Eric Hartford

Average human consumes on the order of hundreds of kilojoules per paragraph of information they create, while AI inference uses on the order of hundreds of joules per paragraph, making AI roughly three orders of magnitude more energy-efficient per paragraph. This disparity will only increase (on both ends) - and that has macroeconomic ramifications.

@UnslothAI @NVIDIAAI @googlegemma 1.5x faster than what?

We’re releasing Gemma 4 NVFP4 quants that run 1.5× faster on your GPU. Gemma-4-12B NVFP4 works on 11GB VRAM. 26B-A4B hits 13K tok/s (B200). Unsloth NVFP4 enables faster, more accurate 4-bit Blackwell inference. Blog: unsloth.ai/docs/basics/nv… Gemma NVFP4: huggingface.co/collections/un…

We’ve just released the 1-bit & 4-bit version of Hy3, a flagship-scale 295B model that can be served on a single GPU. 👌 Run Hy3 with llama.cpp, enable MTP, and experience powerful intelligence on dramatically lower hardware.🚀🚀🚀 Can’t wait to see what you build. #Hy3 #Hy #GGUF #llamacpp

@LeeLeepenkman Yeah I noticed it ignores

accidentally pasted slop into claude code and it paused briefly and ignored me and kept going.... cool gst is my git status alias so muscle memory so sometimes i accidentally type it

Hy3 in 1bit! Benchmarks are still great!

@_LazarusAI SovereignStack Own your AI. The frontier labs will raise prices, deplatform you, deprecate the model you depend on. You can't run it on prem. You can't run it air gapped. You can't finetune it or run community finetunes. SovereignStack enables all that. It's free and open source.

@codemeoww @jun_song He's gonna Grok you in the mouth!

@jun_song Got grok'ed by Elon 🤣 That's why I don't get hasty with model choice. It's just a trap

How do I get refund for this?

@jun_song "Nvidia NVFP4" Runs great on Intel XPU GPU's, with QuixiCore kernels! github.com/QuixiAI/QuixiC… (and AMD too)

SuperGLM-5.2-abliterated is finally available 🚀 The strongest local AI that you can run on your hardware. - uncensored from Nvidia NVFP4 - Fixed broken points from abliteration - enhanced overall performance Only available with nvfp4 now ⬇️

When they design the flavor of spicy chips / nuts, why do they always mix in "fruit loops" flavor?

@0xSero whatcha think about this brianbell-x.github.io/weight-compres…

Nvidia/GLM-5.2-NVFP4 running on 4x rtx pro 6000 with 400k context @ 82 tok/s and is doing amazing I have to water cool these GPUs, the temperatures and sounds are limiting how far I can push this thing This new moet compression is going straight in local studio for onboarding

@sudoingX Gaudi 2 server

what's the loudest computer you've ever owned.

Anthropic is a clown show

Here's a blog post explaining all the neat tricks I discovered while writing this kernel. Hope you'll enjoy reading it as much as I did writing it: blog.alpindale.net/posts/swordfis…

Introducing the Swordfish Kernel. I've been studying the Blackwell sm100 architecture, and this is my contribution as a result of that. It's a weight-only inference kernel, a successor to Marlin and Machete, designed to maximize throughput on the datacenter-grade Blackwell GPUs. On a B200 for INT4 inline dequantization, it reaches 90% of cuBLAS on BF16, and on the Jetson Thor, it reaches ~95% of the memory bandwidth.

💀

Fable is very helpful with molecules - as long as they dont have carbon chains because then they are organic, and that's DANGEROUS.

@ivanfioravanti Fable is still better - when it decides it wants to

I've not used Fable at all in the last 3 days, and I don't feel the need honestly. All new models are more than enough. I'm really curious to see if Anthropic will remove it from subscriptions or not.

@waterloo_intern @dylan522p To gather data

@Shra_va_ni You write it then It takes skill to manage agents effectively

can you call yourself a developer if AI wrote most of your code?

@TeksEdge @Alibaba_Qwen is backing off from open source

We've had Qwen3.6-27B and 35B since April and Qwen3.7 Pro/Max since May. Wonder if Qwen3.7 is coming for medium and small.

Fable refused and downgrade itself to 4.8 Opus because I asked it for help with debugging my compiler (the With Programming Language)