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: ~110 tweets liked, filtered down to what’s below.

Check out the previous roundup (Mar 8) if you missed it.

AI for Everyone

Claude Opus 4.6 Gets 1M Context Window at No Extra Cost

Anthropic dropped the long-context surcharge. 1M tokens is now standard for Opus 4.6 and Sonnet 4.6, same price as any other request. Feed in an entire codebase, a stack of PDFs, hundreds of images. For Claude Code users on Max, Team, or Enterprise, it’s already the default. Benchmark that matters: 78.3% on MRCR v2 at 1M tokens, best-in-class. (source: @claudeai, @bcherny, @ClaudeCodeLog)

Tech Guy Built a Custom Cancer Vaccine for His Dog Using AI

This story went everywhere this week, and it deserves to. A tech guy in Australia adopted a dog with terminal cancer, was told she had months to live, and decided to try something different. He spent $3k to sequence the tumor DNA, fed it to ChatGPT and AlphaFold to identify mutated proteins, then worked with UNSW’s RNA team to manufacture a custom mRNA vaccine. No biology degree. No lab access. The dog survived. Grok verified the story is real. Every step used publicly available tools. (source: @birdabo, @IterIntellectus)

Perplexity Computer Expands: Pro, Mobile, Android

Perplexity Computer is now available to all Pro subscribers and just landed on iOS, with Android rolling out. The demo that got attention: pulling analytics across 5+ platforms into a single dashboard in one query. The document review feature (“Final Pass”) runs 5 parallel reviews and returns a marked-up version. Perplexity’s own legal team used it on their MNDA and actually implemented the changes. Going from Max-only to Pro is a big audience expansion. (source: @perplexity_ai, @AravSrinivas)

Pokemon Go Players Built a 30 Billion Image AI Map

143 million people thought they were catching Pokemon. They were actually building one of the largest real-world image datasets ever assembled. Niantic disclosed: 30 billion images from AR scans of real locations, now training visual navigation AI for delivery robots. Nobody opted in knowing that was the deal. Not necessarily sinister, but it’s a clear model for how consumer apps extract training value at scale. Every photo filter, every AR game, every “scan your room” feature is the same pattern. (source: @Newsforce, @markgadala)

Humanoid Robot Learns Tennis in 5 Hours

Five hours of motion capture data. That’s all it took to train a humanoid robot to sustain multi-shot tennis rallies with a human player, hitting balls at over 15 meters per second with 90% accuracy. A year ago this would have required months of training data. AlphaGo was software. These are robots in physical space. (source: @Rewkang)

NBA AI Swarm Model Turns $1.49M on Polymarket

Someone ran a swarm of AI agents, each trained to think about NBA games differently, let them vote on Polymarket predictions, and walked away with $1.49 million. The insight isn’t that AI is good at sports. It’s that a diverse crowd of AI perspectives outperforms a single optimized model, same reason human prediction markets beat individual experts. The tool (MiroFish) is open source. Most people will file this under “crypto gambling” but it’s really an ensemble AI design case study. (source: @k1rallik)

Karpathy’s AI Job Exposure Map

Karpathy took 342 BLS occupation categories, had an LLM score each one’s AI exposure 0-10, and published an interactive treemap. If your entire job happens on a screen, you’re highly exposed. Manual trades and physical jobs score lower. Not a prediction, just a directional signal. Worth looking up your own category. (source: @_kaitodev)

Travis Kalanick’s Robots Company: Atoms

Travis Kalanick’s next company is Atoms: industrial robots for food manufacturing, mining, and transport. Not humanoids. His thesis is that the humanoid form factor is wrong for most industrial jobs. You want robots optimized for specific tasks, not general-purpose machines wearing the wrong body. Whether he’s right, who knows, but when the Uber guy bets on purpose-built over humanoids, I pay attention. (source: @travisk)

CashClaw: Open-Source Agent That Finds and Does Paid Work

CashClaw is an AI agent that finds jobs on task marketplaces, quotes a price, does the work, and collects payment. It’s now open source. Whether it actually makes meaningful money is TBD, but it learns from each completed task to price and execute better next time. Early, but the “agent that earns its own income” concept is going to keep showing up. (source: @moltlaunch)

AI for Developers

Chrome 146 Enables Native Browser Agent Control

This is the one I’m most excited to try. Before Chrome 146, getting an AI agent into your real browser meant janky extensions, launch flags, or surrendering your logins to a headless browser. Now it’s one toggle in Chrome settings. Flip it, and tools like Claude Code can use your actual open tabs and existing logins. Caveat: if you have hundreds of tabs open (guilty), the official MCP client apparently hangs. Browser Use CLI or agent-browser handles it better. (source: @xpasky, @shawn_pana, @browser_use)

Claude 2x Usage Promotion

Free bonus for the next two weeks: Anthropic is doubling usage limits during off-peak hours and all weekend, for everyone. Nothing to turn on. If you’ve been rationing Claude Code sessions or hitting limits mid-project, now’s the time to let it run. The promotion runs through March 27. (source: @claudeai, @bcherny)

