We’ve been enjoying Radar, working to make it better each week. Here are seven articles from the past two weeks that all converge on the same concept. Every one describes a team that reached for the model, the prompt, or the newest tool, and found the thing that decided the outcome sitting underneath it.
Here's how they stack up for anyone running a design team right now.
Start With What's Under Your AI Features
Two of these are the same argument from opposite ends of the stack. One says the evidence reaching your model is wrong. The other says that's because nobody structured the content in the first place.
Your Model Is Fine. Your Data Plumbing Is Lying to You. makes the case that most hallucinations are retrieval failures. Your dashboard proves the pipeline ran. That's a different claim from proving the truth survived the trip.
Your AI Bug Is a Design Bug Wearing a Model Costume finishes the thought. Eighteen frontier models all got worse as input grew, so the comfort move of stuffing the context window is backfiring. Taxonomy and metadata stopped being hygiene. They're the ranking logic your AI feature runs on.
Then Look at What Your Team Builds With
The tool layer is collapsing and the output is arriving as code. Both of these land on the same warning.
Your Design Stack Is About to Split covers the overlap between Claude Design, Figma Make, and Gamma. If leadership doesn't draw the lines, the team draws them badly, one project at a time.
Your design file wants to be the product, not a picture of it shows three builders arriving at the same guardrail. The model composes. The validated component kit carries the load. Put those two together and a published design system looks less like a consistency project and more like infrastructure.
Then Ask Who's On The Other End
The next two attack the user in your spec. One says that user might be a machine now. The other says the human version was never real to begin with.
The Interface Is Moving. Your Job Follows It. is about what happens when the assistant becomes the front door. Your product does the work and stays invisible, and time in app falls even when you're winning.
The user you designed for was never real goes back to a 1950 study that measured 4,000 airmen looking for the average pilot. Nobody matched. The block on fixing this is almost never missing data. It's what the fix costs to admit.
Know What's Standing In Front Of All Of It
Your team already cracked AI. You're the thing in the way. is the gate on everything above. MIT puts corporate AI failure at 95% and pins it on the companies. Leaders underestimate their own teams' usage by a factor of three.
You can fix the data, publish the system, and redefine the user, and none of it moves if the person who already figured it out is still waiting on an approval. The line worth sitting with is the other one in there. Plausible design looks coherent in a review, uses the right components, and ships because nobody objects.
