Lead on PCB design software TraceMaker (https://xitter.cf/t.co/QXXc9mLbl9), and LLM visualisation llm-visuals (https://xitter.cf/t.co/nhYsCUxUdh).
Australia
Joined July 2026
- Tweets2.2K
- Following2.5K
- Followers1.9K
- Likes13.2K
Nigel Hungerford-Symes retweeted
Turns out there's a livestream straight from the JLCPCB factory where you can watch PCBs being manufactured in real time:
Nigel Hungerford-Symes retweeted
New version is now live with another performance boosts and improvements on Strata
github.com/Niko1221/Strata/r…
Nigel Hungerford-Symes retweeted
Prompt t/s: 2,170 -> 2,653 (+22%)
Output t/s: 64.5 -> 76.4 (+18%)
Enjoy new performance boost! 👀
github.com/Niko1221/Strata/r…
Nigel Hungerford-Symes retweeted
Qwen3.8-Flash-Next first 32K tokens are 4%-10% faster, LOW RAM mode added (for low hardware).
New update here
github.com/Niko1221/Strata/r…
Nigel Hungerford-Symes retweeted
Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance.
Grok 4.8 will be a noticeable improvement.
Grok 4.9 is probably Astra/Fable class.
Grok 5 maybe better than anything. We shall see.
Nigel Hungerford-Symes retweeted
WARNING: Omarchy dot net is not legitimate. It looks like a scam/malware/phishing operation. I alerted their hosting service @spaceship, which acknowledged receipt, but refused to act. Wild that they want to be part of that, but ok!
Nigel Hungerford-Symes retweeted
If the “watermark” is embedded in the statistical sequencing of text ie it can’t be stripped out without rewriting the text - doesn’t that necessarily make LLM text generation less useful?
The model is no longer performing purely for the user’s request. Some fraction of its freedom in choosing words, phrasing, and structure is now being spent on preserving the watermark.
You get an inherent tradeoff between watermark robustness and model flexibility/compliance. Ergo, the stronger & more robust the watermark, the greater the constraint.
In other words: LLMs will suck more and be harder to work with.
Nigel Hungerford-Symes retweeted
Minimax H3, same 4x3090 rig, same clip. Yesterday it rendered in 11:21. Today it renders in 3:45 🤯
The hardware did not change. Two things happened overnight. An AI agent rewrote the attention CUDA kernel and hit a half speed trap inside GeForce tensor cores that most people never heard of. Then a stranger on HuggingFace dropped a LoRA that cuts 20 sampling steps down to 4.
One of these two mattered way more than the other. Breakdown below.
I packed all of it into ready ComfyUI workflows for my rig. If you want them, ask in the replies and I'll share.
Just sent Prime Agent with Kimi K3 to optimize the SageAttention Kernel for more efficient Minimax H3 inference on RTX 3090.
The issue is that the handwritten Cuda Kernel for Ampere loses to the Triton version from the same library. on my shapes it gets 85 ms versus 74 ms with identical accuracy. Triton autogenerator beats human code.
Looks like a tile sizing problem. the 128/64/32/64 dimensions are hardcoded constants tuned for A100. 3090 has 100 kb shared memory instead of 164 kb so the resource balance is completely different.
lets see what it finds.
Nigel Hungerford-Symes retweeted
All that needs to be said about pit bulls was said by S-tier dog trainer Stonnie Dennis in 3 minutes
Nigel Hungerford-Symes retweeted
🚨 Elon Musk says SpaceX will attempt its first Starship upper stage tower catch on Flight 14.
Nigel Hungerford-Symes retweeted
📢Meet Qwen3.8-Max — our most capable model to date.
Next week, the open weights of Qwen3.8-Max will be released, and Qwen3.8-27B is also going open-weights to meet you all!🎉
Qwen3.8-Max, a new bar for coding and cowork at 2.4T parameters:
- Autonomous coding: 10+ days of self-evolving development, from empty folder to production without hand-holding, complete project trace in the GitHub：github.com/qwen-code-dev-bot…
- Real work, real results: Production-quality deliverables across hundreds of professions.
- Long-horizon mastery: System-level autonomous planning with closed-loop adaptive learning, driving 500+ turns of chip design optimization and 365 days of e-commerce strategy.
- Native multimodal intelligence: Vision isn't just input — it's a continuous feedback loop for planning, execution, and self-correction.
💰Pricing:
Input: $2.0 / M tokens
Output: $6.0 / M tokens
Implicit Caching: $0.25 / M tokens
Start building with Qwen3.8-Max! 🚀
📖 Blog: qwen.ai/blog?id=qwen3.8
✅ Qwen Studio: chat.qwen.ai/?models=qwen3.8…
⚡ API: qwencloud.com/models/qwen3.8…
