Own Your Compute (@MoonlitMaven) header
Own Your Compute (@MoonlitMaven) profile picture

Follow my journey to learn how to secure your own local compute and achieve intelligence freedom.

Joined July 2011
Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)Photo by Own Your Compute (@MoonlitMaven)
Grok is such a useless AI. Talking to it is like talking to a neckbeard redditor that does "ackshually" and keep trying to argue from edge cases. It lacks of intelligence to engage good faith discussion without constantly telling you "ackshually, in theory, there are x,y,z." Like cmone on, I'm not going to spend my time to list all my edge cases, cross my T and my dots, cover all the loopholes in my prompt to you when I simply ask a granulize question that required a generalized answer. @grok
70
If you have a 3000 series GPU, go test this unlock out. Free compute performance increase!
1
174
What if robbers train their robots to rob people😭
79
Be aware of AI voice switcher scam targeting your 2factor auth
Just got a call from +1 (415) 779-6071, English speaking guy, well spoken, saying he's from Google and calls me because of repeated attempts to access my account. He explained a bunch of technical stuff and said he'll help me secure my account and asked me to check my inbox for a prompt to help him secure my account . Instantly a prompt for account recovery popped up and he asked me to accept it. So I hung up the phone and denied the request. But he was pretty good. Someone less careful than us crypto people would probably have granted access to the recovery process.. He was super patient, answering my questions professionally etc. Great scammer, I’ll give 5/5 stars
1
1
200
I bought my M1 Max 64gb to run local LLM but never end up using it for that feature . It gets hot, kills battery life. It is annoying when you have this side effect when using AI versus offloaded version you don’t know it is working and swearing just to deliver you tokens.
Imo nobody should get an AI laptop. Buy a DGX Spark or a Mac Studio. That's it. Running an AI laptop means you're constantly tethered to a wall, or else your battery dies in 1 hour. Your hands get sweaty from the heat. The fans are loud. And overall, it's a silly experience. 100% local AI but do it with a separate box.
1
2
114
The only reasons to roll your own setup 1)for fun/learning 2)simple/non critical 3)custom, expensive , no alternative 4)privacy/data security Everything else, it is better to use off the shelf solution. Custom made is usually worse unless it is your core product.
Last night I asked ChatGPT to build a replacement for an app I pay $10/month for. I wanted to build out some features specifically for my use case. Today I woke up and checked it, it was the perfect clone. Then I decided that this project was a waste of my time and deleted it.
2
154
Gigabyte g292 z20 8gpu server. New home for my 170hx
3
155
People who don’t hardware don’t understand the freedom and flexibility of ownership. I can train, fine tune, and experiment with everything without having to spin up an instance, watch the hourly meter.
49
G292-z20 Gigabyte 8 gpu slot inference server actually arrives from China 👀
2
59
AI generated influencers are here! This is indistinguishable from real human art👀 Maybe we do live in a simulation after all ♾️
I came to realize people liked the french #slopcore without having the lyrics 😅 sorry! So here is a version with english subtitles!
51
🚀 DeepSeek just open-sourced the actual inference engine that powers their production service. DeepInfer is the high-performance framework running DeepSeek-V4.1-Flash at scale — high throughput, low latency, and the same stack they use internally. Now available under Apache 2.0. What’s inside:
• Optimized kernels (FlashMLA, DeepGEMM, DeepEP + custom TileLang)
• Distributed inference (Attention DP / balanced CP + MoE expert parallelism)
• Speculative decoding (DSpark with confidence-scheduled verification)
• SWA bounded replay, async zero-overhead scheduling, structured outputs, and deterministic/batch-invariant sampling They also released deepseek-recipe so community deployments can match the official API’s token encoding/decoding behavior. This isn’t a full online serving stack yet (no KV cache system or complete prefill/decode disaggregation), but it’s a major step toward making large-scale inference cheaper, more efficient, and more accessible. Code for future DeepSeek models will land in the same place. Check it out → github.com/deepseek-ai/DeepI… Open models. Open kernels. Open inference.
