Hi there, this is your daily ☕️ Devshot.
In today's Devshot:
⚡ Google ditches Python to turbocharge AI search
🦀 GitHub ported Copilot to Rust using AI
📱 Expo's Exact runs one app on four platforms
🖥️ Netlify launches long-running Node.js servers
🅰️ Angular's Rust compiler cuts build time
🌐 New guide teaches coding agents web basics
Plus: 🎁 5 more stories you might like, 🛠️ 6 tutorials, 🧰 6 tools, and 📚 5 papers.
Google ditches Python to turbocharge AI search LINK
- Google open sourced XSE, the XLA Serving Engine, which strips the Python interpreter out of ML serving by compiling models into native C++ code for TPUs, GPUs, and CPUs.
- Models stay written in Python with JAX, then an export tool packages their computations as StableHLO and trained weights into an "XLA SavedModel" that XLA compiles to native code at serving time, with planned TorchTPU support for PyTorch.
- On short requests, XSE hit two to five times the throughput of Google's best optimized Python server on a quarter of the memory, and already serves over six million queries per second across YouTube Shorts, Ads, and Search.
GitHub ported Copilot to Rust using AI LINK
- GitHub rewrote the shared runtime behind several Copilot products from TypeScript to Rust, moving roughly 430K lines of production code, and did the entire port using Copilot, Copilot CLI, and AI agents.
- Instead of a big-bang rewrite, the team ported one component per pull request behind thin cross-language wrappers, while fleets of coordinating and child agents split large files, like a 30K-line session module carved into 15 parallel sessions.
- The old design forced every client to launch a separate Node process carrying ~100MB of overhead, which hurt apps needing tighter speed, memory, and scale; a single developer finished the Rust port in a few months rather than the years a full team would have needed.
Expo's Exact runs one app on four platforms LINK
- Expo released Exact, a cross-platform application runtime that lets you write an app once and run it natively on web, macOS, iOS, and Linux, with AI agents able to build, run, and test it on each.
- You describe the UI in Contract, a declarative language, and write logic in TypeScript or Rust; the compiler bakes a compact plan that drives the real DOM on web over a ~20 KB runtime, and runs via a Rust runner with AppKit, UIKit, or a GPU painter natively, no webview or JS bridge.
- Agents drive all hosts through just ten operations on a seekable clock to avoid timing flakes, saved edits rebuild in about 99 ms, and first native builds take ten to fifteen minutes before later builds drop to a minute or two.
Netlify launches long-running Node.js servers LINK
- Netlify has launched Netlify Server in Early Access, a long-running Node.js backend that runs inside your project with the same build, preview and atomic-deploy workflow Netlify already uses for frontends.
- Each instance is a long-lived process handling many requests at once, so the event loop serves others while one waits on a database or model, keeping connection pools, caches, SDK clients and WebSockets in memory and scaling down to zero when idle.
- You deploy a single handler file or an unmodified Express, Hono or Fastify app with no ports, adapters or health checks, plus lifecycle hooks and custom scaling signals; it needs Netlify CLI 27.17.0 or later on credit-based or credit-enabled Enterprise plans.
Angular's Rust compiler cuts build time LINK
- Angular's team has rebuilt its compiler's analysis stage in Rust as the new Angular Preprocessor (ngp), decoupling from TypeScript's APIs to clear the path for TypeScript 7 adoption after its Go rewrite dropped the Transformer API Angular relied on.
- ngp runs a parallel Oxc-based Rust frontend to discover Angular declarations before downstream tools handle type-checking and emission; benchmarks across projects of varying size showed 2x-3x faster cold builds, with memory on the largest roughly halving.
- Teams should first update to Angular v20 or later, but keep production builds on the current compiler: ngp adoption is expected in a minor Angular 22 release once TypeScript 7.1 restores the compiler API.
New guide teaches coding agents web basics LINK
- Google shipped the stable 1.0 release of Modern Web Guidance, a set of skills that embed web platform best practices, accessibility rules, and Baseline compatibility data directly into coding agents so they write faster, more accessible, more secure web code.
