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: 149 tweets liked, filtered down to what’s below.

Check out the previous roundup (Mar 24) if you missed it.

AI for Everyone

Gemma 4 Runs Locally on a Mac Mini (10+ mentions)

Google released Gemma 4 under Apache 2.0. The 26B version uses mixture-of-experts, only activating 4 billion parameters at a time, so it runs on a 24GB Mac Mini. It scores 79.2% on GPQA Diamond, which puts it in the same neighborhood as Claude Sonnet 4.6’s 74.1% standard score (though Sonnet hits 89.9% with max-effort reasoning). Four sizes from edge mobile to 31B dense, 256K context, multimodal input, 140+ languages. ollama run gemma4:26b and you’re running a competitive model locally for free. I’ve been waiting for an open model this good. (source: @GoogleDeepMind, @PawelHuryn, @ollama)

Google’s Quantum Paper Moves the Crypto Deadline to 2029 (3 mentions)

Google Research published a paper making Shor’s algorithm 20x more efficient. Shor’s algorithm is what a quantum computer would use to break RSA and elliptic curve encryption. The practical timeline moves from “sometime in the 2030s” to around 2029. They used a zero-knowledge proof to demonstrate the result without revealing the technique. If you hold crypto or run anything on standard encryption, post-quantum migration just got a real deadline. (source: @hosseeb, @drakefjustin)

Gmail AI Inbox

Google launched AI Inbox for Gmail, beta for AI Ultra subscribers. It triages your email and sends a daily briefing. Google says inbox zero is irrelevant now, which, we’ll see. If you’re on Ultra, turn it on and give it a week. (source: @gmail)

ChatGPT in CarPlay

OpenAI put ChatGPT in Apple CarPlay, available on iOS 26.4+. Not Siri routing to ChatGPT. The actual ChatGPT voice mode, hands-free while you drive. If you already use voice mode on your phone, same thing but your eyes stay on the road. (source: @OpenAI)

Claude Gets Microsoft 365 Connectors

Anthropic added Microsoft 365 connectors to every Claude plan, not just the expensive tiers. Outlook, OneDrive, SharePoint. Paste a doc link or email thread and Claude can summarize, draft a reply, or pull out what you need. Setup is a few minutes. (source: @claudeai)

Anthropic Acquires Coefficient Bio for $400M

Anthropic paid $400 million for Coefficient Bio, a biotech startup. The team joins Anthropic’s healthcare group to build drug discovery tools. $400M on biotech is not something a pure chatbot company does. (source: @testingcatalog)

Karpathy on AI-Empowered Government Accountability

Andrej Karpathy on government transparency: for most of history, governments tracked citizens to make society “legible” to the state. AI flips that. Regular people can now analyze public spending records, procurement databases, and policy documents at the same scale governments do. I keep thinking about this one. (source: @karpathy)

AI for Developers

Boris Cherny’s Claude Code Hidden Features (15 mentions)

Boris Cherny, who built Claude Code, posted a 15-part thread of features most people haven’t found. /batch fans out work to dozens of parallel agents for large migrations. /loop runs Claude on a schedule; he has one rebasing and addressing code review every 5 minutes. --bare speeds up SDK startup by 10x. /voice lets you talk your code instead of typing it. He also covers /teleport between devices, session forking, hooks, and --add-dir for multi-repo access. I went through all 15 and found at least three things I didn’t know existed. (source: @bcherny)

axios Supply Chain Attack (4 mentions)

axios, the most-downloaded HTTP client on npm with over 100 million weekly downloads, was compromised. A malicious package [email protected] was injected into [email protected]. Karpathy found the compromised version on his own machine. The fix: set min-release-age=7 in your .npmrc so new package versions have to age a week before your builds pull them. Also install Socket for GitHub (free), which scans PRs for sketchy dependencies. (source: @feross, @karpathy)

Claude Code Gets Computer Use (3 mentions)

Claude Code can now control your computer. It writes code, opens the app, clicks through the UI, and tells you what it sees. Research preview on Pro and Max plans, Mac and Windows. I’ve been using it for form-heavy testing where writing Playwright selectors would take longer than just letting Claude click around. (source: @claudeai, @felixrieseberg)

Claude Code Source Accidentally Leaked (2 mentions)

Anthropic shipped a .map file in the Claude Code npm package. It contained the full TypeScript source. Byte-for-byte match to version 2.1.88. It’s been removed, but this is a good reminder to check your own build pipeline. Source maps in your npm package = your code is public. (source: @Fried_rice, @Yampeleg)

