Our research engine ran 600+ experiments in two days and the dashboard stayed green the whole time. Almost every one was the same experiment. New Borrowed Iron post on how an optimizer quietly collapses into a groove, and the human catch that caught it fast.

"The first time you see an engineer build something in 45 minutes that would have taken a week a year ago, but then see it not ship for another 6 weeks, you will be radicalized." Jack Hanlon, who leads GenAI Media at Meta, wrote that. That gap is the subject of my new piece on Run Data Run. AI made building software the fast part. Everything wrapped around the build did not move: privacy review, security, procurement, legal redline, regulatory sign-off. In biopharma each gate runs serial, three weeks minimum. Six in a row is half a year before anything reaches a patient. The argument: stop shipping faster typewriters and point AI at the governance layer itself. One agent gathers the evidence, another grades it, a human signs off on the fifteen-minute judgment. The next decade of enterprise AI isn't about better models, it's about compressing that review layer. The catch is that the people who must approve this are the same people whose work it eats. Process was built to protect patients. When process is the thing blocking patient-helping AI from shipping, it stops protecting patients. It protects itself. In a patient-facing industry, the queue is not paperwork. The queue is the patient. Every week it stays uncompressed is a week charged to someone who never got a seat in the room. Full piece on Run Data Run.

NVIDIA just handed a small global-health startup I work with a node of eight H100 GPUs for two months. No invoice. The crew putting it to work is five people: me, the founder, and three high-school interns. In a few days that little crew brought the machine up from nothing, staged terabytes of retinal scans, trained a foundation model, watched the first one fall flat, and pivoted to a better one inside the same week. An autonomous research engine I built reads papers and runs experiments on the box overnight. Three AI agents work three lanes at once. The interns stood up their first code repository and now help run a supercomputer. The mission is what I keep thinking about. There are about 200,000 eye specialists on Earth and well over a billion people who need screening. SocialEyes is building AI that reads systemic disease from a photo of the back of your eye, so a health worker with a camera can catch a high-risk case where there is no specialist for a hundred miles. It is grounded in clinics in Nepal. The pace is not longer hours. It is the harness doing the grunt work while the five of us do the deciding. I wrote the whole story on Run Data Run, and used it to properly introduce my technical sister blog, AIXplore, where the engineering build log lives. It is, I have decided, how a sabbatical should go.

The new AIXplore is live. I rebuilt it from the ground up into a lab notebook rather than a blog: interactive widgets you can drive, Tufte-style margin notes for citations and asides, and side quests that open the deep technical detours without cluttering the main read. Run Data Run, my main blog, is still alive and kicking. AIXplore has always been its sister for the more technical builder audience, and now it has the interface to match. The easiest way to see it is the latest article, my builder's read on OpenAI's new GPT-5.6. Drag the cost widget to find where owning your inference beats paying per token, open the side quests on the cyber fine print and the two-week governance timeline, and watch the margin notes fill in the sources as you scroll. The article is the content. How it is built is what I want you to see. It is also home to a series I am proud of and excited about: Borrowed Iron, a build-log of turning a borrowed 8xH100 node into a working research lab, written one engineering session at a time. More on that early next week. Come click around: