X is the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for things I think are actually worth knowing. I use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). This week: 104 tweets liked, filtered down to what’s below.

Check out the previous roundup (Mar 16) if you missed it.

AI for Everyone

Claude Computer Use on Mac (5 mentions)

Claude can now control your Mac: mouse, keyboard, screen, all of it. Kick off a task from your phone via Dispatch, walk away, come back to finished work. It’s slower than a human on purpose. Best for repetitive stuff that doesn’t have an API: copying data between apps, filling out web forms, clicking through legacy software. It tries connected apps (Slack, Calendar) first and only grabs screen control when needed. macOS now, Windows in a few weeks. (source: @claudeai, @felixrieseberg)

Claude /schedule: Recurring Tasks on Autopilot

Tell Claude to check your Slack every morning, pull a weekly report, or scan your email on a schedule. It runs on Anthropic’s cloud even when your laptop is closed. Uses connected apps first, asks for screen control when there’s no integration. I’ve wanted something like this for years. (source: @claudeai, @noahzweben)

Longer-Term Claude Users Get Better Results

Anthropic’s Economic Index looked at how people use Claude over time. The longer someone’s been at it, the less they hand off blindly. They check work, course-correct, try harder things. And they get better results. Newer users expect magic and get disappointed. The adjustment that actually helps: treat it like a junior colleague, not a vending machine. (source: @AnthropicAI)

canirun.ai: Can Your Hardware Run It?

canirun.ai reads your GPU, CPU, and RAM from the browser and grades every major local model from “runs great” to “too heavy.” Filter by task, provider, or license. Free, no signup, thirty seconds. Check it before downloading a 20GB model that won’t run. (source: @CamilleRoux)

Gemini Flash-Lite Real-Time Browser

Google demoed a browser built on Flash-Lite where every page is generated in real time as you navigate. Click a link: the model writes the page. Not cached content, generated content. The demo makes “fast” feel concrete in a way benchmarks don’t. (source: @GoogleDeepMind)

Karpathy’s AutoResearch (2 mentions)

Everyone covered the ML angle. What got missed: AutoResearch works for anything with a repeatable test loop. A marketer ran it on landing page copy overnight, 56% pass rate became 92%, cost $25. If you have a clear success metric and can score it automatically, this tool works. (source: @aakashgupta, @shannholmberg)

AI for Developers

LiteLLM Supply Chain Attack (3 mentions)

LiteLLM versions 1.82.7 and 1.82.8 were backdoored. The malicious code runs on every Python startup, not just when litellm is imported. It collected SSH keys, cloud credentials, Kubernetes configs, crypto wallets, env vars, and shell history, then exfiltrated everything to a spoofed domain. The GitHub issue was closed as “not planned,” suggesting the maintainer’s account was compromised too. Run pip show litellm everywhere. If you find those versions, rotate every credential on that machine. (source: @karpathy, @hnykda)

Claude Code Auto Mode (4 mentions)

Auto mode replaces --dangerously-skip-permissions with something sane. Claude decides what’s safe, safeguards check each action, and you don’t babysit or nuke all guardrails. Team plan research preview for now. (source: @claudeai, @alexalbert__)

/schedule for Developers (6 mentions)

From the terminal, describe a recurring job and Claude runs it on Anthropic’s cloud. Loops run up to 7 days. The team uses it internally for CI triage and library sync. One user set up a Sentry polling loop that finds errors, creates fix PRs, reviews its own work, and emails when done. That’s an autonomous on-call engineer. (source: @noahzweben, @minchoi)

Cloudflare Dynamic Workers (2 mentions)

Dynamic Workers let a Worker spin up child Workers at runtime with AI-generated code. V8 isolate startup in milliseconds, megabytes of memory. Containers are 100x slower and heavier. If you’ve been trying to build a product where AI writes and runs code per request, this is the infrastructure piece that was missing. (source: @Cloudflare, @eastdakota)

Figma MCP Open Beta (2 mentions)

