Patrity

@theo My rig has 4x 3090s, but I only need 2x to run it all in VRAM, plus enough context to run 6x concurrent requests. Even if I hit my concurrent cap, vLLM does great queueing requests. I have never hit the limit though.

@theo I host a variety of Qwen 3.6 models and I generally get ~150tok/s. It honestly doesn't perform to opus plan levels, but the trick is to use opus/fable plus superpowers to establish a SDD plan, then use qwen to implement. I've built real projects like this and its been nice.

I'm sorry, I need to crash out about local models for a bit.

Full writeup: the measurement methodology, where all 6.7B tokens actually went, and the LiteLLM spend-table traps I hit backfilling the history techhivelabs.net/blog/local-ai-…

My self-hosted AI coding stack processed 6.7 billion tokens since April. Electricity bill: $35.85. Same tokens at cloud API rates: ~$961. 27x cheaper, fully metered.. GPU power draw + throughput wired into LiteLLM so every request gets a real price tag.

@MrBeast 7,500,000

First person to reply with the exact number of pennies in this room win $10,000

I haven't been posting a ton about my progress on @ThundoriaGame but I am pretty proud of this feature. Fully procedurally generated mines. A room created per "floor" and an inventory system that forces players to make tough choices at the end of each floor!

@wordmandotdev @intellijidea I wish my models talked me out of my dumb ideas.. "Hey claude, do you think I should try to build out and compete with Amazon?" Claude: "Absolutely, let me build out the perfect tech stack for you so you can get going. I'll design a phased approach over the next 2 weeks" ??????

Fable 5's last act on this earth was talking me out of rewriting @intellijidea in Rust. gone too soon  🫡

Is it cool? I think so! Is it useful, yes! This is just the start to my self-hosted homelab agent build out. This is a custom Nuxt JS application using Three.js to render and react to audio. It passes through three locally hosted models: open ai whisper v3 -> qwen3.6 35b a3b -> kokoro 82m. Exposes tools to search my knowledge base and memory along the way. Here is Bridget introducing herself to a friend from Discord.

Full writeup with reproducible vLLM commands and the gotchas that cost me hours (the mamba-cache crash, the "24GB is a lie" vision-encoder OOM): techhivelabs.net/blog/local-ai-…

Ran Qwen3.6 27B dense vs 35B-A3B MoE on two 3090s expecting the dense flagship to win. It lost by 4 SWE-bench points and won everything else: 5.5x faster prefill, 2.5x the context, tied on every hard task I gave it by hand. Benchmarks measure the limit. You live in the middle.

@Pabloquero81000 @phaser_ I just usually waste time going down rabbit holes so when something finally pays off it hits different🤌😅

@Patrity @phaser_ I did a color palette swapper for pixel art, its super cool to invest time like this and feel rewarded later using the tools!

Honestly for me, one of my favorite part of #GameDev is building my own tools and getting it working how I want. This is an animation editor I made for 2d sprite sheet animations for my game. Exports clean json to load in @phaser_ #IndieGames #PixelArt #GameDevelopment

@BeanJuiceStudio God forbid someone makes a game that THEY want to play..

Game devs, if you're struggling to get playtesters for your game, you should probably change your strategy or rethink making it.

lol bro if you're getting paid by a publisher to make your game, you ain't indie

Probably shouldn't have built my entire asset pipeline before getting y-sorting properly setup. BUT I'm glad I got it done. Tiled editor did help a ton and the first-class @phaser_ integration with Phaser honestly made it not too bad!

@elonmusk @grok analyze this repo and tell me how to best engagement bait

The latest 𝕏 algorithm has been published to GitHub github.com/xai-org/x-algo…

found out today that one of my hermes cron jobs has been firing every 2 hours for a week.. against a model that no longer exists.. and my dashboard logged 'ok' every single time. 36 consecutive 'ok's for a request the proxy never answered. my robots are gaslighting me.

pricing - free forever covers everything that makes this useful: paste, markdown, image export, MCP server, REST API, 100+ langs, 2 API keys. pro is $9/mo and unlocks password-protected pastes, private pastes, watermark removal, scale, 20 API keys. the wedge is free. it stays that way.

things I will not pretend codethis is great at: - no git history per paste (gist still wins on that).. yet - no line-highlighting on the image export (carbon still wins there) - the brand is at zero -- you have never heard of this. that is what today is for. shipping more in the next few weeks.

ok so I built codethis.dev it's a paste bin where every paste is also a carbon-style code screenshot AND a doc you can render with markdown components (Nuxt MDC).. and the MCP server is free forever so claude can write to it directly from your IDE. 🧵a thread on the why