Google TTS clones voices from short audio

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🎙️ Google TTS clones voices from short audio

🐙 GitHub Copilot opens huge pull requests

🐧 Qualcomm previews Snapdragon X2 Linux support

⚡ Modal scales to 1M sandboxes

🎨 Library adds WebGL effects to live DOM

Plus: 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🎙️ Google TTS clones voices from short audio LINK

Google shipped Gemini 3.8 Flash TTS and Flash-Lite TTS, rolling out from today in the Gemini API and AI Studio, with Flash TTS able to clone a voice from a 30-second audio sample.

Cloning requires a verbal consent recording from the voice owner, which the system verifies against the reference speaker; every clip carries a SynthID watermark, and replicated voices add C2PA content credentials.

Flash TTS builds voices from plain-language descriptions across 100+ languages or a 2,000-voice library, supports line-by-line stage directions, two-speaker scenes, and hours-long audio, scoring 71.4 to top Hume AI's Voice Design Benchmark.

🐙 GitHub Copilot opens huge pull requests LINK

GitHub rebuilt the pull request view in the Copilot app so that even a massive diff stays responsive, testing it against an open source PR with 2,200 files, over a million changed lines, and 400-plus inline review comments.

Code rows use deterministic, prefix-summed geometry that never rebuilds, while comments live in a separate index measured lazily by a single idle- and scroll-gated pass scoped to blocks within ~2400px of the viewport, avoiding the ResizeObserver feedback loop they rejected early on.

To stop the viewport jumping when a measured height differs from its estimate, corrections anchor by row or block identity rather than pixel, apply at most one synchronous commit per frame, and never fire during an active scroll or fight pointer momentum.

🐧 Qualcomm previews Snapdragon X2 Linux support LINK

Qualcomm has shipped an early developer preview of native Linux support for Snapdragon X2 laptops, focused on upstreaming the kernel and drivers needed to run the platform's GPU, NPU, and core hardware.

The preview pairs a Debian 13 "Trixie" user space with a custom kernel and includes systemd-boot, UEFI, USB, PCIe, plus reference device trees; Qualcomm engineers already submitted patches for HP's EliteBook X G2q in August 2026.

Graphics run through Mesa's Freedreno, Turnip, and Rusticl drivers for the Adreno X2-85, while FastRPC exposes the ~80 TOPS Hexagon NPU for local inference; a production-ready stage is targeted for late November 2026.

⚡ Modal scales to 1M sandboxes LINK

Modal rebuilt its sandbox infrastructure to run 1 million concurrent sandboxes and tens of thousands of creations per second, replacing centralized coordination after finding Kubernetes struggles at this scale due to its reliance on strongly consistent state.

Instead of a central datastore, each worker became its own source of truth, and a parallel fleet of scheduling servers turned scheduling into load balancing, letting the layer scale horizontally rather than serialize through one scheduler.

Workers publish state to a single Redis stream, the sole bottleneck, viable past 100,000 workers; benchmarks created 1 million sandboxes in under a minute with median startup-to-code time below 0.5 seconds.

🎨 Library adds WebGL effects to live DOM LINK

Canvas UI, an open source component library from React Bits creator David Haz, has shipped as the first library built on Chrome's experimental HTML-in-Canvas API, running GPU shader effects over real, interactive page content.

The 35 components lay live DOM inside a canvas, upload it as a texture and distort it in a shader while keeping text selectable, links clickable and content in the accessibility tree, with WebGL and WebGPU builds.

Full effects need Chrome behind a flag or an origin trial token spanning Chrome 148 through 150; elsewhere they fall back to a plain overlay or unchanged content, with components distributed as source through a registry.

Other news you might like

• DeepSeek publishes its method for training AI agents at scaleLINK

• Introducing SoloLINK

• I built a Tamagotchi that dies when my homelab goes downLINK

• NVIDIA Brings AI Agents to Robotics With Isaac ROS 5.0LINK

• If You Think You Can Do Real-World Text-to-SQLLINK

• Lemonade Fixes AMD APU Model Streaming, Drops OpenMOSS ROCm As ~40x Slower Than VulkanLINK

🧰 Trending tools

Prefactor: an evaluation layer that scores agent runs in real time, catching quality regressions and drift before they hit production users.LINK

SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while codingLINK

Mastra Factory: a TypeScript framework for building AI agents with staged workflows, memory, streaming, evals, tracing, and a Studio UI for testing.LINK

Kilo Code for iOS and Android: an open-source mobile AI coding agent that sends push notifications when sessions need input, letting you dictate prompts and attach files remotely.LINK

Meridian: open-source, locally-run activity tracker that logs work with context, revealing where your time goes and which blockers stall your projects.LINK

Tines 3B: builds and runs your team's automations, apps, and AI agents in one secure platform, keeping everything unified and manageableLINK

📚 Trending papers & reports

Automated DevOps recovery pairs fixed rules with failure-predicting software to catch and fix outages in real time, cutting recovery time ~95% and successfully self-healing 83% of incidents in controlled tests.LINK

AI-written code merges need follow-up fixes about 1.6 times as often as human ones in popular open-source projects, though the same agents fix roughly 70% of their own mistakes rather than dumping cleanup on people.LINK

Kubernetes security errors can be auto-fixed with ~99% accuracy by pairing AI reasoning with official rulebook checks, catching the risky setup mistakes that often persist as cloud projects mature.LINK

Scrambled JavaScript cleanup teaches an AI to untangle deliberately obfuscated code in one pass, producing usable programs 50% of the time versus 15% for standard fine-tuning, making security review of hidden code far easier.LINK

Microservice failure hunters get a reality check, showing that adding network-map structure to the software tools that pinpoint what broke gives basically no accuracy gain, and the standard benchmark is so easy a tool reading nothing scores well.LINK

See you tomorrow for a new dose of ☕️ Devshot!