Round two of a standing habit: when we read something clear-eyed about putting AI to real, public-minded work, we keep it. Five more, no hype, just things worth your time.
01 · The rise of intelligence ownership
The Fermisense team fine-tunes a 9-billion-parameter open model on one high-volume task and beats every frontier setup they tested: 87.3% accuracy to the frontier’s 76.9%, at somewhere between 40× and 340× less cost, by their math. The takeaway: owning a small model trained on your task and your data is now a real strategy. Start on the frontier, then graduate to a specialist you own.
02 · The open-weight models that matter
If you’re going to own a model, it helps to know the field. OpenRouter’s June survey of open-weight models is the most honest we’ve found. Four models, four different strengths: one for cost, one for reasoning, one for image and video, one built and hosted in the U.S. Useful precisely because it’s a map, not a leaderboard.
03 · More AI in federal agencies, results harder to see
Federal AI use cases grew from about 700 in 2023 to more than 3,600 by 2025. Half of them sit inside just five big agencies, and plenty are parked at the pilot stage. A clear-eyed interview on the gap between adopting AI and actually deploying it, which is where the real work (and the real payoff) lives.
04 · “AI-enabled” without the hype
Guidance for rural and community health providers that could have been written for Alaska: point AI at the unglamorous work first (documentation, message triage, prior authorization), keep a clinician’s eyes on everything it writes, and treat “convincing but incorrect” as the failure mode to design against. Our kind of title, too.
05 · In Alaska: a data center inside a power plant, inside a microgrid
Cordova tucked a 170-kilowatt modular data center inside its hydroelectric plant, soaking up the 20–25% of hydropower that used to spill past unused. Local compute on local power, in rural Alaska. The most quietly encouraging thing we read all summer.
Bonus · GUIdebook’s screenshot museum
One for fun. Marcin Wichary’s GUIdebook preserves what software used to look like, screenshot by screenshot, across the first graphical decades. Pair it with Decker, John Earnest’s living revival of HyperCard’s make-something spirit. AI assistants now help people port and bring back old software all the time, and work like that starts in archives and revivals like these.
Reading something we should be? Send it our way through the contact form. And if you’d like a note when the next list lands, the subscribe form on any page does the trick.
— Tyler Arnold
All posts