Figma’s MCP tool lets Claude Code design directly on the canvas with your actual design system: components, tokens, spacing. No more copying specs into prompts. I haven’t tried it yet but the early reports are strong. (source: @figma, @trq212)

Next.js 16.2 Goes Agent-Native (2 mentions)

Vercel found that bundled docs achieve 100% pass rate on Next.js agent evals vs 79% for skill-based retrieval. So 16.2 ships AGENTS.md by default and bundles full docs with every install. @vercel/next-browser gives agents React DevTools access from the terminal. Run npx @next/codemod@latest agents-md on existing projects. (source: @nextjs, @rauchg)

TypeScript 6.0

TS 6.0 turns strict mode on by default. Migration trap: types defaults to an empty array now, so add "types": ["node"] explicitly or builds break. The good stuff: built-in Temporal API (proper dates, finally), Map.getOrInsert, RegExp.escape. TS 7.0 will be native Go. (source: @typescript)

Linear Agent (2 mentions)

Linear shipped the Linear Agent, Skills, and Automations. Triage fires when an issue comes in, the agent enriches it with codebase context and routes it before a human touches it. Skills let you codify repeatable workflows once and forget about them. If you’re on Linear, turn on Triage today. (source: @linear, @cjc)

The No-IDE Workflow: Claude Code + Compound Engineering + Zed

Kevin Rose asked what IDE to set up on a new laptop. Matt Van Horn’s answer: don’t use one. His workflow is Ghostty as the terminal, Claude Code with the Compound Engineering plugin to generate plan.md files, Zed to review and edit those plans, then /ce:make to execute. He also uses Monologue from Every to talk to Ghostty by voice. This is very close to how I build this blog and side projects. The thread is worth reading because it’s the clearest description I’ve seen of what “vibe coding” actually looks like when someone does it for real work. (source: @mvanhorn, @kevinrose)

Honorable Mentions

  • Cursor Instant Grep. Millisecond search across millions of files. Your agents get faster at large codebases with no config changes. (source: @cursor_ai)
  • Claude Code 2.8x faster startup, Agent SDK 5.1x faster over 60 days. New experimental /init interviews you to set up skills and hooks: CLAUDE_CODE_NEW_INIT=1 claude. (source: @jarredsumner, @trq212)
  • App Store Preflight skill for Claude Code. Scans iOS/macOS projects against 100+ Apple Review Guidelines before submission. (source: @QingQ77)
  • Python supply chain attacks accelerating. Beyond LiteLLM, a separate package with 97M monthly downloads was also poisoned. Lock your deps and use hash verification. (source: @TukiFromKL)
  • ChatGPT file Library. Reuse uploaded files across conversations without re-uploading. (source: @OpenAI)
  • Google Drive doc scanner upgrade. Multi-page scanning, auto capture, duplicate detection. Testers say it beats iPhone Notes and Adobe Scan. (source: @minchoi)
  • Google TurboQuant. 6x memory reduction, 8x faster attention, zero accuracy loss. This is why your API calls will get cheaper. (source: @GoogleResearch)

Try This Weekend

For everyone:

  1. Download Claude Desktop and try Computer Use on one tedious multi-step task: expense report, moving data between apps, a web form on an annoying site.
  2. Visit canirun.ai before trying to run any local AI model. Thirty seconds saves hours.
  3. Try the Flash-Lite browser demo in AI Studio. Seeing real-time page generation is better than reading about it.
  4. Set up one recurring Claude task. Morning email summary, weekly report, Friday checklist.

For developers:

  1. Try the no-IDE workflow. Ghostty + Claude Code + Compound Engineering. Generate a plan.md, edit in Zed, run /ce:make.
  2. Set up a /schedule job for one recurring task. CI triage, doc sync, PR reviews. Runs on Anthropic’s cloud.
  3. Audit for litellm 1.82.7/1.82.8. Run pip show litellm in every env. If found, rotate every credential.
  4. Add the Figma MCP to Claude Code if you work with design systems.
  5. Add AGENTS.md to your Next.js project: npx @next/codemod@latest agents-md.