- Version 1.0 adds a modernization workflow that statically analyzes existing HTML, CSS, and JavaScript to find safe upgrades, plus expert-reviewed guides and a public eval dashboard; Firefox and Edge have joined Chrome to make it a cross-browser effort.
- It installs via CLI or through plugin marketplaces including Claude Code, Codex, Cursor, and Copilot; early adopters like MetaMask ran it as a PR review bot that commented on 64 pull requests, with engineers fixing issues in a median 15 minutes.
Tutorials & Advice
> The Codebase Rewrite is Back: AI has slashed rewrite costs, making once-ignored migrations and backlog fixes worth batching, but only strong automated testing keeps agents from quietly breaking things.
> Glue work is still harmful in the age of AI: AI shrinks the uncredited coordination and reporting work by letting solo engineers ship, so avoid becoming the designated glue person.
> Long-horizon agents and durable execution: Durable execution logs each step an AI agent takes, letting weeks-long automated tasks survive crashes and resume rather than restart, making previously uneconomical work viable.
> PropFLIP: Animate layout changes by splitting movement into four separate CSS properties, so each axis of motion and scaling gets its own timing for natural-looking, interruptible transitions.
> Monads in Lean: Monads let functional programmers hide tedious bookkeeping like error-handling and state so code can focus on the actual computation.
> Not everything needs an LLM: Use lightweight decision models that return typed answers with confidence scores to screen inputs before paying for slow, expensive language-model calls.
Other news you might like
- Cloudflare Makes Basin Serverless Analytics Generally Available LINK
- How We Built TLX: A Hardware-Native GPU Compiler for Production ML at Scale LINK
- Linux kernel reviewer steps down from memory management duties due to influx of AI slop patches LINK
- reticlehq/reticle: AI agents can generate code, but still struggle to understand what they build. Reticle brings Jev-style machine-native runtime perception to web & desktop applications. LINK
- Apple's Claris Testing 'Harbor' App for Local AI on Macs LINK
Trending tools
tiun: an AI-native backend bundling authentication, payments, customer data, and analytics into one platform, so teams avoid stitching separate services together. LINK
Kilo Code: open-source mobile AI coding agent that sends push notifications when sessions need input, letting you dictate prompts and attach files remotely. LINK
Inferock: a local proxy for OpenAI, Anthropic, Gemini, and OpenRouter that tracks token usage, failures, and retries to catch billing overcharges. LINK
Onepin: a QA layer for AI voiceover that fixes mispronounced names, prices, and dates across 30+ voice engines without re-rendering takes. LINK
Humalike x GTA RP: offers behavioral infrastructure, APIs, models, and benchmarks that give AI agents the social skills and proactiveness they usually lack. LINK
Claude Opus 5.5: handles agentic coding and complex knowledge work with improved alignment testing, running typical workloads at 40% less cost than Opus 5. LINK
Trending papers & reports
Reasoning-history loss hits AI assistants very differently when a task hops between providers and their past notes get wiped, swinging accuracy from ~4 points worse to ~51 points worse depending on the model. LINK
Engineer AI token spend, the amount of AI assistant usage companies track to gauge productivity, barely reflects actual output, with top spenders burning up to 240x more than peers for similar work. LINK
Automated math-proof cleanup scored ~86 points versus ~33 for reference proofs on a competition leaderboard, while costing under 3 per problem by picking the cheapest capable AI for each step and locking lessons into checked rules. LINK
Enterprise AI agent testing shows that watching an agent's full step-by-step work catches failures best, cutting false alarms 5x versus text-only checks, while functional and security tests each find errors the other misses. LINK
On-device AI deployment gets an automated assistant that gets new models running correctly on phone chips 75% of the time, versus 33% for existing coding tools, cutting the costly engineering work of supporting each new model. LINK
See you tomorrow for a new dose of ☕️ Devshot!