Introduction
THE hot topic is AI, and I am not immune to jumping on the bandwagon to fuel the fire. I’m by no means a cheerleader, but I think it’s prudent to think about the technology like an engineer comparing its assets and vulnerabilities. To be honest, I’m approaching this subject from the perspective of a sceptic and detractor – a bias I will try to temper.
I’m not going to focus on whether it’s a good thing or a bad thing – the discussion has been had in many other forums. I’m also not going to focus on what jobs it may or may not replace – also well-trodden territory. I’m going to center this discussion on how I view the technology as a partner, it’s strengths and vulnerabilities, and more importantly whether we’re in a bubble that is going to burst.
Working with AI
To be clear, I do not work with AI often. Other than interactions through a web search and using it to look up instructions for obscure equipment, I mostly avoid its use. The reason is rooted in the experiences I’ve had when I was trying to use AI for real work.
Working with AI is like working with an intelligent intern that is extremely eager. They can find stuff and learn very fast. But they also stumble on their feet trying to solve problems they are not experienced in solving. There are two problems with this. They stumble with extreme authority, often claiming their reasoning is infallible even when they are wrong. They also produce a LOT of information where volume is the currency of success.
To address the confidence when they are wrong issue, the “manager” (you) needs to have enough knowledge of the task to be able to readily see the BS before it becomes entrenched in your work. That requires a level experience in doing the task yourself so you know what is reasonable without having to dig into details. It also requires experience reviewing engineering work products and looking for the “gotchas”.
To address the volume of information, the “manager” has to ensure that the prompts are sufficiently constrained and the results are properly filtered. You essentially have to slow the AI down without stifling its creativity (like writing a really good engineering specification).
Neither of these solutions means that AI not magically does the work for you, it just changes your focus from the menial stuff that takes a long time (but you’ve done before) more into focusing on defining the problem and constraining the solution. This makes for a very good engineer, but an engineer that has already “paid their dues” with the menial work in the past.
Strengths and Vulnerabilities
From my perspective, AI’s number one strength is the ability to very quickly gather related information into an easily digestible format. If any older engineers had the pleasure of working with real librarians while doing a code study, you will know exactly what this feels like. AI is like that person that can start from the original reference, find all the associated referenced documents, connect them, and summarize them for your digestion. That is pretty magical. The most common example I have, which I encounter often, is putting together regulatory reviews. I may know the starting point for how a product is regulated but there are always referenced codes and standards many layers deep. AI can find all of those and steer me to the correct pages so I can read the applicable aspects. What would have taken me days to piece together can be done in hours.
I’ve mentioned that AI is like an over-eager intern, which I consider a vulnerability, but more importantly it can build a dependency that is unbreakable. If you only know how to use AI to do your job, what happens when the AI is not available. Worst yet, something goes wrong and the AI doesn’t know what it did? This all links back to the AI replacing the interns. Interns “pay their dues” and gain experience when they are working for you. This builds the next generation of managers that have the experience and judgement to oversee the AI. By replacing the interns, we are losing an entire generation of managers so at some point nobody will have the skills necessary to navigate the future AI world.
Are we in a Bubble?
The writing seems to be on the wall for AI’s unfettered growth. There were several ways that the pace of adoption is unsustainable, including but not limited to: infrastructure requirements, environmental concerns, costs, and brain drain. This is a long section, so the TL;DR is that we’re already seeing the bubble burst and it’s not because of just one factor (that could be avoided) but an entire host of unrelated issues that point to the inevitable slow down of this technology.
