Singing Data

This guys deserves all the follows. He’s a smart version of me. He got Qwen3.5-27B running in the browser. Stuff is crazy

Leaked video from inside ICE detention shows women trapped in desperate conditions—smuggled out by husband. Used panties, moldy food, broken shoes, clothes that don't fit because made for men—and a toxic mattress made of insulation. Gabriela Sousa came to U.S. legally granted humanitarian parole from Venezuela—and is married to U.S. citizen husband. Her husband helped the women smuggle out this video they made—with testimonials from several women detained together in these inhumane conditions The video was made secretly inside the Baker County ICE Detention Center in Macclenny, Florida.

“You will be home before the leaves have fallen from the trees.” Kaiser Wilhelm August 1914

@groversmith1 I mean, he has a point. I vote dem AND i think one of bidens squandered opportunities was not talking about infrastructure bill projects and wins

Sad how retards get so much attention on this site

I ran both OpenClaw and Hermes on the same mac mini. this breakdown is accurate and worth reading if you're building with either. the core difference in one sentence: OpenClaw stores everything and searches it. Hermes keeps almost nothing in the prompt and retrieves the rest on demand. to be more specific, Hermes has 4 memory layers: → MEMORY.md + USER.md: ~1,300 tokens of curated facts. that's it. deliberately tiny so the prompt stays stable and cache-friendly → session_search: full SQLite archive of past conversations with FTS5 search. only queried when needed, never injected by default → skills: procedural memory. when the agent solves something hard, it saves how it did it, not what happened, but how to do it again → honcho (optional): cross-session user modeling that attaches to the current turn without mutating the cached prompt Hermes would rather search for a fact at the cost of one tool call than stuff it into every single message and break the cache. I've seen this tradeoff firsthand. my OpenClaw bot was 2 months of accumulated memory, deep context, but every message replays more history. hermes is 2 weeks old with a fraction of the memory, but it responds in 5 seconds vs 60 because it's not dragging its entire life story into every turn. in @NousResearch we trust.

Hugging Face is more than a model catalog. In my new video, I break it down into: 1) Models + Inference Providers 2) Datasets + Data Studio 3) Spaces + Arena demos If you are deciding where to start in the HF ecosystem, this gives you a clear map.

If you're attending GTC, whether virtually or in-person, check out Chris' Agentic AI workshop I am sure many of you would find it interesting and learn something

context if you are new here: Octopus Invaders is a standardized coding test i run on every model and GPU i benchmark. pixel art enemies, particle effects, procedural audio, boss battles, upgrade system. one prompt. full game. Qwen 3.5 35B on a 3090 built it in one pass and it ran. 3,483 lines. Qwen 3.5 9B on a 3060 built 2,699 lines and the screen was black. until tonight. same model family. fraction of the parameters. fraction of the GPU. different prompting strategy. working game. if you want to run this yourself: model: Qwen3.5-9B-Q4_K_M.gguf (5.3GB download) agent: Hermes Agent flags: -ngl 99 -c 131072 -np 1 -fa on --cache-type-k q4_0 --cache-type-v q4_0 fits on any 8GB+ card. 128K context. full GPU offload. flash attention on. bookmark this.

hey if you have a 3060, or any GPU with 8GB or more sitting in a drawer right now, that thing can run 9 billion parameters of intelligence autonomously. and you don't know it yet. 2 hours ago i posted that 9B hit a ceiling. 2,699 lines across 11 files. blank screen. said the limit for autonomous multifile coding on 9 billion parameters is real. then i audited every file. found 11 bugs. exact file, exact line, exact fix. duplicate variable declarations killing the script loader. a canvas reference never connected to the DOM. enemies with no movement logic. particle systems called on the class instead of the instance. fed that list as a single prompt to the same Qwen 3.5 9B on the same RTX 3060 through Hermes Agent. it fixed all 11. surgically. patch level edits across 4 files. no rewrites. no hallucinated changes. game boots. enemies spawn, move, collide. background renders. particles fire. and here's what nobody is talking about. this is a 9 billion parameter model running a full agentic framework. Hermes Agent with 31 tools. file operations, terminal, browser, code execution. not a single tool call failed. the agent chain never broke. most people think you need 70B+ for reliable tool use. this is 9B on 12 gigs doing it clean. the model didn't fail. my prompting strategy did. the ceiling is not the parameter count. the ceiling is how you prompt it. this is not done. bullets don't fire yet. boss fights need wiring. but the screen that was black 2 hours ago now has a full game rendering in real time. iterating right now. anyone with a GPU from the last 5 years should be paying attention to what is happening right now.

I change my mind. I would have bet against @TheAhmadOsman 's Jan'06 prediction that an Opus 4.5 tier model would run on a single RTX 6000 Pro by end of 2026. "Because ultimately the number of parameters matters, and there's no way to escape the physics of vram," I thought. But then Qwen3.5 27B dropped late Feb, which entirely broke the intelligence density barrier. 42 on AA, can run on 3090 with full context, albeit slowly, which now fights Minimax m2.5 which would have required $20k 2x RTX 6K Pros on Feb 12, and Deepseek v3.2 $80k 8x RTX6K in Dec 2025. Now I have little doubt that Minimax 3.3 or Qwen4 or Deepseek v4.5 or whatever will be at or very near Opus 4.5 by Dec 2026. But, also, it's hard to imagine how far ahead Opus 5.6/Gpt-6/etc will be by then. x.com/i/status/20122…

@beffjezos We have full pure Python user space drivers for AMD and NVIDIA in tinygrad. USB4 devices can be mmaped like they are directly on the PCIe bus. This isn't hype, it all works today on any 3000-5000 series NVIDIA or RDNA3/RDNA4 AMD.

This will probably be great for Large single GPUs (e.g. RTX PRO 6000) You’re limited to 40Gpbs initially (during model loading) but then once the model is fully loaded on the GPU it should be extremely faster than Unified Memory speeds for inference

Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI youtu.be/kwSVtQ7dziU?si… via @YouTube

problem: the united states has no near term plan to open the strait

So many of these, but this one is a classic Chuck Norris joke. 🐅 #RIPChuckNorris

Leon Wyczółkowski - Spring, the interior of the artist's studio (1933)

It's Past Time to Privatize U.S. Air Traffic Control and TSA—Copy Canada’s Model marginalrevolution.com/marginalrevolu…

Assuming the USS Boxer is heading to the Middle East, and considering that she only left San Diego two days ago, she won't be near the Persian Gulf until mid-April. If the reinforcements are intended to reopen the Strait of Hormuz, then the oil market faces a very long wait.

Chuck Norris went out in the ultimate Chuck Norris way - working out in Hawaii, posting about enjoying life at 86, then boom - checks out quickly. Reminds me of the classic: "Chuck Norris has two speeds, walk and kill."

@real_bill_gross Hoping for fresh prophecies from the oracle...

The elite 8 (9/10?) are being cut quickly — ORCL, MSFT, IBM. Maybe it’s not a single elimination game but it feels like it. I’m staying away. Be satisfied with single digit returns — WES, VZ, T, VEU.