This week was actually a little quieter in terms of news (I'll call it out since it happens so rarely!). Still, someone turned Lead into Gold, Claude Code got lazy tool loading, and the US administration gets into AI image generation (I guess that would clearly break Spain's new planned deepfake law).
I also connected up steampunk.ai to Commune, which is a cool new discussion community site started by Fran Mendez, (of Async API fame). It's a conversation space to promote discussion. You can check it out via the button below:
There are lots of other cool newsletters being added there as well.
Here are this week's main stories. I'll keep it to four posts this week:
- Claude's new constitution. Anthtopic, as far as we know, is the only major model provider to prominently use a set of core consistent principles as a driver for training. For their new models, they released the new version of the constitution they use. It's unknown whether grounding LLMs in this way really affects their behaviour deeply enough to prevent serious failures down the line, but I do think trying to be explicit about the values a model should embody is valuable. As Anthropic says in the post, it tries to express why a model should behave in a certain way, not just how to behave.
- Brex CTO James Reggio on the AI Engineer Podcast. This is a great snapshot in time for what large tech-first (but non-AI) companies are doing with AI adoption. Brex is no doubt more tech-forward than most large companies, but the patterns are very relevant: supporting employee adoption, looking at product reinvention, and trying to thread the needle between buying AI tech in and building from scratch.
- Welcome to Gas Town. Hat tip to Michal C for leading me to this. This post by developer Steve Ygge has gone viral in engineering communities as the current pinnacle of "AI Agents gone crazy," but it also represents a glimpse of the future. The post walks through how to harness many (Claude Code and Codex-like) coding agents to work together in a structure called an orchestrator. The post details a complex multi-agent system with deliberate roles, linkages, and underlying tools (such as issue tracking systems for agent communication). Running a system like this clearly burns a huge amount of AI tokens, and the code generated is unreviewable. Yet... It's a compelling vision of the kind of code engine room we might have with more AI progress.
The image from the blog post says it all:
- Weaponizing Calendar Invites: A Semantic Attack on Google Gemini. Researchers found a clear vulnerability in Google Calendar invites that allowed someone to send a calendar invite that tricked Gemini into creating an invite containing listings of personal meetings. No doubt the precise hole will get close, but the nature of the attack is hard to shield against: human-written text that takes a powerful LLM to decode. The key challenge is that the firewall needs to be smarter than the agent on the other side.
Wishing you a great weekend.
