The Sensor That Deleted the Model

The best thing I read this week is an argument against the thing I sell, and I think it's right.

A group at Queen Mary University of London, working with Florence, Trieste, and Trento, built a fingertip that can feel the raised lettering on a penny. A hundred-micrometer resolution. A tenth of a millimeter. It maps depth and size in real units, not just relative bumps, which is more than most soft tactile sensors will give you.

Here is the part I have not stopped thinking about.

100-micrometer resolution, with no computational latency.

Not low latency. Not optimized. None.

How You Are Supposed To Do This

The standard architecture for a soft touch sensor goes like this. Put a camera behind a deformable gel. Press the gel against something. The camera sees a distorted pattern. Then run a model that infers, from the distortion, what force was applied and where.

Four stages. Camera, deformation, inference, answer. The third stage is where all the engineering goes, and it's where all the trouble lives: the model has to be trained, it drifts when the gel ages, it needs recalibration, and when it produces a number nobody expects, somebody has to go figure out why.

The QMUL team deleted the third stage. Not optimized it. Removed it.

Their sensing layer is a Bragg reflector, made by exposing a light-sensitive film to a red laser until it sets up alternating bands of polymer density. Stretch that structure and it reflects a different wavelength. Which is to say: press on it and it changes color, and the color is the pressure.

The camera does not infer force from distortion. It reads a color. The physics of the material already did the math, at the speed of light, for free, in a layer of polymer.

They are honest about the catch, and I'll repeat it because it's the thing that would bite you: soft materials under repeated use are a durability question, and they have not validated it in a real robotic hand yet. This is a lab result. Fingertips get slammed into things ten thousand times a day.

I Have Been Apologizing For The Wrong Thing

I monitor equipment for small buildings. A sump pump in a basement, two years running now. Forty devices in a small commercial building.

The sump pump system has no sensor on the pump. There is a fifteen-dollar smart plug between the pump and the wall, and everything I know about that pump comes from how much current the motor draws and for how long.

I've written before about the fact that this is not the cheap version of a real system, it's the direction the whole field is heading. What I have not admitted in public is a smaller and more uncomfortable thing.

The parts of my system with a model in them are the parts I babysit. The parts that are just physics and a number have run for two years without me.

The alarm that has never once been wrong is not clever. It is: this pump has drawn current for longer than it has ever drawn current before. That's it. That's the rule. There's no inference in it, no training data, nothing to drift. It is a threshold on a physical quantity, and it is the single most valuable line of code I have running.

The pattern work sitting on top of it, the cycle-shape comparison, the slow-degradation trending, is genuinely useful and it is genuinely where my attention goes when something looks strange. That's the tradeoff, stated plainly: the smart layer costs ongoing attention, and the dumb layer does not.

I used to be a little sheepish about how much of my value came from the dumb layer. The QMUL result is the same insight at a much higher level of sophistication, published by people with lasers: when you know the machine, the best place to put the intelligence is as close to the physics as you can get it, and ideally inside the physics.

Why This Doesn't Mean The Robots Are Wrong

I want to be careful, because there's a lazy version of this argument that says all the AI stuff is hype and the old ways are better, and that's not what I think.

The same newsletter that pointed me at the tactile sensor also linked Rodney Brooks' piece on the four time scales of technology. It's the most useful framework I've read this year, and it explains the whole tension:

Look at the second row. Hype is the only one of the four measured in months. Brooks points out that "AI agents" went from obscure in mid-2025 to inescapable by this month. Meanwhile Linux shipped in 1991 and Microsoft adopted it in 2012. Self-driving cars were demonstrated in 1987.

In the same week, Google's newest robotics model posted somewhere between 45.7% and 76.3% on humanoid manipulation tasks. That is a real result and I don't want to be cute about it. It is also, read honestly, a coin flip to three-quarters on things a person does without thinking about it, which is precisely what a technology looks like early in a twenty-year deployment curve.

General-purpose embodied AI is real and it is coming and it is on the twenty-year clock. My customer's pump is on this month's clock. Those are different businesses, and confusing them is how you end up selling a research program to somebody who wanted a working basement.

The Owner Who Counts Closed Work Orders

One more thing landed this week that sharpened this for me, from a very different direction.

Kilroy Realty runs close to four thousand water sensors across eighty buildings. That's a serious deployment, roughly fifty points per building. When they talk about it, here is the number they lead with:

Whether the engineers actually closed the resulting work orders.

Not detection rate. Not sensor uptime. Not accuracy. Close rate.

They also, and this is the part I'm still chewing on, turned the same alert history into a record they bring to insurance negotiations. Same sensors. Same data I'm already collecting on every site I run. An entirely separate argument for what it's worth.

In the same roundup: Boston University started a controls retrofit at a track and tennis facility, chosen because the systems were simple and it was close to the shop, funded it with leftover money from a chiller replacement, and cut utilities twenty percent.

Simplest system. Closest building. Leftover budget. Twenty percent. That is not a technology story at all, and it's a better sales case than any technology story I could tell.

What I'm Doing About It

Last week I said I was going to compute my own lead time: for every catch on both sites, the days between the first anomaly in the data and the thing it predicted. I'm doing that. But this week convinced me it's only half the number worth publishing.

The other half is Kilroy's. Of the alerts I've raised, how many resulted in somebody actually doing something.

I don't know that number yet, and there's a real chance it's worse than I'd like. An alert that nobody acted on is not a detection, it's a notification, and I've been counting notifications. If the close rate is thin, that's a more useful thing to learn than another month of tuning the pattern matching, and I'll publish it either way.

Here's where I've landed, and it's the opposite of where I expected to be a year ago.

If you own a small building, the intelligence is not the product. A threshold on a current draw, checked every second for two years without ever needing attention, is a better thing to own than a model that's right more often and has to be maintained by somebody who isn't you. The London group put the intelligence in a sheet of polymer. I put mine in a fifteen-dollar plug and one comparison. Kilroy put theirs in a work order system.

All three of us made the same bet: put the smart part where it can't break, and spend the effort on whether anybody did anything about it.