Compiling code an Raspberry Pi 5 is painfully slow

OpenAI's decision to use MXFP4 datatypes as the standard for gpt-oss is a big deal and sets the tone for the rest of the industry. Is it perfect? No, but it's a huge improvement over FP4 or INT4. Find my deep dive @theregister .

That's my arm if you're wondering. Not sure why I felt the need to share that, but now you know.

Main thing confusing me: why give up an optimized training stack built around Nvidia to take a chance on Huawei? Like other than national pride. The 910C doesn't support FP8, so back to BF16. Were they using them for RL?. Latest for @TheRegister .

Gentle reminder that GPUs aren't just for AI. In the right hands they can advance life saving science. Boffins at LLNL used the No. 1 ranked supercomputer, El Capitan, to build a real-time tsunami forecast system. My latest for @theregister .

Need one of these for review. I want to type -tp 8 into vLLM.

Got GPT-OSS 120B running tonight in Llama.cpp. ~60Tok/s with minimal expert offload to DRAM. Had to compile a pull request to make it play nice to Open WebUI, but I have tool calling now. Need to figure out a system prompt cause my god this thing loves tables.

RT @TheRegister: s nuclear strategy is 'slow, inefficient, and costly'

Meanwhile, if you haven't figured it out already GPT-5 is a huge improvement, just mostly for OpenAI's bottom line. This thing is a big cost cutting measure and not necessarily in a bad way:. My latest for @TheRegister .

Still need to do some cable management and it remains to be seen whether that PSU will be adequate, but it posted. 24 Zen4 cores.128GB DDR5 5600 RDIMM.2x RTX 3090 TI 24GB w/ NVL. And don’t worry those GPUs are power limited to 250W. The PSU shouldn’t explode.

Well of course the Chinese don't want H20s anywhere near their sensitive government workloads. Why give Uncle Sam any more reasons to think kill switches are a good idea. My latest for. @theregister.

Challenge: go a week with gpt-oss (either 20B or 120B) as your only AI assistant. Could you do it?. I have a feeling it's gonna be painful, but I'm going to try.

I did a thing…

When Llama 4.1?.

Okay that's a wrap on this thread. Almost time to tune into the GPT-5 launch preso. [Fin].

But while gpt-oss is a big step up for US open weights models, the best models in this category are still Chinese. Qwen3-Coder and Qwen3-A235B-A22B-Instruct is very, very good for its size, and its just one of many open weights models out of China to choose from. [5].

Now the ball is in Meta's court. Fingers crossed Llama 4.1 and 4.2 are more impressive than what we got back in April. It'd be great to see Google, Microsoft, and IBM step up their game with more capable Gemma, Phi, and Granite models too. [4].

But if you need an American made model for regulatory and compliance reasons, gpt-oss is hard to beat. Plus, I'm sure application-specific fine tunes will be more reliable. [3].

They are not perfect. The gpt-oss-20B halucinates a bit too much and its refusal to admit when its wrong is problematic. [2].

After having a few days to mess with OpenAI's gpt-oss — specifically the 21B parameter version — I gotta admit, these are the most impressive open models to come out of the US so far this year. A 🧵 [1]