Meta drops AI from performance reviews

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

📊 Meta drops AI usage from performance reviews after backfire

🌐 Chrome moves to two-week release cycle

⚡ Expo Modules 2.0 boosts native call speed

🧵 Neovim adds async library vim.async

Plus: 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

📊 Meta drops AI usage from performance reviews after backfire LINK

Meta scrapped AI tool usage as a performance-review metric after the criterion backfired, with executives saying internal reviews will now weigh the quality, speed, and complexity of engineers' work instead of dashboards and token counts.

The old rule drove "tokenmaxxing," where employees burned AI tokens in bulk to climb internal leaderboards, pushing Meta's internal AI spending toward billions in 2026 and prompting plans for budgets and a central usage dashboard in 2027.

The shift redirects evaluation away from raw token counters, meaning engineers no longer face pressure to inflate consumption just to look productive, and are instead judged on actual output rather than how heavily they lean on AI tools.

🌐 Chrome moves to two-week release cycle LINK

Google has moved Chrome to a two-week stable release cycle starting with version 153, dropping the four-week schedule it adopted in 2021 across desktop, Android and iOS.

The tighter cadence ships interoperability changes, bug fixes and security patches faster, aligning with a strategy that uses Gemini-powered automated tools to find and fix vulnerabilities, plus smaller, quicker downloads.

Organizations avoiding the biweekly pace can stay on the Extended Stable channel, which gets weekly security fixes and major feature updates every two months; Chrome 154 arrives September 22 and is in beta now.

⚡ Expo Modules 2.0 boosts native call speed LINK

Expo Modules 2.0 lets you write React Native native modules as plain annotated Swift or Kotlin classes, dropping the 1.0 DSL entirely, while making runtime calls faster than the API it replaces.

A build-time macro reads your Swift signatures ahead of time, eliminating 1.0's reflection-style per-call allocation and dynamic type conversion, making synchronous calls 2.5 to 5.6× faster than the 1.0 API on the same SDK.

You can try it now on iOS in SDK 57, which covers modules, functions, records, shared objects, and events via Swift macros; views and Android are still coming, with an official beta arriving in SDK 58.

🧵 Neovim adds async library vim.async LINK

Neovim shipped vim.async, a native structured-concurrency library in its Lua standard library that orchestrates asynchronous workflows without blocking the main event loop, replacing ad-hoc callbacks and third-party coroutine wrappers like plenary.nvim and async.nvim.

Routines run inside Tasks via vim.async.run(), scheduled cooperatively over stackful coroutines; vim.async.await() suspends a frame and yields to the loop, while child tasks attach to the parent scope and propagate unhandled errors to siblings.

Flow-control primitives include vim.async.semaphore() for permit limits, vim.async.timeout() for cancellation deadlines, vim.async.iter() for completion-order results, and vim.async.pawait() as an async pcall; Task:wait() and Task:detach() bridge synchronous code and independent tasks.

Other news you might like

• Kubernetes v1.37: Advancing Workload-Aware SchedulingLINK

• Leap second proposal will keep software stacks in syncLINK

• Super Smash Bros Melee gets fully decompiled after over six years of effort — ambitious and technically impressive project delivers GameCube classic as C codeLINK

• AI Made A Lot Of "Hideous" Code But Found Major Bottlenecks For Faster Linux CompilationLINK

• A “proof” of Fermat’s Last Theorem that fits the marginLINK

• Harness rebuilt its Git repository for nonstop AI agent trafficLINK

• Agents identifying as OpenAI systems wrote 17,000 posts to a wiki no one was supposed to write toLINK

🧰 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

SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while codingLINK

Humalike x Hermes: gives AI agents turn-taking, timing, and memory APIs to decide when to speak, wait, or interrupt, useful for game characters and roleplay.LINK

Kilo Code for JetBrains: open-source AI coding agent for JetBrains IDEs that helps you build, ship, and iterate on code with support for multiple modelsLINK

Meridian: open-source, locally-run activity tracker that logs your work with context, showing where your time went and the reasons behind it.LINK

📚 Trending papers & reports

Code review file ordering matters more than tools assume, with a survey of 1,355 developers finding only 10% like the default alphabetical order and 64% worry it makes them miss bugs.LINK

Prompts as engineering artifacts get a study protocol testing whether writing them like real software requirements, not just quick instructions, predicts coding success, using an experiment with 30 developers plus community and mined data.LINK

Self-healing cloud infrastructure predicts failures before they happen and automatically responds to incidents, cutting downtime and manual firefighting so enterprise systems stay online and scale without teams babysitting them around the clock.LINK

Coupled code generation teaches AI to write software whose pieces only work correctly together by learning from what actually happens when the code runs, beating leading rivals on seven of eight tricky problems.LINK

Coding-agent software catalogs get built fast but need heavy follow-up engineering, since the biggest risk is not crashes but silent failures where the tool confidently returns plausible yet wrong or incomplete answers.LINK

See you tomorrow for a new dose of ☕️ Devshot!