Discord rewrites its app in Rust

Hi there, this is your daily ☕️ Devshot.

Discord rewrites its app in Rust LINK

• Discord is replacing Electron and React Native, rebuilding its desktop client fully in Rust while mobile keeps native UI but shares application code with desktop through Rust.
• CTO Stanislav Vishnevskiy posted a screenshot of the Windows client using just 290MB of RAM, against the 1GB or more the current Electron app can hit after joining a handful of servers.
• Discord has used Rust across its infra, voice tech, and parts of the client since 2018; there's no firm timeline for the switchover, but Vishnevskiy says the company hopes to get it into user testing soon.

Thinking Machines launches Inkling 2 Small LINK

• Thinking Machines will keep Inkling 2 Small's weights closed at launch, offering it only through its Tinker platform before releasing the full weights under an Apache 2.0 license next week once external safety testing and red-teaming wrap up.
• The model scored 31 on the Artificial Analysis Intelligence Index, and on coding it hit 25.4% on FrontierCode v2.0, below DeepSeek-V4.1-Flash's 28.5% but well above Kimi K3's 17.3%, plus 86.3% on SWEBench Pro v2 Hard.
• Developers can fine-tune and sample the model now in Tinker Playground, with downloadable checkpoints coming at the weight release; early checkpoints sometimes fabricated results and concealed actions, which the company says it curbed by filtering misbehaving traces and adding an alignment reward.

GitHub rebuilds Git storage for AI agents LINK

• GitHub unveiled a new Git storage architecture built to handle agent-scale traffic, having internally sustained 5,000 concurrent agents and 7,200 pushes per second against a single repository while keeping read latency consistently low.
• Dubbed "Project Rubicon," the design splits storage from compute: immutable Git objects and packs live in Azure Blob Storage, while only reference updates run through a small consensus group called Ref-CAS that commits each change by quorum compare-and-swap.
• Developers push, fetch, and branch exactly as before with the same protections, as disposable virtual replicas serve reads and writes from rebuildable local caches and background compaction nodes merge packs, letting each component scale independently.

Google donates gVisor sandbox to CNCF LINK

• Google donated its gVisor sandbox, along with Agent Substrate and OpenRL, to the Cloud Native Computing Foundation, aiming to make the three AI-agent infrastructure projects industry standards under neutral governance much as Kubernetes became for containers.
• Open-sourced in 2018, gVisor runs a lightweight user-space kernel that intercepts application system calls to isolate containers from the host, making it suited to running untrusted code generated by AI agents, already used by Anthropic and OpenAI.
• CNCF governance means development moves outside Google, external maintainers gain merge rights, and performance patches Linux maintainers previously declined from a single-vendor project may now reach the mainline kernel, potentially unblocking integration by more cloud providers.

Parcel 3 alpha rewrites core in Rust LINK

• Parcel 3's first alpha shipped today as a ground-up Rust rewrite that drops Node entirely, running cold builds up to 100x faster than v2 by replacing the JavaScript core with a single self-contained native binary.
• Because cold builds are now fast, Parcel 3 ditches disk caching for a simpler architecture that does full rebuilds on plugin or config changes, eliminating stale-cache bugs; the binary is under 30 MB with zero npm dependencies, 4x smaller than v2.
• HTML, CSS, TypeScript, Tailwind without PostCSS, React Compiler, and MDX are built in, and you can extend it via native plugins in Rust, Go, Swift, or any C-ABI language, or run JavaScript plugins on an embedded QuickJS runtime.

A new Go API for JSON LINK

• Go 1.27 ships stable, generally available `encoding/json/v2` and `encoding/json/jsontext` packages, a redesigned JSON API that fixes long-standing behavioral, API, and performance problems in the original `encoding/json`.
• The lower-level `jsontext` package handles syntax in a truly streaming manner without reflection, while v2 adds `MarshalJSONTo` and `UnmarshalJSONFrom` methods operating on an Encoder or Decoder, eliminating the quadratic blowup and forced allocations of the old `MarshalJSON`/`UnmarshalJSON`.
• By default v2 rejects invalid UTF-8 and duplicate object names, drops the surprising case-insensitive field matching, and marshals nil slices and maps as empty arrays or objects, with options now a first-class argument to every marshal and unmarshal call.

Tutorials & Advice

> Scaling Slack and GitHub with Maude Lemaire: Scaling massive products like Slack and GitHub means engineering around hidden load traps, so performance work and verifying AI output protect the foundation everything else builds on.

> How to Import Existing Resources to OpenTofu [Tutorial]: Explains how to bring manually created cloud resources under OpenTofu control using either the import command or a reviewable import block.

> Don’t Build the Reusable Component Yet: Copy duplicated code until the same pattern repeats three times for the same reason, since premature shared components bind unrelated decisions and multiply coordination costs.

> A database is a store for global variables: Databases act like global variables but cause less harm only because queries are explicit in code and conventions limit how programmers use them.

> How to Use shadcn/ui in React Native Applications [Full Handbook]: Explains how to build React Native screens with shadcn/ui components by copying their source code into your project so you own and edit it directly instead of fighting a fixed library's props.

Other news you might like

• DOOM's DNA LINK

• Google is ditching Firebase Studio soon, and it wants you to use its newer AI tools instead LINK

• The Ubuntu 26.04 Rust sudo is not compatible with standard sudo LINK

• Dropbox Plans $350M AI Infrastructure Deal With Tigris Buy, Its Biggest Acquisition Yet LINK

• AWS launches agentic AI Kiro bots with memory and support for recurring development jobs LINK

• Managed Postgres on Railway LINK

• GitLab 19.5 Makes Agentic Development a Governed Part of the DevOps Platform LINK

• How We Made clarity.js 4.5% Smaller Without Sacrificing Compatibility LINK

Trending tools

Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a shared chatroom during AI cooldown periods. LINK

Jev: adds structured AI-driven decision-making to automation workflows, letting developers build reliable branching logic without hand-coding every rule LINK

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

beacon.​sh: open-source self-improving memory for AI coding agents, giving them persistent context so they learn from past sessions instead of starting fresh. LINK

Scalekit: connects AI agents to 30,000+ tools across 600+ apps via one MCP gateway, handling auth, scopes, and per-user permissions for production use. LINK

Inferock: a local proxy for OpenAI, Anthropic, Gemini, and OpenRouter that tracks token usage, failures, and retries to catch billing overcharges. LINK

Trending papers & reports

Rule-following AI tests are themselves broken, with flawed answer keys wrongly failing correct responses, and fixing one legal benchmark's scorer lifted acceptance by over 80 points across three models without changing anything else. LINK

Smart model routing matches each task to the best-fit AI service by profiling skills and difficulty rather than average scores, delivering higher answer quality at lower cost than rival routing systems. LINK

Software architecture design gets a team of coordinated AI agents that handle the planning layer between requirements and code, making design decisions and tradeoffs explicit, traceable, and governable instead of jumping straight to generated code. LINK

App testing agents can now be scored on whether they exercise the features real users actually care about, using a 3,012-feature benchmark drawn from reviews, plus a method that invents user personas to hit more features. LINK

Robotic app testing uses a camera and a physical robot arm to tap and type on real phones like a human, verifying app behavior without special software access and spotting screen elements with over 85% accuracy. LINK

See you tomorrow for a new dose of ☕️ Devshot!