Weekly Links: AI Marketing Disasters, Anthropic IPO, and Tips for Agentic Engineering

This week, Agents get therapists, Supabase hits 10.5B in valuation, and Amazon product search begins to invent products you can't buy (yes, it's deliberate). Finally, Anthropic suggests the world pause AI innovation, but in the same week, raises $65B and files for an IPO.

On to the main stories this week:

• How a Starbucks marketing stunt spiraled into mass boycotts in South Korea. This is a wild story, and the key to it all: "Marketers chose the slogan after consulting an AI tool, looking for suggestions, Shinsegae Group said. It turned out some managers who approved the campaign never opened the email attachments showing the marketing material." Yes, the campaign was "relevant" to the date, but not in the way you would have hoped.
• The Newest Instagram "Exploit" is the Goofiest I've Seen. Not to be left out, Meta had a serious incident this week with a new AI system for handling user account resets. The security situation is laughable, and the exploit still is not properly fixed. In fact, the fix itself smacks of AI coding (the button for triggering account help has been removed, but API access for the same thing remains). It's scary that even the biggest companies seemingly don't know how to triage functions that are critical and security-related from run-of-the-mill functionality.
• How AI has de-skilled translation. This week's column rapidly risks turning into a full-on "AI is terrible" edition, but I wanted to include this because it's a clear illustration of how some job roles are being eliminated. In translation, in particular, first Google Translate and now LLMs have gotten so good that human adjustment is barely needed for some types of work. Increasingly, translation jobs are becoming "fix what the AI missed," and understandably, the most skilled translators are choosing to step out of the market altogether rather than do this.
• NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI. Accelerating the trend for all-in-one models that bake in an understanding of the world, NVIDIA's new open-source model combines a physical world model with an LLM and a framework for action-taking. This is the current path many robotics companies are betting on to enable physical AI. It's interesting to see NVIDIA so engaged in this, but logical since physical AI may end up being even bigger than data center AI.
• BDD, ADR, PRD, WTF: Capturing Decisions for Humans and AI Alike — Michal Cichra. My Safe Intelligence colleague has a great talk up on the AI.Engineer YouTube that is getting a lot of love. The key point: if you're going to use Agents to code, you need to rethink how you communicate boundaries to them. Definitely worth checking out!

Lastly, this week: if you're a LEGO fan, you might love this - a 12,060-piece set of the Sagrada Família in Barcelona.

Thank you to various colleagues for surfacing great links this week!