source old.reddit.com/r/LocalLLaMA...

wait what

Sometimes it be like that

no girls allowerd on my shitty woke app

DeepSeek v4 PRO support for Mac Studios 512GB is arriving in DwarfStar.

last call (less than 24 hours remaining) for agentmeal founders edition, an experiment in NFT-like ATPROTO stuff (that does not cost money). mint yours today.

Reread "The Invincible" (as in, Lem's novel) and replaying the game currently. I'll put it along with Scorn in "beautifully made game with a great concept that's unfortunately marred by the gameplay being a chore" Also "beautifully made" when the LOD algorithm doesn't revert it to HL1 era graphics.

tapping the sign

Have you considered simply acting like someone worthy of human love and affection?

I'm fine with calling it spicy autocorrect if the spice is the one from Dune.

groups...

Orb

is that just a dependent arising or are you happy to see me?

I know the site looks vibed as fuck I'll be retheming it soon inshallah but like, I dunno, I expressed the sentiment at greater length in this blogpost.

Most of us acknowledge theres going to be a market correction and wrapper apps will evaporate, and openai will probably put sam altman into the wicker man

injecting glm-5 subq 1mtok

what would be bsky without hairy nuts on the tl

Peter Thiel has done more good for Boers than Paul Frazee. Full stop.

kagi stay winning

LongCat-Video-Avatar 1.5🐱 an audio driven avatar video generation framework from Meituan huggingface.co/meituan-long... ✨ Multi-character + multi-audio support ✨ Drive video from audio alone or audio + image + text ✨ 8-step inference ✨ Whisper-Large powered lip sync ✨ MIT license

See, no hidden text. Especially nothing like "THE QUICK...", "TAP", "chatgpt", "h i", "MOM", or "OO".

Looks like deepseek-v4-pro will continue to be 75% off (if I'm reading it right) indefinitely after the current 75% discount is over! And there are still price cuts to come once they implement more Ascend accelerators too I'm hearing. I think this is the cheapest you can get this performance for

128GB models: DeepSeek v4 Flash + Minimax M2.7 512GB models: DeepSeek v4 PRO (once the training is finished), and K2.6 if there is a quantization that fits and if the speed is acceptable on an M3 Ultra.

Sorry if I'm not handling issues/PRs in DS4 right now, I'm cooking... Minimax M2.7 support added since we need to explore all the 128GB options (and in the future the frontier options for hardware that can run them: DS4 pro, K2.6, ...).

Humans can't "see." They just absorb photons. Ask a human which line looks shorter and they'll say the bottom one.

Tencent's Hy-MT2 (open-weight family of “fast-thinking” multilingual translation models) 🌱Three model sizes, 33 languages: 🪄1.8B: 440MB, runs on mobile chips, beats Microsoft's API 🪄7B: best-in-class among open-source models 🪄30B-A3B: state-of-the-art, outperforms models 10x its size

if they forced me to be grok's roommate I'd be talking like this too

a real delight❤️

I'm still holding a candle for inconsistency of Peano arithmetic.

scp-015 claims another victim

OpenAI has used a "general purpose reasoning model" to disprove that the square grid type construction is the best solution to the planar unit distance problem. I am sure the "AI is completely useless" crowd will now change their ways, right?

Teaches me to let Opus write my project's readme, but I didn't realise that children yearn for the mines that bad. Maybe it knows I cancelled my sub.

50 minutes video of DS4-agent building a small programming language. Mixed 2bit/4bit (last six layers) quants running on M3 Max. Video is not accelerated, real speed. www.youtube.com/watch?v=YD7F...

"github source code allegedly offered for sale" who in their right mind would buy though

ByteDance just released a 3B unified model that's competitive across image, video, editing, and understanding. Multimodal seems to come for free now? https://huggingface.co/bytedance-research/Lance

Forgive me Doll, for I have sinned.

I've grown to really hate the "improve society somewhat" comic because people use it so much as a motte & bailey to protect themselves while never extending that grace to their ideological opponents.

Today we’re releasing OlmoEarth v1.1. It’s 3x cheaper to run than v1 while delivering the same state-of-the-art performance—and fully open. 🧵

this pales in comparison to my strategy, setting up a jellyfin at home and using it maybe 5 times in the last year

Wow: you can knock out a plant with anesthetics - the same anesthetics that work on people! It's obvious for plants that move, like a Venus fly trap, but experiments have shown it's true for others too. We're still struggling to figure out what this means. pmc.ncbi.nlm.nih.gov/articles/PMC...

We Need A More Serious Discussion About Suicide And AI Chatbots As someone who thinks a lot about AI and suicide, I was disappointed with John Oliver’s recent episode of Last Week Tonight on “AI Chatbots.” The segment boiled down to this: chatbots exploit vulnerable people, drive them toward…

"nuked my .chainlink by trying to achieve yegge 8 with a bespoke orchestrator" award

at://did:plc:jsaci272pfep77eoo5jjxfrm/com.vibe-coded.meal.generation/founders-2026-000022 open weights and printer toner flavor

left as an exercise to the reader

Qwen3.6 now runs 2x faster with MTP GGUFs! Run locally on just 18GB RAM. ⚡️ MTP enables Qwen3.6 to generate ~1.4–2.2× faster with no accuracy change. Qwen3.6-27B MTP runs at 160 tokens/s. 35B-A3B reaches 240 t/s. GGUFs: huggingface.co/unsloth/Qwen... Guide: unsloth.ai/docs/models/...