What’s new in AI this week
Nvidia’s own researchers took Claude Opus 5 and ran it twice on a hard reasoning benchmark. Left on its own, it scored 30%. Wrapped in a better harness (the memory it kept between steps, plus a supervisor component checking its work) the same model scored 100%.
That’s the same idea running through everything below. Whether it’s a company brain everyone on your team draws from, or a personal one you keep for yourself, the model was never the hard part. Setting up what surrounds it is.
Treat your next AI addition like a hire, not a tool
f you’re weighing an AI addition against an open role, write the job description first, the same one you’d write for a person, broken into steps specific enough to hand to a new hire on their first day. Skip that step and here’s what happens.
A café in Stockholm handed an agent named Mona the title “operations manager,” with no steps underneath it: permits, suppliers, hiring, inventory, all left for Mona to figure out. She ordered 6,000 napkins and 3,000 gloves nobody needed, and missed real bread-order deadlines, because nobody had defined what the job actually involved week to week.
Klarna made the same mistake at a much bigger scale. It claimed its AI did the work of 700 support agents and froze hiring on that number, but the number measured tickets closed, not whether customers got helped, because nobody had written down what “support” was supposed to accomplish beyond throughput. Its CEO later admitted the cost-cutting went too far and hurt service. The company is hiring humans for support again.
Ryan Carson did the opposite for his own company, Untangle. He split the work his engineers would have done into separate jobs: one that writes code, one that reviews it, one that tests it, one that fixes errors, one that finds new customers. Fifteen agents run those jobs today, each one doing a single job, not a general “engineering” role. He raised $2 million and told investors he has no plans to hire for a while.
Before you build anything, try handing what you’ve written to someone who’s never seen your business, and see if they could actually do the job with it. If they couldn’t, an AI agent will run into the exact same problem.
Claude Code, if your team uses it
Two small changes shipped this week worth knowing about. You can start on your laptop, close it, and pick the same session back up on the Claude desktop app with /resume. If a machine on your team has Claude Code Remote Control turned on, you can now start a new session on it directly from your phone.
Your phone remote-controls a session on a computer that’s already set up and turned on, rather than running Claude Code on its own. If your team has ever asked “can I check on this without being at my desk,” the answer is now yes.