Hi there, this is your daily ☕️ Devshot.
In today's Devshot:
🪁 Cloudflare open-sources AI agent browser
🧠 Alibaba's Qwen4-Max cuts task costs 62%
⚙️ GitHub rebuilds Copilot's diff viewer
🦙 llama.cpp hits 1.0 after 3.5 years
🟢 Node.js 26.10 adds debounce and throttle
Plus: 🎁 6 more stories you might like, 🛠️ 5 tutorials, 🧰 6 tools, and 📚 5 papers.
Cloudflare open-sources AI agent browser LINK
- Cloudflare has open-sourced Kitesurf, its stateless browser built for AI agents, publishing the code on GitHub under the Apache 2.0 license so developers can run and deploy their own browser engine on Cloudflare Workers.
- Rather than spinning up a browser container, Kitesurf parses HTML, applies CSS, runs page JavaScript, and renders screenshots using Rust packages compiled to WebAssembly, aiming for lower CPU and memory use than Chromium while supporting the Chrome DevTools Protocol.
- You can connect Puppeteer or point an AI harness at the local CDP port, and deploy to your own account with a single command, though a Workers Paid plan is required because Kitesurf uses Dynamic Workers to isolate pages.
Alibaba's Qwen4-Max cuts task costs 62% LINK
- Alibaba released Qwen4-Max, a 2.4-trillion-parameter flagship model that it says completes professional tasks for 62% less than its predecessor, Qwen3.8-Max.
- The mixture-of-experts model keeps the same parameter backbone but activates roughly a quarter fewer parameters per token, supports a 1-million-token context window, uses 35% fewer tokens at high-effort reasoning, and roughly cut both its input and output list prices.
- Analysts urged treating the savings as a vendor benchmark, recommending enterprises measure cost per accepted result including retries and review time, and warned that upcoming open-weight Qwen4 models lose their safety training once fine-tuned, requiring fresh adversarial testing.
GitHub rebuilds Copilot's diff viewer LINK
- GitHub has rebuilt the diff surface in the Copilot app so reviewers can open a pull request with 2,200 files, over a million changed lines and more than 400 inline comments without the review stalling.
- The team split the document into two domains: fixed-height code rows whose geometry is computed once, and variable-height comment blocks keyed by identity, measured in a single idle-gated batch scoped to roughly 2,400 pixels around the viewport.
- An early version wrote each block's height back into the layout, creating a feedback loop; the shipped design defaults to flagging instead, writing synchronously only when a user toggles a section so content below moves in the same frame.
llama.cpp hits 1.0 after 3.5 years LINK
- llama.cpp has reached 1.0 after 3.5 years, with the Hugging Face team formally committing to a stable, versioned API for the many downstream apps that build on the library for local LLM inference.
- Both llama.cpp and ggml now follow semantic versioning, signaling API and ABI incompatibilities with major-version bumps, while the engine runs on virtually any hardware via dynamically loadable ggml backends and supports over 150 model architecture variants plus mainstream speculative decoding.
- Expanded CI infrastructure using HF Buckets and HF Jobs plus self-hosted runners across NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel validates every change, and the team also shipped Llama desktop apps for MacOS and Windows.
Node.js 26.10 adds debounce and throttle LINK
- Node.js 26.10.0 pulled debounce and throttle rate-limiting helpers directly into the core utility module, long one of the ecosystem's most commonly vendored dependencies from lodash or the throttle-debounce package.
- James M Snell landed both in a single pull request and reworked the cancellation signal to abort every future invocation rather than fire once, plus added a leading-edge option so the first call can run immediately.
- The release also surfaces private keys and certificates from .p12 and .pfx bundles via a new OpenSSL-backed API, extends the experimental Virtual File System, and continues Node's push to cut small userland dependencies out of typical trees.
Tutorials & Advice
> The Path to Rails 8.1: From Vue and jQuery to Stimulus: When replacing browser JavaScript, first ask whether a plain HTML element or the server can do the job, leaving minimal code to maintain.
> Reading JSON at 277 GB/s in Python: Use Python 3.14's threading to parse JSON across all cores at once, reading only the fields you need for far faster processing.
> Probabilistic data structures in Go: Approximate data structures like Bloom filters and count-min sketches answer membership and frequency questions using a fraction of the memory exact storage needs.
> Your coding agent shouldn't grade its own homework: Use a separate, independent AI agent to test code live in a real browser, since coding agents miss their own mistakes.
> Sometimes three boring statements beats one clever expression: Split dense chained expressions into several plain statements because they read faster and let you log or debug each step individually.
Other news you might like
- Remix v3: Fast Pages, Thin Ecosystem LINK
- DuckDB has an answer to “bloated” database messages LINK
- Relax: Rust without the wait LINK
- Announcing the GraphQL over HTTP specification LINK
- Developer creates a 100% valid ZIP file that doubles as a BIOS-bootable operating system LINK
- Migrating our Email API to Go LINK
Trending tools
Jev: adds structured AI-driven decision-making to automation workflows, letting developers build reliable branching logic without hand-coding every rule LINK
Monospace from Directus: wraps any SQL database with instant REST and GraphQL APIs plus a no-code app for collaborative, team-friendly data management. LINK
iFixAi: audits AI agents with 250 inspections across 69 misalignment categories, flagging failures and giving engineers evidence to investigate and fix issues LINK
Floot MCP: connects Claude and ChatGPT to build full-stack apps from chat, handling databases, logins, hosting, emails, and deployment automatically. LINK
Beam: streams individual windows from a remote Mac as native local apps, each with its own Dock icon and ⌘Tab slot over an encrypted connection. LINK
AUDR by Chargebee: an open standard for recording who initiated agent runs and their costs across every system a run touches, joining usage to business context. LINK
Trending papers & reports
Coding chatbot leaderboards can now rank AI by actually running the code and letting people see the results, making human quality judgments far more reliable across 14,000 sessions in 10 programming languages. LINK
AI-built app deployment gets a safety net that automatically fixes common launch-killers like exposed passwords and missing health checks, with ~88% of repairs using preset templates rather than unpredictable AI-written code, passing independent verification. LINK
Coding-assistant tuning tested seven cheaper ways to customize code-writing models up to 16 billion parameters, finding that one low-cost method usually gives the best bang for the buck while full retraining still wins on raw accuracy. LINK
HTML5 game-building test grades AI agents on whether they actually deliver playable games, finding top systems pass under 50% of requirements and stumble most on AI opponents and overall game feel. LINK
Coding-agent evaluation shows that giving an AI a map of a software project's architecture boosts its accuracy to fix issues by ~16% to ~86%, often beating a newer model that lacks the map. LINK
See you tomorrow for a new dose of ☕️ Devshot!
