Nokia's AnyJev turns any open LLM into a Jev-style decision model for free. We tested it on 500 Onion and real headlines: Jev won by 12 points, but AnyJev still had a few surprises.

Whistle is bigger than the ESP32-S3’s flash, but we trimmed it to fit, and now this $10 chip understands spoken commands in seven languages, fully offline.

An LRU cache becomes a bookcase. A rocket engine comes apart on hover. Anatomy is an open-source Claude Code skill that turns technical concepts into interactive machines. We tested it.

We gave a robot duck a 15MB brain with Needle 3 from Cactus Compute and fine-tuned it to only obey when you say please. Be rude and it gets angry.

Your WiFi can tell if anyone is home, and even whether you're sitting, walking or doing squats. We built a $30 RuView setup to test it, and it was right about 80% of the time.

Parakeet Redux shrinks NVIDIA's speech model from 1.2GB to just 178MB using 1.58-bit weights. I tested the accuracy, timestamps, multilingual audio, and how it compares to Whisper.

Claude Code can build your app. GoLive tries to handle everything after—hosting, DB, DNS, email, Stripe, verification, even teardown. I tested the full workflow.

We ran a 125B AI model on one RTX 5090 with Strata and it hit 250+ tokens per second, 3-5x faster than Qwen 3.8 Flash in the cloud. Here's how it works and how the tests went.

KittenTTS 2 can clone your voice from seconds of audio, but it's nothing like the tiny model devs originally loved. I tested it locally on a Mac and compared it with ElevenLabs.

The internet's master key changed on Sunday, and if your router missed it, your internet just stopped working.

An hour of audio transcribed in 21 seconds. On a MacBook Air. Completely offline. That's what the 164MB Phonon-2 claims. We compare it to Whisper, Redux, and NVIDIA Parakeet. There's a catch.

You can block cookies, go incognito, and use a VPN, and a website can still tell it's you: your browser has a fingerprint you can't delete.

Effect 4.0 rewrites the runtime, shrinks bundles, speeds up tasks, and makes failures part of the type system. I tested retries, timeouts, typed errors, and what actually changed.

Docker Agent lets you build AI agents in YAML with no framework code. We used it to debug an outage with two agents, and then had GPT-6 Luna and Claude Haiku 5.5 battle it out at Battleship.

Deno is joining Cloudflare, but the Deno runtime stops being developed after one more year. We break down why

🍏 Our mobile apps are now fully native on iOS and Android. → Critical alerts that cut through Do Not Disturb → Looping alerts that ring until you pick up → Live Activities with the escalation countdown on your Lock Screen → AI SRE chats that stream in real time Already have the app? Just update it from the App Store or Google Play.

“Which errors are new since yesterday’s deploy?” Your agent can answer that now. The Better Stack MCP server connects Claude Code, Cursor, or Codex to your errors, logs, and metrics, so it writes the query, runs it, and sends the chart back in the chat. Ask which pages are throwing 500s. Or just ask for a dashboard. It builds it in Better Stack and hands you the link. One line to install, OAuth in the browser: betterstack.com/docs/getting…