A researcher asks a question that sounds simple: "Which AI model should I use for my DNA variant, BOLT-LMM or DNABERT?" It's a trap, and it's the most common mistake I watched teams make over years in genomics. Those two tools don't belong in the same conversation. One scans a whole population to find which genetic differences track with a trait. The other reads a single stretch of DNA and turns it into something a computer can work with. Asking which is "better" for one variant is like asking whether a microscope or a telephone is better for measuring temperature. Here's the part that should bother anyone who relies on benchmarks. "Genomic AI model" is one phrase covering at least nine genuinely different kinds of tool, each answering a different question. And the leaderboards everyone uses to choose a model can't catch the mistake, because a leaderboard ranks tools within a category. It assumes you already picked the right category. The error happens one step before the benchmark ever runs. Pick the second-best tool in the right category and you lose a little accuracy. Pick the wrong category entirely and you get a confident answer that means nothing, formatted and plotted and ready to drop into a slide. The fix is a habit, not a tool: before you ask which model is best, ask what kind of object your task actually needs. Are we even in the right column? I built a small open thing to make that concrete. But the lesson outlives the tool, and it's not only about genomics. The next time someone hands you a "which is best" question, the more useful question is usually hiding underneath it.

Busy week away from the usual headliners. Google shipped DiffusionGemma (text generation by diffusion, ~4x faster) and Gemma 4, Microsoft revealed its MAI frontier models, and Bezos raised $12B for an 'artificial general engineer.' All in today's wire.

the refusals are friction, i hit a few too. but the switch is less about model quality than whose harness your loop already lives in. codex won because the loop was already built around it. i spent a day with fable inside my own harness and it just disappeared into the work.

I built a finance app for myself in two nights. It works, it reads my real accounts, and it would help other people who want the same thing. It was never going to leave my laptop. Not because the idea is bad. Because making a private tool public means a weekend of dull, dangerous cleanup: hunting for secrets buried in old files, scrubbing your personal data out of the history, renaming everything, building a brand, writing a front page. One missed line and you've published your bank details to the internet forever. So you don't. The tool stays yours. Most useful private things die exactly this way. For scale: months ago I built its sibling, EmberPlan (emberplan.com), a FIRE-planning app I run for the financial-independence community. On the models of that moment it took maybe ten times the effort. This one took two nights. Yesterday Anthropic shipped Fable 5, a new tier of model. Today I handed it the whole app and asked for a clean public version. In one session it swept the history and found my real brokerage positions hiding in a forgotten test file. I would have shipped those by eye. It caught them, cleaned the project, redesigned it, named it, generated the brand, and open-sourced it. The leak it caught is the whole point. The work that kept this tool private wasn't hard, it was tedious and risky at the same time, which is the exact combination that makes "I'll do it later" the rational call. Take that away and a different question lands on the table: that useful thing you built for yourself and never shared, is it about to start helping people you'll never meet? For the first time the work in the way is small enough to actually do.

Apple just rented its brain. At WWDC this week, the company that owns every layer of its stack, the chip, the OS, the store, the cable, did the one thing you'd assume it never would. It built the new Siri on Google's Gemini, distilled it into its own models, and put it on a billion phones. The same week, the largest prediction market on which company has the best AI model gave Google an 8% chance. Anthropic sat at 90. Apple didn't rent the best model. It rented one that was good enough, available, and willing to sign. That's the strategy, not a slip. The model is becoming a commodity. A new frontier one lands every few months, and they get more interchangeable by the week. What you can't rent is everything around it: a privacy boundary outsiders can verify, the logic deciding what runs on the phone versus the cloud, and a billion devices people already carry. Apple rented the engine and built the car. For most people, this iPhone is where they'll meet real AI for the first time. Not a separate app they had to go find. The assistant in their hand, doing the thing they wanted done. The brain underneath will be a commodity they never see. The question I can't answer yet: is Apple's wrapper a real moat, or a beautiful interface over a rented model? On day one, the reporting can't even agree how much of this is Google. We find out the first time Google ships something Apple can't distill its way around. Rent the brain. Own the car. Just make sure it's one only you could build.

AI Is Building AI. Now Read the Footnotes.