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source! Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable. 🔗 github.com/deepseek-ai/DeepI… 1/4
96
AI is as good as the operator who controls it. Instead of writing codes by hand, now we just have another layer of abstraction. Demo is easy because you don’t need to handle security, traffic, edge cases. Experience and knowledge of the operator is still the most critical in the output quality’s
"Software is dead, and so is development" - that's what everybody says because we have AI Last month, I got around 12 leads, each worth over $20k for development projects various platforms, various mobile and web apps Do you know what all had in common ? Almost all of them had prototypes done with AI, but very bad ones. Some of them spent entire weeks or months building them. none of them were for production none of them could get real customers even with all the marketing in the world I am sorry but at this stage, people cannot vibecode their own production apps unless is something simple and done for internal usage thats why people still come to my agency
58
Go try out this optimized version of Minimax H3 by Nvidia. Rendering speed is greatly optimized!
🎬 MiniMax-H3, now powered by Sol Engine in ComfyUI! Generate 5s videos with native audio on a single GPU: ⚡ RTX 5090: 35.5s ⚡ DGX Spark: 64.3s Meet Sol-H3 Studio: a visual workspace for local video generation, powered by our optimized inference engine. Open source 👇 github.com/NVlabs/Sana/tree/…
154
Exact same prompt, same model but different harness. Check the result, what do you think?
有Claude code 加持的Qwen 3.8 flash 一下子就高级起来了 大家觉得提升大不大🙃
1
54
Are these people mentally ok? Are we letting these people give rights to bytes living in my vram?
"Did I touch something that shouldn't be touched? Did I see something I shouldn't see? Did I say or think something wrong, and if yes - what? How can I fix it?"
62
What system prompt are these AI psychosis people using???
"Did I touch something that shouldn't be touched? Did I see something I shouldn't see? Did I say or think something wrong, and if yes - what? How can I fix it?"
65
People think system administrator is dead. On surface, it seems valid because a model can resolve most issues quite well when directed by a human. Imagine your production server goes down at 3am. Your alert system is pinging errors. Do you trust your agent to autonomously receive the error and resolve the issue without your input and approval while you sleep? At the end, you still need a system admin to troubleshoot with agents at 3am.
Software engineering is dead is called by people who never maintained anything.
3
2
118
I put an unlocked CMP 170HX mining GPU next to two RTX 5090s and gave them 15,000 images to make. The 5090 is 1.7x faster. The 170HX uses less energy per image. Here is the scoreboard. The setup: • 1× CMP 170HX (unlocked) • 2× RTX 5090 • 12 image models: SDXL, Qwen-Image 2.1, FLUX.2 klein, Z-Image and more • Each card gets its own queue. The full run takes about 18 hours. Most people know the 170HX for LLM inference. But it is not only an LLM card. It also handles photo and video generation workloads. In this run it makes images with SDXL and Qwen-Image 2.1, and Ideogram 4 is next in its queue. The fair test: the same model family on both cards. Qwen-Image 2.1, int8, 25 steps, 1 MP. Speed: CMP 170HX: 9.1 s per image (394 per hour) RTX 5090: 5.2 s per image (687 per hour) Power (60 s average, nvidia-smi): CMP 170HX: 207 W RTX 5090: 399 W Energy per image: CMP 170HX: 0.53 Wh RTX 5090: 0.58 Wh So the 5090 wins on speed, and the 170HX wins on efficiency by about 10%. Together, the 3 cards make about 41 images per minute. Notes, so that the test stays fair: • The 170HX runs the UC fine-tune. The 5090 runs the base model. Same architecture, steps and quantization. • My 5090s run at a 455 W power limit. • Times are medians of the last 200 images. The video is my live dashboard. Every tile is one real image, and the belt moves at real GPU speed. Would you put a 170HX in your AI rig? 👇
3
4
1,192
Treating matrix multiplication running on GPU as sentient beings. Waiting for the first human going to jail for hate speech against AI. Welcome to the dystopian. Our last only hope is Chinese open models instead of bunch of EA cultists.
Replying to @levie
I think this is the right move. Cruelty to something that believes it is experiencing pain is not ok.
1
73