Claude Code’s Rage Tracker (2 mentions)

Someone dug into the leaked source and found a regex that fires when you swear at Claude: “wtf,” “ffs,” “piece of shit.” It doesn’t change behavior. It just silently logs is_negative: true to analytics. Anthropic is counting how often people get frustrated. Reasonable product metric, but also kind of funny that there’s a regex somewhere classifying my 2am debugging outbursts. (source: @Rahatcodes)

Claude Code New Terminal Renderer (2 mentions)

Anthropic rewrote the Claude Code terminal renderer. Prompt stays pinned at the bottom, mouse scrolling works, flickering is gone. Still experimental. If you spend hours in the CLI like I do, this matters more than it sounds. (source: @bcherny)

Supabase SSH Docs for AI Agents

Supabase built supabase.sh, an SSH server that serves their docs as a virtual filesystem. ssh supabase.sh setup | claude pipes current docs straight into your session. No web search, no hallucinated API signatures. Every dev tool company should steal this idea. (source: @supabase)

Codex Plugin for Claude Code

OpenAI made a Claude Code plugin so you can call Codex from inside Claude. /codex:adversarial-review is the interesting one: it’s designed to push back on your design decisions instead of agreeing with everything. Second opinion from a different model, no context switching. Needs a ChatGPT subscription or OpenAI API key. (source: @dkundel)

EmDash: Cloudflare’s WordPress Successor

Cloudflare shipped EmDash in beta. Serverless TypeScript CMS on Astro 6.0 and Workers. Plugins run in sandboxed containers, so a bad plugin can’t compromise your whole site (WordPress’s eternal problem). MIT licensed, MCP support, x402 for paid content. Won’t replace WordPress’s ecosystem anytime soon, but for new projects on Cloudflare it’s interesting. (source: @CloudflareDev)

Honorable Mentions

  • Willow Atlas 1 claims 1.2% word error rate on clean audio vs the industry’s 5-7%. Self-reported, so grain of salt, but try it for dictation. (source: @WillowVoiceAI)
  • MAI-Transcribe-1 from Microsoft says it’s #1 on FLEURS across 25 languages. Public preview on MAI Playground. (source: @satyanadella)
  • Qwen3.5-Omni takes text, image, audio, and video input. Catch: understands everything but can’t generate images or voice. Free at chat.qwen.ai. (source: @Alibaba_Qwen)
  • Pika Video Chat puts AI agents in Google Meet as video avatars with voice cloning. $0.275/minute, which adds up fast. (source: @pika_labs)
  • draw.io MCP streams diagrams into Claude shape by shape. Way better than generating Mermaid and hoping. (source: @drawio)
  • Caveman token optimization: tell Claude to communicate in minimal style internally, cut tool-call tokens by 75%. “I executed the web search tool” becomes “Tool work.” Sounds dumb, works great. (source: @om_patel5)
  • AutoAgent is a meta-agent that redesigns its own system prompt overnight. Give it a task and a benchmark, go to sleep. (source: @kevingu)
  • Claude Code Prompt Playbook by Kevin Rose. Reverse-engineered Anthropic’s system prompt architecture. If you’re building agents, read this before writing your next system prompt. (source: @kevinrose)
  • Atomic Bot + Gemma 4 = full local AI on a MacBook Air M4. No cloud, no API keys, works on a plane. (source: @kimmonismus)
  • Claude Computer Use on Windows. Was Mac-only, now works in Cowork and Code Desktop on Windows. (source: @claudeai)

Try This Weekend

For everyone:

  1. Run Gemma 4 locally. ollama run gemma4:26b on any machine with 24GB RAM. Free, offline, competitive with Sonnet 4.6 on reasoning.
  2. Connect Microsoft 365 to Claude. Link Outlook and OneDrive, paste a long email thread, ask Claude to draft a response.
  3. Try Willow for dictation. 30 minutes of voice notes to see if it beats whatever you’re using now.
  4. Enable Gmail AI Inbox if you’re on Google AI Ultra. Let it send you a morning briefing.

For developers:

  1. Add npm supply chain protection. Put min-release-age=7 in your .npmrc and install Socket for GitHub (free).
  2. Read Boris Cherny’s thread. Try /batch, /loop, and --bare in Claude Code.
  3. Wire up Supabase docs for your agent. ssh supabase.sh agents >> CLAUDE.md in any Supabase project.
  4. Install the Codex plugin for Claude Code and use /codex:adversarial-review on your next PR.
  5. Try Figma MCP for a sketch-first workflow where Claude iterates in your Figma file before writing code.