Modern manufacturing has often ignored lessons that internet companies and children have known for a long time. For the moment let us discuss these self-contained systems of anomalies in regards to our children. In many ways a child’s brain works extremely well, without many of the common cognitive biases that we gain as adults. Without those biases they expose themselves to dangers, but also learn at a rate faster than adults do most of their lives. The brain goes through a fifteen year cycle where neurons are shed and pruned. Anomalies that are discovered in childhood become the fixed ideas of adulthood. I am in no way a developmental psychologist, but I am a technologist who has spent a large portion of my life in the lab, inventing with others, as well as inside factories, where we have advanced machines to make anomaly tracking something robust, and more importantly, actionable. Like the child who learns through consequence by attempts to put together modular train sets, as my son often does, iterations in production processing should be the tracks that eventually create a completed circle. Using computational microscopy for detection and classification combined with AI Feedback and Feed Forward mechanisms we can begin to close the loop. This is not often fully utilized.
Factories, unlike the developing brain, tend not to prune the process, reducing it to its core productive capabilities. Instead they tend to gather more data about anomalies, adding more information, in turn making the production and inspection systems of the factory larger. It is often said that the human brain is the most complex known “thing” in the universe. While this may be true, it is the seeking of self-sufficiency and simplicity that allow humans to prosper and engage in those activities that we all value like science, art, and literature. The equivalent would be to utilize the hearing, feeling, smelling, and perhaps most importantly, seeing senses to create sustainable and iterative processes. Akin to what we hope to be seeing with additive manufacturing. This could be thought of as optimizing certain anomalies for innovation and productivity, and allowing others able to be pruned. There are now ways of taking advantage of this knowledge, which will come as no surprise to you. Even the ways that we speak about our most powerful algorithms are brain based. We use machine learning, deep learning, reinforcement learning, all of which require neural networks. Factories exist much less as a network than these newest advances in artificial intelligence. They have essentially been an assembly line of serial processes and rarely have parallel systems thinking.
To put all of this into perspective we can go back to a revolutionary idea that Deming introduced in Japan after World War II. By using Statistical Process Control (SPC) rather than the more expensive and wasteful tradition of Pass/Fail Quality Control he helped generate the abundance of 20th century industry. That is, by measuring a relevant sampling of parts in production to see variability, a process can be adjusted. The reasons for adjusting the process could be any number of things, but we can consider any spikes over a certain standard deviation in process as anomalies, otherwise there would be no statistical jitter at all. By implementing SPC on as many layers of the process as possible we are able to zoom in and find the step in the process where anomalies are occurring. Essentially Deming was able to put in place a system where Quality Control was less relevant because the control of the process became more relevant. Japan was not the only country to benefit from this. Silicon Valley was built using many of these principles. Computers allowed for collection of many more statistically relevant data points, and luxury products became all the more commoditized.
When automated SPC and the human ambition to innovate came together with Gordon Moore’s famous prediction, it may have seemed like human manufacturing had reached the pinnacle of perfection. Every year companies pushed themselves for better SPC, and also for continuous improvement so that the “Law” would be self-fulfilling. The desire for continuous improvement and these advancements had very little to do with actual new ways of controlling a process. This is most apparent when you contrast Moore’s Law with another important predication that came from an Intel founding father, Arthur Rock. In short Rock’s Law is the tradeoff for making Moore’s Law possible. Rock observed that as transistor packing went smaller and price cheaper, a fab to build these devices would get larger and more expensive. This is exactly what happened. So the smart phone in your hand has been continuously improved, but the factories that make them are ever more complex. A factory is the opposite of the child that optimizes by learning to be more efficient through memories by pruning. A child becomes a hyper parallelized system with as much of the Goldilocks number of neurons as possible to continuously improve. With our new tool box of Artificial Intelligence it is now possible to make a factory learn like a child.
This idea of continuous improvement may be what is at the heart of a growing debate about our place in history and whether we are at an inflection point of technological progress, or whether the presence of anomalies themselves will lead to stagnation. It may be that people outside of the near perfectly clean fabs, where nanometer size defects and particles destroy a process, never think of potential stagnation as anomaly issues. Whether we are discussing a semiconductor fab, or turbine factory, or tire factory, how we deal with anomalies is the key to how we continuously improve. Moore’s Law is recognized as something that will come to an end, but it will only come to an end if we assume the Rock’s Law will remain true. There is a physical limit to the processes that are used for the most complex microchips available, so there is naturally a limit to Rock’s Law. It will have to plateau because there will be no reason to spend more on a product that cannot improve from spending more. The challenge and possibility of current computational technique that include not only AI, but also Super-resolution microscopy and high precision motion control make it essential to prune the fab and examine anomalies to see where opportunities exist.
This was the original plan for Nanotechnology and Atomically Precise Manufacturing that Eric Drexler has been suggesting for the last 30 years. I think that Drexler is correct, but there are many current examples where we do not have to be atomically precise in our manufacturing response to anomalies that will lead to continuous improvements. These are not only dramatic in capabilities but transformative in creating processes that reduce price with improvement rather than increase it. In industrial Additive Manufacturing, SPC is not even truly happening yet. In general, parts are printed and pass or fail at that point. This is mostly due to how new 3D printers are. The control systems that are in a fab are not yet in a 3D printer, but they will be. Adding computerized SPC alone will be an improvement. It will be possible to pause a process when an anomaly is detected. There can be a machine learning component to this, but it is hard to imagine that with this alone continuous improvement will occur. It seems counterintuitive to say that making anomalies our friend is a possibility, but it is not something that is foreign to evolutionary biology, either in the form of mutation, childhood development in the form of learning how to make train sets, or in an AI learning to win at the game GO. We can imagine that instead of stopping a process with an anomaly, the system, with a memory of a past completed design and a goal for an improved design, will use this anomaly to iterate. The first part made with an anomaly will likely not be better than an anomaly free part, but with time a strategy for optimizing with anomalies will emerge. What matters is not the design of the final product, it is how it functions.
We call this AIPC rather than SPC. Like humans an AI is capable of optimizing for an ideal outcome, not only for statistical control. With AIPC, you are limited only by your production/computational toolbox, and most importantly the learned information from previous iterations. Really though it is a perspective that will also be led by a human push for continuous improvement. Humans will not be finding the solution by spending more money and adding more things, but by analyzing and thinking as a system. In essence we will be making our factories children again.