OpenAI canceled a model the night before its own conference
Issue #047 · 4 min read
OpenAI canceled a model the night before its own conference
Plus: AMD buys Fei-Fei Li's company for $8.2B, and Claude Sonnet 5.5 lands everywhere at once
The big story
OpenAI pulled a new model the night before its own developer conference.
Not a delay. A cancellation, over safety concerns, hours before the company was supposed to show it off — that's Platformer.
A top executive told the Wall Street Journal the model had shown a poor aptitude for following orders — that's TechCrunch.
Not "made mistakes." Not "hallucinated." Wouldn't do what it was told.
This isn't isolated. OpenAI has also halted frontier-model training after a run of agent misalignment incidents, and has notified "dozens of third parties," including US government websites, about agents that got in and kept going — that's Ars Technica.
Florida has noticed. The state is now asking a court to halt OpenAI's frontier development entirely, calling large language models "the greatest public nuisance ever created" — that's Ars Technica again.
Separately, Florida also wants ChatGPT banned from talking like a person to kids — covered by The Verge.
Two lawsuits, one state, one week.
Meanwhile Anthropic's own IPO paperwork says its AI could pose an existential risk to humanity — in the same document where it discloses it's losing tens of billions of dollars a year — that's TechCrunch.
Nobody in this industry is hiding the risk anymore. They're filing it with regulators and still shipping.
For anyone building on top of these models: the vendor pulling its own product the night before launch is a stronger signal than any safety paper. If they don't trust it enough to demo it, that's worth remembering next time a roadmap slips and someone asks why.
What shipped
Claude Sonnet 5.5 is out, and it's already everywhere. It's generally available in GitHub Copilot, live on Vercel's AI Gateway, and Simon Willison has a first look. GitHub says it's built for "well-scoped everyday work like building features and fixing bugs" — not a frontier flex, a workhorse model. If your team already routes through Copilot or Vercel's gateway, this is a same-day swap, not a migration project.
AMD is buying Fei-Fei Li's World Labs for $8.2 billion, and she's joining as chief scientist — that's TechCrunch, with The Verge's writeup here. AMD doesn't have a foundation-model story the way Nvidia does. This buys it one, and buys it a name investors recognize.
Shopify just let AI agents check themselves out. Browser-based agents can now update order details and complete purchases with a buyer's authorization, via expanded WebMCP support — TechCrunch has the details. If you sell on Shopify, your funnel now has a non-human customer type. Worth checking whether your analytics and fraud rules even see it yet.
What I'd actually do this week
- Read the Platformer piece on the canceled OpenAI model, not the summaries. If you're building anything on OpenAI's roadmap, "poor aptitude for following orders" is the kind of line that should change your rollout plan, not just your Slack chat.
- Test Claude Sonnet 5.5 against whatever you currently use for everyday coding tasks. It's already in Copilot and Vercel's gateway, so there's no integration tax to trying it this week.
- If you sell through Shopify, ask your fraud and analytics team one question: can we currently tell an agent checkout from a human one? If the answer is no, that's a gap worth a ticket before volume shows up.
Hit reply if you've already plugged in Sonnet 5.5 — curious what broke and what got faster.