GoKawiil
Tech News
A technical blog post walks through building a custom attention kernel for Nvidia's Blackwell B200 GPU entirely from scratch, using CUDA and some PTX code. Through a 14-step progression of optimizations illustrated with 60 diagrams, the kernel reaches 94.4% of FlashAttention-4's performance on 4K, 8K, and 16K sequence shapes, and is demonstrated by plugging it into a video-generation model.
Google is reportedly preparing to launch Gemini 3.8 Flash imminently, a model DeepMind's internal testing found outperforms Anthropic's Opus on coding tasks. The release would follow Gemini 3.7 Flash, launched just weeks earlier, and comes as Google shelves plans for a separate Gemini 3.5 Pro model, pushing the next Pro-tier release to Gemini 4.0.
Google has upgraded Gemini's video analysis with a new 'agentic' approach that lets the AI selectively choose which frames, audio, or transcript segments to review instead of processing every frame. The company says this cuts token usage by up to 88% while improving accuracy by up to 7%, and enables tracking of object counts, movement, and sub-second events across multi-hour videos. The feature currently works through the Gemini API in AI Studio and the Enterprise Agent Platform, with plans to reach the Gemini app and YouTube's 'Ask YouTube' tool soon.
A new open-source tool called slotstream enables Apple Silicon Macs to run the 104GB Qwen3.8-Flash-Next language model even when the machine doesn't have enough RAM to hold it, by streaming weights from disk and dynamically sizing memory use, down to an 8.1GB floor. On a 48GB Mac it delivers about 12 tokens per second in warm decode with a roughly 3-second cold start, capping memory use at 32GB so the rest of the system stays usable. It requires macOS 14+, Apple Silicon, and about 110GB of free disk space for the one-time weight download.
Smartwatches from Apple, Google and Samsung all include a basic flashlight feature that simply maxes out the display's brightness and turns it white. Despite mockery of the feature compared to a phone's dedicated camera flash, it proves genuinely useful for quick tasks like finding a dropped pill, precisely because the watch is already on your wrist rather than buried in a pocket or bag.
A major flash flood struck Bright Angel Creek in the Grand Canyon on Saturday, killing a 46-year-old man whose body was found near Crystal Rapids, and leaving about 15 people missing or unaccounted for. The surge pushed boulders into the Colorado River, prompted the airlift evacuation of more than 60 people, and destroyed the park's only pipeline supplying potable water to residents and tourists.
The vLLM project released version 0.28.0, combining 584 commits from 270 contributors. The update centers on deep performance work for Kimi-K3, including new decode context parallelism, fused kernels, memory-saving expert sharding, and ROCm support, alongside DeepSeek V4 improvements such as end-to-end sparse MLA and AMD Quark NVFP4 support.
Japanese researcher @dydt_Nao successfully demonstrated a dinner-plate-sized USB flash drive built from magnetic core memory, transferring both a text file and a small monochrome bitmap image between two PCs. The device, which uses a Raspberry Pi Pico as its controller, stores only 128 bytes of data and displays activity through a 32x32 LED array that lights up as each magnetic core is accessed. The creator plans to showcase the device at Maker Faire Tokyo 2026 in early September.
Modern flash storage relies on NAND memory cells that trap electrical charge, with newer 3D NAND designs stacking cells vertically and packing multiple bits per cell to boost capacity while lowering cost per gigabyte. These improvements have made flash ideal for SSDs and portable devices, since it lacks the moving parts of hard drives, giving lower latency and better durability. However, flash remains fundamentally different from RAM, serving as persistent storage rather than the fast temporary memory a system needs while running.
Labor unions are increasingly pushing back against data-center construction, warning politicians who support these projects that they risk losing union backing. The move signals that blue-collar workforce concerns are becoming a central point of contention in the broader public backlash against data-center expansion.
Samsung's new budget-tier 990 M.2 SSD, a QLC-based Gen 4 drive released after recent AI-driven storage price hikes, is now available at Amazon for $339 instead of its $529 list price. The 2TB model uses Samsung's PiccoloQ controller and V9 QLC flash, delivering performance close to the 990 EVO Plus with solid latency and bandwidth in testing.
Z.ai has confirmed it built the mysterious 'Ox Alpha' model that surged in popularity on OpenRouter last week, rebranding it as GLM-5.3-Flash. The model, served largely on Chinese-made chips, is open-weight, cheap, and aimed at coding and agentic tasks like browsing and controlling desktop apps, and its weights are now downloadable on Hugging Face. Z.ai's stock jumped following the reveal.