Cursor builds web apps from prompts

Hi there, this is your daily ☕️ Devshot.

Cursor builds web apps from prompts LINK

• Cursor can now take an AI-generated app from a natural-language prompt all the way to a live, shareable URL, handling hosting, databases, and authentication so users never touch the underlying infrastructure themselves.
• The feature runs only through Cursor's cloud agents, deploys on Vercel for Platforms without users creating Vercel accounts, provisions PostgreSQL when needed, and works from either a new Cursor Origin repo or an existing GitHub one.
• Before publishing, Cursor runs deterministic and agentic security checks for exposed secrets and vulnerabilities, but it's unclear how well these catch misconfigured database permissions or access controls, so sensitive internal deployments still warrant extra scrutiny.

PyTorch 2.15 survives a single bad GPU LINK

• PyTorch 2.15 shipped a rewritten NCCL backend so a single failing or slow GPU no longer forces a distributed training job to restart from scratch, letting an orchestrator drop the laggard and swap in a replacement worker.
• The new ProcessGroupNCCL enforces one communicator per process group at initialization, killing the crosstalk, intermittent deadlocks, and unpredictable performance of the old version, which could spawn dozens of communicators per group.
• A "flight recorder" now snapshots all operations for post-failure inspection instead of a generic stall message, and the legacy backend ships only as a rollback, slated for removal in 2.16 early next year.

GKE timeout traced to Linux kernel bug LINK

• A Google support case that began in January 2026 with intermittent GKE pod timeouts to the Kubernetes API server ultimately unravelled into three separate bugs spanning Private Service Connect, GKE's TCP stack, and the Linux kernel, with the final fix shipping as a kernel patch.
• Packet captures showed fresh SYNs getting back a stale ACK from a prior connection, traced to a February 2026 USPS rollout that silently dropped RFC 5961 challenge ACKs, something GKE's stack quietly relied on to clear stale state, worsened by NAT causing frequent 5-tuple collisions.
• About 800 developers were blocked from deploying to production for weeks as europe-west clusters stalled; a February rollback of the USPS change restored challenge ACKs and recovered clusters by early March 2026, but left unexplained why the KCP VM lingered in FIN_WAIT_2 over two minutes.

Lima adds experimental Windows VM driver LINK

• Lima 2.3 shipped on Friday with a new experimental `hcs` driver that lets VMs start natively on Windows through the same Host Compute System behind WSL2, dropping the need for QEMU or a Unix toolchain.
• The `hcs` driver talks directly to the Windows HCS API, but currently supports only `plain` mode, requires an explicit `--dns` flag, allows just one instance at a time, and offers no file mounts or dynamic port forwarding yet.
• Alongside Windows gains, v2.3 brings real-IP networking to Linux VMs via a privileged `lima-privileged-net` helper, block-device attachment for macOS VZ guests, a `limactl menu` TUI, and tighter AI-sandbox controls like `--sync-exclude`.

OpenTelemetry JavaScript SDK 3.0 is stable LINK

• OpenTelemetry JavaScript has shipped SDK 3.0 as stable, alongside experimental packages and a new API release, marking Logs as stabilized so that Traces, Metrics, and Logs are now all stable in JS.
• The release raises the minimum Node.js version by dropping support for Node.js 18 and 20, ships proper ESM and CommonJS exports for every package, consolidates Tracing SDK packages, and adds complex attribute support per OTEP 4485.
• Several public types and class names were renamed or moved and many deprecated packages removed, though the API package has no breaking changes and zero-code instrumentation users need not migrate; 2.x stays supported for one year.

Tutorials & Advice

> React.lazy, Suspense and the Hidden 300 ms Startup Delay: Render your own loading placeholder instead of letting React suspend, because React deliberately holds suspended content back 300 milliseconds, delaying startup needlessly.

> parse_number() as a SQLite extension: C vs Rust: Prompting Claude to compare and then optimize implementations shows a hand-tuned C parser can beat Rust's faster number-parsing, yielding a far smaller, quicker SQLite add-on.

> How do we auto-approve and merge 60% of PRs?: Shift human review from code to design and user experience so agents can safely auto-merge most low-risk changes, freeing engineers from rubber-stamping.

> true or false in CSS: Store true/false keywords in registered CSS variables and read them in container style queries, making multi-condition styling logic far easier to read.

Other news you might like

• Run Linux applications in hypervisor-isolated micro-VMs with Hyperlight LINK

• Inside the new JNI layer in Expo Modules 2.0 LINK

• GCC 17 Feature Development Ends Next Week LINK

• Linus Torvalds sees signs AI-prompted patch flood has peaked LINK

• vercel-labs/agent-browser: Browser automation CLI for AI agents LINK

• Open Source pg_vectorsearch Brings More Scalable Vector Search to PostgreSQL LINK

• Homelabber uses second-gen Optane memory sticks to run a 2.8-trillion-parameter AI model on CPU alone 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

iFixAi: audits AI agents with 250 inspections across 69 misalignment categories, flagging failures and giving engineers evidence to investigate and fix issues LINK

Humalike x GTA RP: provides behavioral infrastructure, APIs, models, and benchmarks that give AI agents the social skills and proactiveness they typically lack. LINK

Claude Opus 5.5: handles agentic coding and complex knowledge work with stronger alignment testing, running typical workloads at 40% less cost than Opus 5. LINK

FastRouter.​ai: routes requests across 200+ AI models through one OpenAI-compatible API, with intelligent routing, automatic failover, and governance for enterprise teams. LINK

Trending papers & reports

Design-agent benchmark tests whether general-purpose AI assistants can handle a full professional design job from understanding client intent to production-ready files, finding they finish tasks but often miss specific client requirements. LINK

AI coding assistants invent library features that don't exist in up to 40% of answers, and cheap code-checking tools catch only some of these errors, leaving a hard ceiling no automated check can fully fix. LINK

Software planning agents can draft the requirement documents and effort scores that normally start a project, tested on a real fintech build as a feasibility check, though the scores aren't yet verified against actual development time. LINK

Database query foresight tests whether AI agents can predict what a database command will actually do before running it, finding top models wrong 8 to 30 percent of the time, even on queries they wrote themselves. LINK

AI agent safety rules can silently stop working after routine operations like reapplying a plugin or saving a permission, letting a banned action slip through even with every safeguard still switched on. LINK

See you tomorrow for a new dose of ☕️ Devshot!