Chromadera retweeted
说个暴论,真搞不懂现在怎么还有那么多教人用 AI 的教程。
现在的 AI 早就不是前几年那种理解力有限,非得背一套提示词模板才能干活的时候了。打开软件,直接用日常语言像聊天一样把想法说清楚不就行了吗?连正常说话都要看教程学,缺的根本不是什么工具技巧,就是基本的表达和沟通能力。
真正该花时间去学的,反而是专业本身的底子。就拿写代码来说,现在的门槛根本不在怎么用 AI,而在软件工程知识。项目结构怎么拆合理,边界条件怎么兜底,还有 AI 生成的代码到底靠不靠谱、有没有暗坑,这些全得靠自己的工程能力去判断。放着正经的软件工程不学,天天研究怎么在聊天框里提问,纯属本末倒置。
Chromadera retweeted
Well chaps, that was certainly more fun than a three-day symposium on emerging market equities.
Chromadera retweeted
my multiplayer arcade volleyball game 'just volleyball' is OUT NOW🏐
catch the launch discount and pick it up for $4.95 while you can!
Chromadera retweeted
In a year from now people are gonna be saying you're NGMI if still using markdowns for your agents btw
Chromadera retweeted
This light is driving me crazy.
I gave Gemini 4 Carbon a shot through @factory and honestly, it’s a good model. This scene came out of 1 prompt with zero fixes.
Every candle is a real light. Drag one and the shadows swing, the wine throws red caustics across the cloth, and the glints on the glass are insane. 1 HTML file, no assets, all of it built in code.
YuE2 music on my RTX 3090
In 2026, you still think Docusign MOAT is their app?
DocuSign wants $8,100/yr from a 15-person team just to sign PDFs. So I built SignuDoc: open source, one-click deploy to your own @vercel account, $0 per seat.
White-label branding, full audit trails, tamper-evident seals and your documents never leave your cloud.
1
101
Chromadera retweeted
Video models will get eaten alive by LLMs.
Fable 5.2 made this anime sakura storm in code.
Over 18k cherry blossom trees, 450k floating petals, and 43 million blossoms individually rendered in less than 3 minutes on an NVIDIA L4, using vork
Chromadera retweeted
3 days of testing. 12 different quants ‼️
The best GLM 5.3 quant for 4x DGX Sparks is @wrldsuksgo2mars's K4!
His quant beats Nvidia's NVFP4 by staying closer to BF16 while being much smaller, so it fits more comfortably on 4 Sparks.
I tested about six 2.75-3 bpw quants, and they all failed my requirements.
I also tested the GR8 engine with the up1 and bit338 quants. Its default runtime uses bit325 (3.25 bpw), which would have given worse results as expected, so for a fair comparison I only compared the top 2 quants.
up1 is actually slightly better than K4 and technically fits on 4 Sparks, but the extra memory GR8 sets aside on top of it doesn't. So in practice you need TP6 to run it.
Overall, my quant of choice for GLM 5.3 is GLM-5.3-EXL3-K4-MC by @wrldsuksgo2mars (tj). Link to the quant below.
I'll be working closely with him and using his quants in other recipes too.
The GLM 5.3 recipe is in development. It won't be the fastest out there because the quant is heavy, but it will be the best quality you can run on 4x DGX Sparks.
Chromadera retweeted
if you ever want to feel like a total loser I recommend being a google enterprise customer
I can only blame myself for thinking this company could be remotely competent
Chromadera retweeted
There are simply not enough use cases being shared in local AI community, so let's see if we can change that
I have around 7~10 different models loaded up at any given time, into an environment I built called NetSphere. It allows me to give tasks to all/group/individually, make them cross-check the work, or have a leading model delegate a large task to smaller models fit for the job.
For this post, I'll show my research use case. I give my models a topic, and have them create a tree chart that breaks it down to relevant sections with sources being pulled from web and local vault. Once its been cross-check by the group for accuracy, I can generate deliverables with visuals created from my 6000s dedicated for comfyUI work.
This method also makes for a good model bench, since it allows me to clearly see that certain 9B model's research was nearly as good as GLM 5.3/QwenFN that gives me confidence to delegate similar work to it.
I'll start sharing more use cases from now, and I encourage yall to do the same



























