Not every gap is a prediction problem
A vague sense that something should be smarter, and a model gets proposed before anyone has asked what kind of problem this actually is. Some problems are prediction problems, genuinely suited to a model from the start.
Dyson released a launch video walking through how their $500 CameraJet™ electric toothbrush came together, and the order surprised me.
It didn't start with a camera or a model. It started with a jet. They built it, made it faster, then faster again, until it was firing a burst in about a millisecond. It worked. Then they hit a wall that speed couldn't fix: the jet had nothing telling it where to aim. Some gaps between teeth were never getting hit.
The obvious next move, if you're building a product in 2026, is to reach for AI. Point a model at the problem, let it find the gaps, done. That's not what they did.
They added a camera first. A plain sensor. And it didn't work either, because the brush head vibrates something like a thousand times a second, and every frame came out smeared. So they solved that with more engineering: a higher frame rate, an LED strobe timed against the vibration, until the image held steady. No model involved. Just better instrumentation, tuned until the thing they were measuring stopped lying to them.
Only after that, jet built, camera built, image finally trustworthy, did they ask the question that gets a model built: what if the camera still misses one? Its job is narrower than seeing or aiming. It's there to catch the cases where seeing and aiming, done as well as physics and optics allow, still aren't enough.
Whatever that $500 is actually paying for, the build order suggests it isn't mostly the model. It's a jet with a millisecond trigger and a camera stable enough to aim it. The model is the layer that shows up last, does the narrowest job, and gets the most attention anyway. That's where the real attention is misplaced.
I keep watching product teams do the opposite of this, before they've even shipped. A vague sense that something should be smarter, and a model gets proposed before anyone has asked what kind of problem this actually is. Some problems are prediction problems, genuinely suited to a model from the start. Others are measurement problems, or mechanism problems, or aiming problems, and a model bolted onto one of those just gets asked to compensate for something better engineering would have solved for free. It compensates badly, because it's often the least reliable brick in the whole structure. Then it gets top billing anyway, because "AI" is what gets funded and what gets covered.
Timing was never really the point. What matters is asking what kind of problem you actually have before deciding a model belongs in it at all. A jet needs to fire faster. A camera needs to hold still, neither is a prediction problem, and no model fixes either one. Find the actual gap first, then match the tool to it, not to what's fashionable to announce.
