Jürgen Schuck

#ordinaryguy #culture #science #technology #laughing #hiking #motorbiking

[Jürgen Schuck @otabuzzman]

@gerogerber me too, provided there's a cool protective case for it… 🥳

[Jürgen Schuck @otabuzzman]

@gerogerber @DB_Bahn sounds like a bad ux… 😬

[Jürgen Schuck @otabuzzman]

@violaodorata2 @tcboyle einmal manta, immer manta… 😉

[Jürgen Schuck @otabuzzman]

@tcboyle they know their *belle cuisine* business there, too.

[Jürgen Schuck @otabuzzman]

@tcboyle currywurst, french fries, topped with ketchup and mayonnaise. in the „ruhrgebiet“ they call it „mantateller“.

[Jürgen Schuck @otabuzzman]

@tcboyle currywurst, french fries, topped with ketchup and mayonnaise. in the „ruhrgebiet“ they call it „mantateller“.

[Jürgen Schuck @otabuzzman]

@tcboyle the car that coined a dish…

[Andrej Karpathy @karpathy]

nanochat now trains GPT-2 capability model in just 2 hours on a single 8XH100 node (down from ~3 hours 1 month ago). Getting a lot closer to ~interactive! A bunch of tuning and features (fp8) went in but the biggest difference was a switch of the dataset from FineWeb-edu to NVIDIA ClimbMix (nice work NVIDIA!). I had tried Olmo, FineWeb, DCLM which all led to regressions, ClimbMix worked really well out of the box (to the point that I am slightly suspicious about about goodharting, though reading the paper it seems ~ok).

In other news, after trying a few approaches for how to set things up, I now have AI Agents iterating on nanochat automatically, so I'll just leave this running for a while, go relax a bit and enjoy the feeling of post-agi :). Visualized here as an example: 110 changes made over the last ~12 hours, bringing the validation loss so far from 0.862415 down to 0.858039 for a d12 model, at no cost to wall clock time. The agent works on a feature branch, tries out ideas, merges them when they work and iterates. Amusingly, over the last ~2 weeks I almost feel like I've iterated more on the "meta-setup" where I optimize and tune the agent flows even more than the nanochat repo directly.

[Lize Raes @LizeRaes]

Could this loop execute in parallel, and run orders of magnitude faster? Project Babylon (experimental) allows running Java code directly on the GPU (yes, even the one in your laptop). See how in our @Jfokus talk with @ammbra1508 https://t.co/sZI74ByO9K

[Jürgen Schuck @otabuzzman]

@mac_and_i 4/ 2025 präsentiert Theorie und Praxis von Stable Diffusion. Probier’s mal aus, und bei Fragen: fragen… 😀

https://t.co/dJ14h81MlH

[TornadoVM @tornadovm]

New TornadoVM release v1.1.1!

Packed with new features and optimizations that unlock GPU acceleration for Llama3 in pure Java!

🧬JIT-compiled kernels for high performance.
👏 Thanks to our amazing contributors.

👉 https://t.co/FiQtx0m5SA

#opensource #AI #LLM #Java #TornadoVM

[Jürgen Schuck @otabuzzman]

Here is the transcript of the PDF made with AWS Textract, awk and manually. Details are in the repo. A reasonable next step could be a port to NASM. Anyone in the mood?
#altairbasic #microsoft #retrocomputing #pdp10

https://t.co/ayWQzJVf30

https://t.co/VsolXaAqFX

[Jürgen Schuck @otabuzzman]

In der @mac_and_i 3/ 2025 bringe ich in der DevCorner den kleinen GPT-2 von OpenAI mit Metal 🤘 auf Trab. Nur den Forward-Pass. Wer hat lust auf den Backward-Pass? 😎
#gpu #metal #apple #macandi #heise #parallelcomputing #macos #ios #ai

https://t.co/8XDQZVJ8Oc

[Jürgen Schuck @otabuzzman]

@geeksforgeeks 💯 digital, books included.

[AI at Meta @AIatMeta]

Today is the start of a new era of natively multimodal AI innovation.

Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality.

Llama 4 Scout
• 17B-active-parameter model with 16 experts.
• Industry-leading context window of 10M tokens.
• Outperforms Gemma 3, Gemini 2.0 Flash-Lite and Mistral 3.1 across a broad range of widely accepted benchmarks.

Llama 4 Maverick
• 17B-active-parameter model with 128 experts.
• Best-in-class image grounding with the ability to align user prompts with relevant visual concepts and anchor model responses to regions in the image.
• Outperforms GPT-4o and Gemini 2.0 Flash across a broad range of widely accepted benchmarks.
• Achieves comparable results to DeepSeek v3 on reasoning and coding — at half the active parameters.
• Unparalleled performance-to-cost ratio with a chat version scoring ELO of 1417 on LMArena.

These models are our best yet thanks to distillation from Llama 4 Behemoth, our most powerful model yet. Llama 4 Behemoth is still in training and is currently seeing results that outperform GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused benchmarks. We’re excited to share more details about it even while it’s still in flight.

Read more about the first Llama 4 models, including training and benchmarks ➡️ https://t.co/9G3QgVdCkB
Download Llama 4 ➡️ https://t.co/eVomRvEr0w

[TornadoVM @tornadovm]

New release #TornadoVM v1.1.0 is out!

Key highlights:
✅ Mixed precision computations (FP16 to FP32)
✅ New memory and buffer management features for hardware accelerators (persist & reuse data)

🔗https://t.co/kfWQMdx2Ot

Thanks to all contributors!

#opensource #Java #AI

[Jürgen Schuck @otabuzzman]

Der zweite Teil meines Tutorials in #iX 4/ 25 bringt #LLaMA mithilfe von #ExecuTorch auf‘s iPhone, und eine Mini-App für #Xcode in #Swift zeigt, wie sich auf dem Fon mit dem #LLM chatten lässt.

https://t.co/4BVALoy7Zj

[Jürgen Schuck @otabuzzman]

Das neue @iXmagazin 2/ 2025 bringt Teil 1 meines Tutorials zu ExecuTorch. Das Framework und API macht große LLMs passend für Smartphones und Edge-Devices. Vielen Dank an die Redaktion. #executorch #ix #heise #LLM
https://t.co/gPIiEG0MGa

[Jürgen Schuck @otabuzzman]

Is there anybody out there in the mood for porting #CUDA kernels to #Metal compute shaders in #llm.c? Link to repo included in blog post. PRs welcome. 🙂Any other feedback as well. #Swift #LLM #GPT2 #parallelcomputing #macos #ios

https://t.co/705GC6xFIw

[Jürgen Schuck @otabuzzman]

some good thoughts on the subject (imo) https://t.co/p2WP9OjWCc