Expo runs iOS simulators in the cloud

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

☁️ Expo runs iOS simulators in the cloud

🔗 CUDA now runs on AMD GPUs

🍺 Homebrew 7.0.0 adds vulnerability scanner

🐉 Agoda replaces SQL cache with DragonflyDB

🔷 C# 15 adds union types

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

☁️ Expo runs iOS simulators in the cloud LINK

Expo's EAS Simulator now runs iOS Simulator sessions on cloud macOS infrastructure and streams them to a browser tab, letting developers build and test iOS apps from any Windows, Linux, or ARM laptop without a Mac.

A development build pointed at Metro over a tunnel delivers Fast Refresh straight to the cloud iPhone, and the web preview can inject a simulated camera feed, a test pattern or your own photo or video, into apps using standard AVFoundation APIs.

Coding agents in Claude Code, Cursor, or Codex can drive the device via the eas-simulator skill to build, tap through screens, and return screenshots, while every session records a replay you can link in pull requests as evidence.

🔗 CUDA now runs on AMD GPUs LINK

A new open-source project, Speedstu's "CUDA-for-AMD-Windows," lets CUDA-exclusive workloads run natively on AMD Radeon GPUs under Windows without virtualization or dual-booting, bridging the ZLUDA translation layer to AMD's HIP/ROCm SDK.

PowerShell scripts detect the GPU, pull a pinned ZLUDA v6-preview.69, and remap the CUDA driver API plus cuBLAS, cuSPARSE, and cuFFT to AMD equivalents; the author trained a 2.2M-parameter PPO network end-to-end on an RX 9060 XT.

Scope stays narrow: cuDNN, TensorRT, and NCCL don't resolve yet, so cuDNN-heavy tools fail, and the official ZLUDA path hit ~13,300 steps/sec versus ~3% slower for a legacy binary overlay.

🍺 Homebrew 7.0.0 adds vulnerability scanner LINK

Homebrew 7.0.0 has shipped with BrewUI, a native graphical interface for browsing, searching, and managing packages and their dependencies, available on macOS 26 "Tahoe" and later.

A new `brew vulns` command scans installed formulae, a single formula, or a Brewfile, sending upstream repos and versions to OSV.dev, then checks whether Homebrew already backported a security patch.

Sandboxing tightened as home directory access is now blocked by default and network downloads are separated from offline installs, while concurrent download and install make upgrades faster.

🐉 Agoda replaces SQL cache with DragonflyDB LINK

Agoda swapped its tier-one hotel Price Cache off a 72-shard Microsoft SQL Server setup to DragonflyDB, handling ~1.5 TB of volatile pricing across 300,000 reads and 1.5 million writes per second while cutting shard-routing complexity.

The move delivered roughly 8x better P99 read latency, with DragonflyDB serving about 300,000 requests per second at ~8ms and processing ~1.6 million writes per second at ~10ms P99 across a three-shard-per-cluster design.

Agoda validated the cutover with dual reads that compared supplier counts and price-data lengths at over 99.9% parity, then A/B-shifted traffic; two clusters now failover automatically when pods detect a 10-point cache-hit-ratio divergence within two minutes.

🔷 C# 15 adds union types LINK

Microsoft shipped .NET 11 Release Candidate 1 last week with a go-live license for production use, making C# 15 the default language version for projects targeting the new runtime across Visual Studio 2026 Insiders and VS Code.

C# 15 stabilizes features from the preview cycle, including union types, collection expression arguments, closed class hierarchies, non-virtual static interface members, labeled break and continue, and extension indexers, while awaiting inside unsafe contexts is now allowed.

The SDK extends dotnet test to Android, iOS, macOS, and Mac Catalyst projects, adds reproducible container image digests when SOURCE_DATE_EPOCH is set, and skips uploads for manifests already present in the target registry.

Other news you might like

• AI’s best coding agent fails 60% of the time — and the data backs it upLINK

• Your AI coding spend bought 25% more output. Duplication rose 81%.LINK

• Developer builds viral 3D source code visualizer that consumes 21GB of RAM — flies around 2.5 million lines of code at over 120 frames per secondLINK

• Amazon Linux 2027 Enters Public Preview with SELinux Enforcing by DefaultLINK

• Talos Linux takes aim at VMware with open source hypervisorLINK

• Grab's Agent Framework LLM-Kit Accelerates AI Agent Production DeploymentLINK

🧰 Trending tools

Prelint: reviews AI-generated pull requests against your ADRs, docs, and past decisions, catching product drift before it reaches productionLINK

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

Dograh: open-source voice agent platform with a visual flow builder, telephony, human transfer, and self-hosted QA monitoring across 30-plus model integrationsLINK

Coldtea.ai: automates QA and monitoring using AI agents that catch post-deployment bugs, flag issues, and generate actionable fix tasks for your teamLINK

Kane CLI: drive real Chrome with plain-English commands to test web apps or extract data, avoiding brittle selectors and scripts for developers, QA, and AI agentsLINK

Openbase: helps developers choose npm and open-source packages by comparing popularity, maintenance activity, and reliability metrics alongside real user reviewsLINK

📚 Trending papers & reports

Tool-using AI agents invent answers ~45% of the time when a tool silently returns broken data without an error flag, yet one added sentence forcing a pass or fail check cuts that fabrication to under 1%.LINK

Secure code generation gets a runnable testing benchmark of 255 Python tasks plus an agent that spots risky spots before writing, making AI-written code measurably safer without breaking whether it actually works.LINK

Automated test writing gets faster by having a new type of fast-generating AI reuse the repeated structure across tests for the same code, cutting generation time without lowering the quality or coverage of the tests produced.LINK

Autonomous coding agents often draw broader conclusions than their experiments actually support, so this work builds a traceable handoff that records what each test proved, its limits, and when a decision should be reopened.LINK

Skill library audits stress-test how an AI agent's toolkit works as a whole, showing that overlapping or vaguely named skills, not weak individual ones, cause up to ~13 point accuracy drops and quadrupled costs.LINK

See you tomorrow for a new dose of ☕️ Devshot!