Hi there, this is your daily ☕️ Devshot.
In today's Devshot:
⚡ Anthropic ports model to AMD
🤖 Claude code adds experimental code mode
🐘 Next.js memory bug traced to timer
📦 KubeVirt graduates from CNCF
🦀 Rust 1.100 stabilizes the never type
Plus: 🎁 7 more stories you might like, 🛠️ 4 tutorials, 🧰 6 tools, and 📚 5 papers.
Anthropic ports model to AMD LINK
- Anthropic ported Claude to AMD GPUs in about three days with almost no human intervention, using its own frontier model, Mythos, to break its reliance on Nvidia and run Claude across all four major chip families.
- Mythos ported or wrote nearly 100 percent of the kernels itself, collapsing a process that previously tied up dozens of engineers for months and that most of the industry had judged not worth attempting.
- If accurate, the figures from Anthropic's unverified investor presentation suggest CUDA's software lock-in is far weaker than assumed, letting labs switch vendors for better price-performance and sidestep Nvidia's unusually high 70%-plus margins.
Claude code adds experimental code mode LINK
- With Code Mode turned on, Claude writes JavaScript that calls its tools inside a script instead of issuing each tool call directly, which lets it chain several operations together within one block of code.
- Writing JavaScript to invoke tools cuts down the back-and-forth of standard tool calling, where Claude must stop and wait after each separate call, letting it loop, filter, and combine results before returning an answer.
- Developers who want to try this JavaScript-driven tool calling can follow the instructions at code.claude.com/docs/en/code-mode to switch it on, and administrators can use the same docs to turn it off for an entire organization.
Next.js memory bug traced to timer LINK
- Zepto engineers traced a Next.js memory problem, pods pinned at roughly 5 GB with garbage collection freeing nothing, to react-query cache-expiry timers firing during SSR, each one keeping a finished request alive until it expired minutes later.
- Every server-rendered query starts a 2-to-5-minute eviction timer that never gets cancelled because subscriptions run inside effects that don't fire during SSR, and Node pins the request's AsyncLocalStorage store, response, buffers, spans, to each pending timer as a GC root.
- Giving the per-request server QueryClient a cache with `cacheTime` set to Infinity skips the timer entirely, cutting production memory per pod by about three-quarters; v4/v5 only change the default, so hooks passing an explicit finite `cacheTime` still need this guard.
KubeVirt graduates from CNCF LINK
- KubeVirt, the project that runs virtual machines as workloads on Kubernetes, has graduated from the CNCF, reaching the foundation's top maturity tier seven years after it entered the sandbox.
- The graduation lands as KubeVirt lets a single Kubernetes control plane schedule VMs, containers and AI workloads together, with a consistent API spanning edge devices and the cloud and much recent feature work driven by AI demand.
- Next on the roadmap is passing GPUs, CPUs and memory through to VMs, and maintainers point to virtualization's isolation model as newly valuable for AI; Nvidia was an early adopter and Microsoft is now ramping up.
Rust 1.100 stabilizes the never type LINK
- Rust 1.100 shipped today, finally stabilizing the never type `!` after ten years and five failed attempts, giving developers a stable way to represent diverging computations that never resolve to a value.
- Any code following a never type is treated as unreachable; the release also stabilizes core parts of the Allocator API and adds Cargo reporting for unused dependencies.
- A new Cargo `min-publish-age` setting lets you impose a dependency cooldown, skipping crates published in the last few days to dodge malicious packages often caught within hours of release.
Tutorials & Advice
> Build once, run on any agent: A practical guide to agent skills: Capture repeatable tasks as portable, versioned skill files that work across different AI agents, keeping project settings separate so teams reuse them without rewriting.
> Agents are users, so your software needs evals: Build repeatable grading tests for AI agents using your software, because unlike human users they can complete a request while silently delivering the wrong result.
> Your JS Error Is Lying to You: Preserve the original failure in an error's cause chain rather than trusting the top-level message, because each wrapping layer hides what actually broke.
> Top Types of Hidden Coupling and How to Deal With Them: Identify dependencies that architecture diagrams miss, like shared databases or synchronized deployments, because hidden coupling quietly threatens system stability and complicates changes.
Other news you might like
- How a false smoke alarm almost drowned Jane Street’s only NYC datacenter LINK
- Tim Peters, RIP LINK
- Local vector search for coding agents: From 507 MB to a 7-line loop LINK
- Introducing Expo Simulators: cloud simulators built for agents LINK
- Hobbyist turns $45 USB-C hub into a working Nintendo Switch 2 dock by patching its firmware LINK
- Uno Platform and Studio Get Major Agentic AI Updates LINK
- OpenAI Codex Helps Uncover & Fix A Nearly 20 Year Old Radeon Linux Driver Bug LINK
Trending tools
tiun: an AI-native backend combining authentication, payments, customer data, and analytics in one platform, so teams skip stitching separate services together. LINK
Kilo Code: open-source mobile AI coding agent that pushes 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 corrects mispronounced names, prices, and dates across 30+ voice engines without re-rendering takes. LINK
Humalike x GTA RP: provides behavioral infrastructure, APIs, models, and benchmarks that help AI agents develop the social skills and proactiveness they typically lack. LINK
Claude Opus 5.5: handles agentic coding and complex knowledge work with stronger alignment safeguards, running typical workloads at 40% less cost than Opus 5 LINK
Trending papers & reports
Stack Overflow edit screening can flag the half of volunteer post-edits most likely to be rejected while catching ~81% of bad ones, letting moderators review far fewer submissions without missing problems. LINK
Quantum circuit checking verifies that compiler optimizations preserve a program's meaning, getting correct verdicts on ~96% of test cases versus at most ~69% for rival tools while cutting total runtime by roughly half. LINK
Automated pull-request testing can now generate tests that specifically guard new code changes, catching up to ~90% of them versus ~35% for the prior best tool, reducing future bugs from unprotected updates. LINK
Shared 3D modeling classrooms let software students inspect designs together in a browser, but the study found usability scores that don't add up, flagging shaky evaluation reporting and navigation and distraction problems before any learning claims can be trusted. LINK
Front-end coding tests reveal that AI building apps scores ~88% on individual tasks yet delivers a fully working app only ~52% of the time, exposing a gap between partial credit and real success. LINK
See you tomorrow for a new dose of ☕️ Devshot!