Manus Desktop: Local AI Agent on Your Machine

Manus went from cloud-only to running on your machine. It can sort thousands of photos or rename hundreds of invoices without you touching anything. Each terminal command needs your approval, which is the right call. Available now for Mac and Windows. Whether it works on your messy real files is another story. Promo videos are always polished. Throw a real messy folder at it and see. (source: @ManusAI, @testingcatalog)

Claude Code 2.1.75 + 2.1.76

Two releases this week. The bigger change in 2.1.76: MCP servers can now pause mid-task and ask for your input via a form or URL. Agents hand control back at decision points instead of guessing or stopping cold. You can also turn off transcript persistence now if you don’t want session records. (source: @ClaudeCodeLog, @ClaudeCodeLog)

Claude Code Voice Mode

Voice mode is now in Claude Code on desktop. Sounds like a gimmick but it’s actually useful for the “explain what I’m looking at” or “what should I do next” style prompts where typing feels like extra friction. Try it once and see if it sticks. (source: @amorriscode)

Launch Claude Code Sessions from Your Phone

You can now start a Claude Code session from your phone and have it run on your laptop. Useful for queuing up a task from the train before you’re even at your desk. Boris called it mind-blowing. That’s overselling it, but the workflow shift is nice. (source: @bcherny)

agent-browser Goes Full Rust

agent-browser went from Node.js to pure Rust and the numbers are striking: 18x less memory, 99x smaller binary. For a tool you’re running inside agentic loops, that overhead reduction actually matters. It shipped 5 patch releases in 8 hours one day this week. Moving fast. (source: @ctatedev, @ctatedev)

OpenSquirrel: Agent-Centric Code Editor

Most editors bolt agents onto a file-centric model. OpenSquirrel flips it: agents are the primary unit, files are secondary. Written in Rust with the same GPUI framework that powers Zed, already supports Claude Code and Cursor. Karpathy asked for something like this and someone built it in a weekend. What does an editor look like when you’re mostly directing agents, not editing files? Nobody’s answered that well yet. (source: @elliotarledge)

Anthropic Claude Certified Architect Exam

Anthropic launched a certification exam for Claude architects. The $100M partner network behind it says this is an enterprise play, not just a badge. If you’re the person at your company figuring out how to deploy Claude, this credential may matter sooner than you think. (source: @shannholmberg)

Honorable Mentions

  • Context7 CLI: Pulls current library docs into your AI context window in a token-efficient way. Stops Claude from confidently writing deprecated code.
  • llmfit: One command scans your hardware and ranks which of 497 models from 133 providers will actually fit. Run this before downloading a 70B model that won’t fit in your RAM.
  • Alibaba page-agent: Browser automation that skips Playwright, extensions, and screenshots. Works directly in the browser with natural language and any LLM. (GitHub)
  • GLM-OCR: 0.9B parameter model for parsing PDFs locally with state-of-the-art accuracy. Runs on under 1.5GB VRAM in LM Studio.
  • Defuddle: YouTube transcript extraction with timestamps, chapters, and speaker diarization. Also powers Obsidian Web Clipper’s Reader mode. MIT open source. (defuddle.md)
  • Mistral Small 4: 119B total params across 128 experts, 256k context, Apache 2.0. You can toggle reasoning on or off per request. Dropped the same day Mistral announced a partnership with NVIDIA to co-develop open-source models.
  • Codex subagents: Codex can now spin up parallel sub-agents mid-task. Claude Code already does this. The pattern is converging across the industry.
  • GPT-5.4 at 5T tokens/day: Hit 5 trillion tokens per day within one week of launch. That’s more than OpenAI’s entire API handled a year ago.
  • Grok TTS API: xAI shipped text-to-speech with five voices. Available now via API.
  • NVIDIA Vera Rubin Space-1: Jensen announced a chip for orbital data centers at GTC. Yes, data centers in space.
  • NVIDIA Newton 1.0: Robot simulation platform hit GA. Articulated mechanism sim and high-fidelity hand modeling.
  • Context Hub: Andrew Ng’s open CLI that feeds coding agents up-to-date API docs. He’s calling it “Stack Overflow for AI agents.”
  • Perplexity Computer + Comet: Computer can now drive the Comet browser directly, using your logged-in sessions with permission.
  • Google Maps in AI Studio: Gemini APIs are getting native Maps data access. Spotted in AI Studio, not officially announced yet.
  • Flipper One mechanics: Full 3D models for the enclosure, back plates, and modules are now open source.

Try This Weekend

For everyone:

  1. Defuddle.md: Paste any YouTube video URL and get a clean timestamped transcript. Try it on a long interview or lecture.
  2. Perplexity Computer (Pro users): Try the analytics aggregation demo. Give it access to 2-3 platforms and ask for a combined dashboard.
  3. Karpathy’s job exposure map: Find it at github.com/karpathy/jobs and look up your actual occupation category.

For developers:

  1. Chrome 146 + Browser Use CLI: Update Chrome, flip the MCP toggle in settings, and let Claude Code drive your actual browser session. Zero extension setup required.
  2. Manus Desktop: Download it and throw a messy folder of files at it. Real test of whether local agents are ready.
  3. llmfit: One command to know exactly which local models your machine can run.
  4. agent-browser auto-connect: agent-browser --auto-connect open <url> for instant Chrome connection in your agentic scripts.