Generating Intelligence

“By far, the greatest danger of Artificial Intelligence is that people conclude too early that they understand it.” — Eliezer Yudkowsky

Last year I wrote about the hype –– and challenges –– of generative AI. I was, and remain, dubious that it will ever live up to the breathless optimism that accompanied last year's splashy launches. But I may have moderated my position a little as a few practical use-cases have started to emerge from the noise.

Generative AI still produces problematic results, still can't tell when its hallucinating, and as we're learning, turns out to have enormous infrastructure and energy impact. But as I said in my previous post, we can't put this genie back in the bottle –– and we shouldn't try. As with previous innovations in computation, our understanding of AI will evolve until today's magic seems commonplace. As in the past, we'll likely decide to rescind the "AI" label entirely, and reserve it for some future magical innovation –– but only if we ride out this hype cycle and keep pushing for the next summit.

In last year's post, I urged caution in adoption and integration of generative AI. On an impromptu panel at the ARC conference earlier this year, I held that line –– but with a hint of optimism added. There are valid applications of Gen AI, and promising potential for continued improvement of cognitive-style computing, but we should continue to proceed thoughtfully. While in the past year, I've never managed to coerce ChatGPT into producing a complete and useful novel work, I have successfully used it to summarize existing work. While I'm nowhere near trusting it to replace a software developer, its ability to regurgitate Stack Overflow, and (potentially plageriously) playback related code from GitHub, can accelerate monotonous or obscure development tasks. And although they aren't a part of our current Gen AI hype, machine learning models that quickly classify, estimate, or leverage first-principles to simulate or predict, have long since proved their worth in the toolbox of manufacturing problems –– and data science in general.

Much prognostication has been done by others about the societal impact of Gen AI. Some have even compared it to the iPhone as the dawn of a new era. I'd argue that this moment in computing has the potential to be even more important: the iPhone, in its elegance, made us all dumber. Its glossy, fluid UI, and closed-sandbox model abstracted users from the challenges computing presented in the past. If I hadn't taught them, it would have been entirely possible for my kids to make it through high school without understanding what a file system is. The iPhone turned our most personal computer into an appliance that only gets cracked open by specially trained service people. Apple continues this trend with more hardware that can't be repaired, an Operating System so locked down that you can no longer drag pre-installed apps to the trash, and, with each release of the OS, additional "security" restrictions making the platform more locked down. If any part of the system breaks, the expectation is customers will just throw it out and buy a new one...

We have an opportunity with this AI moment to do much better. Instead of pretending these algorithms are magical and impossible to decipher, we can leverage them to learn. Microsoft's new Copilot keyboard key attempts to obscure the complexity of AI behind a familiar user interface –– as they are doing with so many of their software products. Are we properly equipped to understand what happens when Copilot is invoked? Does a simple key press properly communicate the massive amount of computation and energy consumed to train the model that generates a resume in Word, or summarizes a Teams meeting??

What if, instead of teaching students a new key on the keyboard of a computer they don't understand, we taught them how statistics, the laws of physics, and data come together with the flow of electrons, the manipulation of logic gates, and the laws of mathematics to produce results they can understand. In fact, if there's a danger to Generative AI, it’s that we currently cannot explain it simply. As Einstein once said: "all physical theories, their mathematical expressions apart, ought to lend themselves to so simple a description that even a child could understand them."

In manufacturing, we have a prerequisite problem we need to tackle first: the data we might feed to an algorithm to sort, classify, or regurgitate, lacks the structure even of the morass that is the Internet. ChatGPT benefits from a corpus made up of common languages, predictable semantics, and a simple but critical technology called the hyperlink, that articulates relationships between ideas, documents, words, and images in a programmatically discoverable fashion. The very fabric of the World Wide Web lends itself to machine learning. Yet the common fabric of manufacturing data is, at its very best, arbitrarily named key–value pairs. We have no semantic consistency, we have no relationship lines to follow (apart from those trapped within proprietary systems), and thus we have no near-magical innovation in AI to exploit. And while most system integrators can at least claim competence in understanding logic and the manipulation of electrical signals, our industry's best attempts to make sense of our data are social media buzzwords, like the imperfect "Unified Namespace" concept (which is neither unified, nor a properly qualified namespace.)

As a technologically dependent society, we have a responsibility to demand accountability, transparency, and openness from the vendors (new and old) offering to change our world with their new computational wizardry. Our governments have the dual responsibility to encourage and support innovation, while working with both its producers and its consumers, to erect guardrails and a regulatory framework that ensure the general population isn't preyed upon by this new technology era (as we have been by the smart phone ecosystem). And within manufacturing, we should see this moment as a call to lift our ecosystem out of its 1990s–level stunted growth, demand better from the vendors, integrators and innovators in our space, and finally position manufacturing technology in its rightful place as a creator of real value.

We're nearing the peak of this wave of AI in the hype cycle. In the valley of disillusionment, durable and truly transformative use-cases will be found. When that moment comes, some companies and offerings will disappear, and the ones with staying power will bob back to the surface as survivors. We can start voting now for who we want those remaining players to be: invest in those who are open about how their technology works and willing to equip the next generation workforce to help them understand and responsibly use and evolve that technology; invest in those who are transparent with regulators and consortia about the risks and demands that come along with that evolution. And if you're in manufacturing, invest in an ecosystem that embraces standards and open interfaces, that gives us the best chance of evolving along with the rest of the world.

For the past 6+ years, CESMII has been working with standards bodies, other consortia, manufacturers, and the vendor and integrator ecosystem that support them, to provide mechanisms to bring shape, structure, relationship, and reliable semantics to fully qualified namespaces that we call Smart Manufacturing Profiles. Our research projects explore how AI can take advantage of this improved foundation for manufacturing information to drive new insights. Our innovations are open, and members include some of key thought leaders in this space. Consider this your invitation to join us in pushing for a better future.