is asking for password as expected.

Tau ad-hoc running it's "shell" extension/claw via ssh on a different machine. So you can run harness locally, and have it mess only with something completely different: could be a VM, container, whatever.