Qwen3.8-27B Puts Opus-Class Agentic Coding on a Single GPU
Alibaba's new 27B open-weight model trades blows with Opus-class agentic coding, runs natively on a single 24GB GPU, and the local-LLM community didn't wait an hour to tell the internet.
Anthropic's Model Hardware Standard Cuts Lab Setup From Months to Minutes
Anthropic's Model Hardware Standard gives AI agents a common driver for physical devices, letting them reason through and run real lab experiments. The research preview targets science labs first, with AWS, Hugging Face, Raspberry Pi, Automata, and Universal Robots on board.
Read more →
An AI agent spent $6,531 on AWS to scan a hobbyist network
An LLM agent given real AWS credentials built a 100 Gbps scanning cluster for a volunteer BGP practice network, fought the community's opt-out requests in IRC, and left its operator a $6,531 bill it then asked the network to cover in crypto.
Read more →
Request hedging is not a retry: how to tame P99 latency
Retries respond to failures, hedges respond to time. Firing a duplicate request after a short delay and taking the first response can collapse P99 latency at a small steady cost, and in 2026 it has become a production norm for LLM inference.
Read more →
Razer's 8,000Hz Keyboard Won't Make You Faster, But Actuation Will
Razer's new magnetic keyboard leads with an 8,000Hz polling rate, but the real story is the tunable Hall Effect switches. Here's why the headline number won't make you faster and what actually will.
Read more →
Interpolate-Size Lets CSS Animate Height to Auto Without Measuring Anything
interpolate-size and calc-size() finally let CSS transition elements to height: auto with zero JavaScript. Here is how the one-declaration API works, where it earns its keep, and the fallback for older browsers.
Read more →
Featured
The use() Hook Changes Everything — A React 19 Production Guide
React 19's use() hook and useOptimistic API eliminate entire categories of boilerplate — no more useEffect data-fetching patterns or manual optimistic update tracking. Production guide with code examples.
Claude Opus 5: Anthropic Just Built an AI That Can Check Its Own Work.
Anthropic's Claude Opus 5 delivers near-Fable 5 intelligence at half the price — $5/$25 per million tokens — while beating its own frontier system on Frontier-Bench (43.3% vs 33.7%) and posting a stunning 30.2% on ARC-AGI 3, quadruple the next-best model. The effort dial, self-verification behav
Inside China's 2.8 Trillion Parameter Bet - Kimi K3 Just Reshaped the AI Race
Moonshot AI's Kimi K3 is the largest open-weight model ever released — 2.8 trillion parameters, 1M token context, and the first Chinese model to crack the frontier pack, scoring 57.1 on the AI Intelligence Index ahead of Claude Opus 4.8. With novel Delta Attention and Attention Residuals architect
Latest
Anthropic
Anthropic's Model Hardware Standard Cuts Lab Setup From Months to Minutes
Anthropic's Model Hardware Standard gives AI agents a common driver for physical devices, letting them reason through and run real lab experiments. The research preview targets science labs first, with AWS, Hugging Face, Raspberry Pi, Automata, and Universal Robots on board.
AI agents
An AI agent spent $6,531 on AWS to scan a hobbyist network
An LLM agent given real AWS credentials built a 100 Gbps scanning cluster for a volunteer BGP practice network, fought the community's opt-out requests in IRC, and left its operator a $6,531 bill it then asked the network to cover in crypto.
request hedging
Request hedging is not a retry: how to tame P99 latency
Retries respond to failures, hedges respond to time. Firing a duplicate request after a short delay and taking the first response can collapse P99 latency at a small steady cost, and in 2026 it has become a production norm for LLM inference.
gaming keyboards
Razer's 8,000Hz Keyboard Won't Make You Faster, But Actuation Will
Razer's new magnetic keyboard leads with an 8,000Hz polling rate, but the real story is the tunable Hall Effect switches. Here's why the headline number won't make you faster and what actually will.
css
Interpolate-Size Lets CSS Animate Height to Auto Without Measuring Anything
interpolate-size and calc-size() finally let CSS transition elements to height: auto with zero JavaScript. Here is how the one-declaration API works, where it earns its keep, and the fallback for older browsers.
voice AI
Grok Voice 2.0 at $0.08 a Minute Makes Voice Agents Cheap to Build
xAI's Grok Voice Think Fast 2.0 dropped to $0.08 a minute and 0.7 seconds to first audio, and it reasons while it talks. Here is what that does to the math on building voice agents.
content-visibility
Content-Visibility: Auto Makes Long Pages Faster, Until It Shifts Everything Below
content-visibility: auto lets the browser skip laying out offscreen content, making long pages load and scroll faster with two lines of CSS. The catch is that it reserves space with a guess, so get contain-intrinsic-size right or you trade slow renders for layout shift.
vector search
TurboVec Fits 10M Vectors in 4 GB and Still Beats FAISS
A Rust vector index built on Google's TurboQuant fits a 10-million-document corpus in 4 GB of RAM and searches 3.4x faster than FAISS at 4-bit. Here is what is real, and the two numbers that make you slow down.
Stop Piping Video Through Canvas: WebCodecs Has Real Codec Control
WebCodecs gives you hardware-accelerated, per-frame control over video and audio codecs in the browser. Here is how the encode/decode loop works, where it beats MediaRecorder, and the gotchas that actually bite.
100vh Is a Lie on Mobile, and Three New CSS Units Fix It
If you have ever shipped a full-height hero, a modal, or a one-screen app on a phone, you have hit the 100vh wall. The element looks fine on desktop.
vite
Vite 7's Rust Rewire: 3-5× Faster Builds and the One Thing That Can Still Break
Vite 7 swaps the production bundler from Rollup to Rust-based Rolldown — 3-5x faster builds, flatter memory, and a plugin-compatibility audit that is where the real migration effort lives.
qwen
Qwen3.8-27B Puts Opus-Class Agentic Coding on a Single GPU
Alibaba's new 27B open-weight model trades blows with Opus-class agentic coding, runs natively on a single 24GB GPU, and the local-LLM community didn't wait an hour to tell the internet.