In all the hype about AI (artificial intelligence) I think we may have wandered astray.

These days when I get into these higher level discussions of automation or quality, people will bring up that they are implementing “AI” across the organization as part of their digital transformation, or industry 4.0 or something or other. AI is often the first thing they say. To me, this is like saying “hammer” or “saw” or “skillet.” “My strategy for 2025 is to use a hammer.” Why hammer? Are we using nails to attach something? Should we use a screw instead? Is it a picture? What does it weigh? If we are using nails, a nail-gun will nail it faster. Why use a skillet? Obviously then we are talking about cooking something. Is it dinner?

Is the end goal to use AI for the sake of using AI, or to use it as a tool to accomplish something specific?

They are then sometimes surprised to learn that they already use AI, to solve part of a particular problem.

NVIDIA, which makes the chips that process data fastest to render AI decisions, is now the world’s most valuable company. They are worth $3.28 trillion (as of end of 2024). Amazing applications have come and will continue to come from the chips they make.

An Application Example

Back in 2018, we used a NVIDIA GPU to evaluate welds on a surgical instrument. It worked really well. Traditional machine vision methods hadn’t been able to do it, besides to catch situations where the weld was entirely missing. But entirely missing was only part of the real world problem. Incomplete welds were very dangerous to the patient, as well as uneven welds. The tangible reality of the situation suggested the use of a better tool. That tool was TensorFlow, an open source deep learning tool from Google, and an NVIDIA GPU (although a CPU can process it too). The result was a neat machine vision system that worked well at solving a real world problem.

The application actually hinged on making sure all parts passed through the system, so after really studying the line, a very compact system with an interlock got put on the station where the handle was added to the part (at the other end). This ensured that every tool which got a handle had a good weld and allowed the inspection to occur out of the way of the assembler without adding a step to the process or extra time to a step. In terms of the overall economics of the production line, this was the way to do it, so we used very short working distance s-mount lenses and a custom Delrin fixture on a machined plate.

We used the GPU to train the model, because it’s massively faster for training models than a CPU, due to 100s and 1000s of simpler parallel cores, rather than more capable 1-32 cores a CPU might have. GPUs, graphical processing units, grew out of all the parallel computations needed to compute every pixel on the computer monitor. So it’s basic math, but tons of it at once, to generate video for playing video games. By morning we could see the model converged and it was going to work. That was the critical moment.

To run the model, or “run inference” as AI engineers would say, it took single digit milliseconds on a GPU, and 1-2s on the CPU. But as the key part in the process of assembling the handle took about 20s, there was no point. Running the model on the GPU would have been more hardware, lots of power consumption (as the world is finding out while running these GPUs to mine bitcoin and do “AI”), another point of failure, and more heat which would make all the other electronics fail sooner. So we took it out.

AI has gone into lots of vision systems we’ve built. It’s amazing, for example, for dealing with anything that generates stray reflections (metal and glass). But it should, in my opinion, be farther down in the chain of decisions to be made.

For instance, for another customer making a pharmaceutical product, the rim of the glass vial has to be as perfect as possible so the product can be sealed properly. However, they have to use Borosilicate glass (Pyrex effectively) instead of soda-lime glass (every bottle at the grocery store). Borosilicate glass is more challenging to make bottles out of. Two halves need to be fused because it doesn’t have the right viscosity to blow mold it.

Getting a perfect rim in a fused product is tough. It has a weld seam going up both sides where the two halves go together, and therefore two weld seams cross this rim. Ouch.

Ideally the solution, should have been a series of solutions that used what are now called “traditional machine vision” approaches. Why?

It would be ideal as much as possible to measure each weld seam as it crossed that rim in order to be able to say that one weld seam was bigger than the target of 100µm. Or that both were bigger. Or that the weld was misaligned. It gives the person running the process the information needed to make the right adjustment to the process. Which knob to turn? As anyone who’s been to any factory knows, there isn’t just one knob. And making quality parts involves controlling the process and understanding the relationships between the physical reality of what you make and the exact set up of all the knobs.

It could be both halves are perfect and are not getting properly aligned to fuse them. It could be that they are too cold when fused. It could be that one half is a “short shot” (molded with too little glass).

Unfortunately, getting this information out of clear glass is tough, and really tough to do at 500ms a part. And tough to do in one “final inspection”, rather than a series of upstream inspections for specific problems. We had to turn to AI. We also had to buy an enormous GPU from NVIDIA. There was no other way.

Again, the GPU worked remarkably well. The system never let through a critical closure defect in any of our tests or the customer’s. But no one could really prove it worked. It didn’t really give you a knob to turn. It scrapped a decent amount of rims that I would just describe as odd-looking but not actually out of any applicable tolerance. The AI inspection had created a cosmetic tolerance that was not part of the set of required tolerances.

While AI was the fastest way to get to the macro solution, a micro-level approach would have guided everyone faster to an understanding of the knobs involved in the process.

The more micro approach also maps up better to the end performance needed: a good seal. Is AI really “thinking” about that, or is it regurgitating what its human handlers have trained it to do? What would it do for other tangible scenarios that impacted seal quality? Could I say with certainty it understood the underlying concept of what we wanted. Maybe. A complete weld leaves behind more visual evidence, and I’m confident that algorithm was doing what we wanted for the other application. A good seal is more speculative about what will happen when mated with the lid and a more geometric question.

In the end, AI itself is not a solution, it is a tool, and you really have to think through whether and where it’s the best tool for the job.

Is NVIDIA worth $3.28 trillion in terms of its potential to human advancement? Maybe. Boeing is worth $132 billion (Dec 31, 2024). Airbus is ironically worth $122 billion Euros (Dec 31, 2024). So about the exact same at current exchange rates after 100+ years of competition. I quote these numbers because a plane is a very useful tool and a very expensive tool. It allows travel without highway or rail infrastructure in between the two points. All that’s needed is a 3km runway in Denver and a 3km runway in Frankfurt or Tokyo, one of those tools (called planes), jet fuel and a pilot. If tomorrow people or things need to go to Mexico City, no physical infrastructure changes. That’s incredible.

The market is saying AI will solve things that are at least 10X more valuable than that. Maybe it will. I’m not an AI skeptic, but see it as a tool. How can it contribute to a specific solution for a real-world problem? And even then, what type of AI will win, and what chip-type will it be computed on?

For applications in quality, I think we have to be careful that we aren’t using it to short circuit our understanding of the problems, how those problems relate to each other, and how all of it relates to building better product.