AI inference is obviously profitable

Richard Geldreich's Blog: Basis Universal v2.6 update

Basis Universal v2.6 will expose its real-time analytical encoders through new APIs and tooling, with a new command line tool called basisu_pack shipping in January 2027. These encoders don't require SIMD, yet many are faster than other older search-based SIMD encoder designs. The update also plans official Python wheels for several platforms, transcoding performance increases, and large compression ratio improvements for XUASTC. CLI tools ship as both native binaries and secure WASM WASI executables.

7 MoE routing strategies in production, clearly explained

+ a practical learning checklist to learn AI engineering.

Linux 7.4-rc2 Released With FAT32 Optimization, AMDGPU HDMI 2.2 Improvement

Linus Torvalds released Linux 7.4-rc2, calling it a big rc2 following the largest ever rc1. The release includes a notable FAT32 performance improvement, especially for synchronous mounts, AMDGPU driver fixes for HDMI 2.2 FRL, and numerous fixes for long-standing core memory management bugs. Torvalds noted the diffstat looks benign and is largely driven by automated coccinelle script updates that add noise, and that AI-generated fixes and bug reports continue to pour in. He said Alexei claims fixes on the bpf side are now coming in faster than new reports. Linux 7.4 is aiming for a stable release around the end of December.

How to run an LLM locally

Run an LLM locally with Ollama. See how much memory you need, where LM Studio and llama.cpp fit, and how to call the model from code.

RAG vs. Vectorless RAG, clearly explained!

...explained visually (with tradeoffs in production).

DeepSeek Super-Agent Architecture Explained

How DeepSeek Harness's Super-agent works: a Router, isolated Workers, event triggers and a Watchdog let one AI agent run many tasks. Diagrams inside.

The End of AI Frontiers

Large language models remain far less efficient than biological brains at learning, representing knowledge, and planning, the post argues. A child hears roughly one token per second, reaching a billion tokens only after 31 years, while GPT-4-class models reportedly use four orders of magnitude more data. The Phi model series, trained on 10 trillion tokens, still trails GPT-4 along most axes. A Portia spider plans prey-catching behavior with a brain a million times smaller than a human's, and squirrels recall cached food after months, suggesting that existing models may be far less representationally and memory efficient.

I changed one llama.cpp flag, and my 8GB laptop GPU now runs a 35B model faster than I can read

Một lập trình viên đã tìm ra cách chạy model Llama 35B trên laptop GPU chỉ 8GB VRAM với tốc độ đáng kinh ngạc. Nguyên nhân kỹ thuật nằm ở việc chỉnh sửa cờ n_gpu_layers trong llama.cpp, cho phép tận dụng hiệu quả bộ nhớ VRAM thay vì để nó trở thành điểm nghẽn. Hệ quả là model 35B parameters chạy nhanh đến mức người dùng không kịp theo dõi, vượt xa các kỳ vọng ban đầu về giới hạn phần cứng. Điều đáng học hỏi là đôi khi vấn đề không nằm ở cấu hình phần cứng mà ở cách tối ưu hóa phần mềm để khai thác tiềm năng thực sự của hardware.

Bài viết tiết lộ cách tối ưu hóa một cờ trong llama.cpp để chạy mô hình 35B trên GPU 8GB nhanh hơn tốc độ đọc của bạn.