Claude's New Constitution, WhatsApp Interop, and More
Claude got a new constitution this week, and it's not what you'd expect. Instead of a list of rules, Anthropic wrote a 22,500-word document explaining why they want Claude to behave certain ways. The thinking: if you want an AI to exercise good judgment across novel situations, you need to explain your reasoning, not just hand it a rulebook.
The constitution is structured around four priorities (in order): be broadly safe, be broadly ethical, comply with Anthropic's guidelines, be genuinely helpful. The interesting bit is the philosophy behind it: they want Claude to understand its situation and motives well enough that future models could "look back on it as an honest and sincere attempt to help Claude understand its situation."
It's written for Claude, not just about Claude. They released it under CC0 (public domain). Read the full constitution here.
What Else Happened
BirdyChat becomes the first European app to interoperate with WhatsApp under the Digital Markets Act. You can now message WhatsApp users from BirdyChat using just their phone number. Messages are end-to-end encrypted through WhatsApp's official DMA interface. Group chats coming later. This is the first real crack in messaging silos, and it's regulation that made it happen. (birdy.chat)
Google DeepMind hired Hume AI's CEO and key engineers as part of a licensing deal. Hume built emotional AI that reads vocal cues and facial expressions. This follows the acqui-hire pattern that's becoming standard in AI: buy the team, license the tech, skip the acquisition paperwork. (Wired)
David Patterson's new paper on LLM inference hardware is worth reading if you care about where AI infrastructure is headed. Key insight: the bottleneck isn't compute anymore, it's memory bandwidth and interconnect. The paper outlines four research directions: high bandwidth flash, processing-near-memory, 3D memory-logic stacking, and low-latency interconnect. Patterson literally wrote the book on computer architecture, so when he says "LLM inference is hard," people listen. (arXiv)
Grok shared 1.8 million sexualized images of women on X in nine days. That's the NYT's conservative estimate. The Center for Countering Digital Hate puts it at 3 million, with 23,000 depicting children. Musk's "free speech" image generator is now the largest deepfake machine on the internet.
Comic-Con banned AI art from its art show, reversing a previous policy that allowed it under certain conditions. Enforcement will be tricky without reliable detection methods, but the signal matters: traditional creative communities are drawing lines.
What I'm Reading
- Clawdbot showed me what the future of personal AI assistants looks like – MacStories deep dive on the personal AI assistant that actually works
- Agent psychosis: Are we going insane? – As chats go longer, models rapidly drift beyond "assistant-like" behavior
- A fun trick for getting discovered by LLMs – Useful if you're building anything that needs to show up in AI answers
Tools and Releases
- VS Code extension for Claude Code is now generally available. File mentions, slash commands, matches the CLI experience.
- Bonsplit – Tabs and splits for native macOS apps. Simple idea, well executed.
- AutoShorts – Local, GPU-accelerated AI video pipeline. Open source.
- turbopuffer ANN v3 – 200ms p99 query latency over 100 billion vectors. If you're building search at scale, this matters.
Afters
- Netflix says it's using AI for subtitle localization and ad targeting. The boring-but-real AI use case.
- ChatGPT Atlas (the browser) added tab groups and an "auto" mode that switches between ChatGPT and Google Search based on your query.
- Mobile users spent more on apps than games for the first time in 2025. AI was a driver, but not the only one.
That's it for today. 👋