The infrastructure limitations could be an entire chapter to a very boring book. There are three big hitters here, land, water, and electricity. The required data centers are huge, and there is a large NIMBY movement making finding space and developing these facilities increasingly more difficult. This may be why companies are looking at putting these data centers underwater or in space. We can also debate whether the technical advantages are real, but I think they’re just looking to avoid people. Water usage is also a big political topic. Regardless of your view on climate change, it is uncontestable that most places where data centers want to exist are more arid. Taking water away from the people in those situations is not insignificant and faces tons or resistance. The last issue is probably the one that is most limiting. These things require lots of power. The current grid is not nearly robust enough for them to just “plug in”, so either they have to drive a (very expensive) grid upgrade or be powered off grid. Asking the general public to pay for grid upgrades just so large companies can cash in on billions of dollars is facing lots of resistance. Creating off grid power that can power these facilities is very expensive and in some cases waiting for technology that doesn’t exist yet (nuclear).
The environmental concerns are closely coupled to the infrastructure concerns. Large construction sites will do some damage to local ecosystems, and the waste produced just during construction will be significant. Water usage will not only take water away from the ecology, but in most cases return water in a more polluted state than it left (mostly by heat). In terms of power, there is no free lunch. Creating that much power, whether from the grid or locally will have a negative impact on our future environment – some technologies more than others. Clearly using a million diesel generators is a worst case scenario, but even building nuclear power plants raises the specter of environmental concerns into concrete production, waste heat rejection, and radioactive storage.
The cost of AI has been, to this point, relatively low for the output. Even as a skeptic I have to be realistic about that. But I also have to think about where this is going. AI companies are in the red because they undercut their operating costs to push implementation. That means they got companies and everyday people hooked on this low cost AI feature. This cannot go on forever or the AI companies will go out of business, so they will necessarily have to raise prices. We’re already seeing some of that happening and the backlash is real. Companies are going from “token maxxing” to limiting the amount of AI employees can use. That means there is a ceiling to how many tokens people will buy and it’s connected to the price of those tokens. As token prices go up, demand for those tokens will decrease, effectively killing the mad dash towards expanding AI usage.
The last point I want to make is about brain drain. When companies run by executives that didn’t actually do the work first saw AI, their first thought was “we need less people” and went through systematically slashing workforces. This seemed really bad (and it was for many), but really just put a much worse situation in motion. What they soon found out is that AI does not replace people, just some of the menial labor those people were doing. As a knee-jerk response they hired back a lot of people, but basically just those that are “managers” (as described earlier). Those managers now had to not only do their old job, but also babysit a bunch of eager “AI interns”. This creates an environment where there is not a new generation being trained as the managers now spend their time training and overseeing AI agents. It also burns these employees out because of the increased pace of the work (for what currently seems to be very little output increase). What happens when they retire? Will AI agents take their place? That might work for a couple of generations until something goes wrong and no human knows how any of it works. This is the dystopian future some AI execs have promised. I'm less pessimistic and instead think that is will make companies learn some very hard lessons and either fold due to unresolvable issues or lawsuits, or they will reduce reliance on AI just to ensure enough human training is happening, again lessening our need and growth of AI.
Final Thoughts
I’m excited about the potential for AI to help us solve some of the world’s hardest problems. But I will not be replacing interns with AI for two important reasons. The first was covered here and it’s because I feel an obligation to train the next generation. The other is darker. The intern can be fired. If an intern brings me a bunch of useless crap and wastes my time, they know, deep down, that their days are numbered. AI has no such sense of consequence. It will endlessly waste my time until it is successful or I just give up. So not only do I have an overly eager intern, I have one that doesn’t really care how much of my time it wastes. Sounds terrible (and direct experience tells me it is).
I’ll leave you with one more thought that is not unique to me, but really captures my overall sentiment about AI. I was at a conference at a local design firm and the department head for a university industrial design school briefly discussed the design process and the value of slowing down. It makes for a more human solution that is more rooted in the consequence of the design decision than meeting requirements. That was briefly connected to one of the real concerns with AI is that it is an “solutions accelerator” by design. It is trying to speed up the design and decision process to produce more stuff. This flies in the face of conscientious design. You can debate whether this is good or bad, but I know where I stand. And I think more and more people are starting to think the same way, which is why AI demand is decreasing and will likely continue to do so until it is relegated to being a reference gatherer.