Silv's AI Weekly: Claude Computer Use, a scary supply chain attack, and TypeScript 6
Hey — here's this week's AI roundup from silv.blog.
Curated from ~104 tweets liked by @mattsilv, Mar 16-24, 2026:
AI for Everyone
Claude can now control your Mac. Mouse, keyboard, screen. Kick off a task from your phone, walk away, come back to finished work. Best for repetitive stuff without an API. Read more →
Claude /schedule runs recurring tasks on autopilot. Check Slack every morning, pull weekly reports, scan email. Runs on Anthropic's cloud even when your laptop is closed. Read more →
Experienced Claude users get better results by iterating more, not delegating more. Anthropic's Economic Index found longer-term users check work, course-correct, and attempt harder tasks. Read more →
canirun.ai tells you which local AI models your hardware can run. Reads your GPU/CPU/RAM from the browser, grades every model. Free, 30 seconds. Read more →
Gemini Flash-Lite generates web pages in real time as you browse. Every click creates a fresh page. Try the AI Studio demo. Read more →
Karpathy's AutoResearch works for anything with a test loop. A marketer ran it on landing page copy overnight: 56% pass rate became 92%, cost $25. Read more →
AI for Developers
LiteLLM 1.82.7/1.82.8 were backdoored. Malicious code runs on every Python startup, steals SSH keys, cloud creds, crypto wallets, everything. Run pip show litellm everywhere. Read more →
Claude Code auto mode replaces --dangerously-skip-permissions. Sane middle ground: Claude decides what's safe, safeguards check each action. Team plan preview. Read more →
/schedule for devs: describe a recurring job, Claude runs it on cloud. Loops up to 7 days. One user built a Sentry polling loop that creates fix PRs autonomously. Read more →
Cloudflare Dynamic Workers: AI sandboxes 100x faster than containers. V8 isolate startup in milliseconds. The missing piece for AI code execution at scale. Read more →
Figma MCP open beta. Claude Code designs on canvas with your actual design system. Read more →
Next.js 16.2 ships AGENTS.md by default. Bundled docs hit 100% agent eval pass rate vs 79% for retrieval. Read more →
TypeScript 6.0: strict mode on by default. Add "types": ["node"] explicitly or builds break. Built-in Temporal API. Read more →
The no-IDE workflow: Ghostty + Claude Code + Compound Engineering + Zed. Matt Van Horn's answer to Kevin Rose asking what IDE to install: don't use one. Read more →
Honorable Mentions
- Cursor Instant Grep — millisecond search across millions of files
- Claude Code 2.8x faster startup + new experimental
/initinterview mode - ChatGPT file Library — reuse uploads across conversations
- Google Drive doc scanner — multi-page, auto capture, beats iPhone Notes
- Google TurboQuant — 6x memory reduction, 8x faster attention, zero accuracy loss
Try This Weekend
For everyone: Download Claude Desktop and try Computer Use, visit canirun.ai before running local models, try the Flash-Lite browser demo.
For developers: Try the no-IDE workflow (Ghostty + Claude Code + Compound Engineering), set up a /schedule job, audit for litellm 1.82.7/1.82.8.
Read the full post with all sources and links
Know someone who'd find this useful? Forward this email or have them subscribe:
You're receiving this email because you subscribed to the silv.blog weekly AI digest. Unsubscribe anytime.