From Fourier Transforms to AI: A Personal Journey of Perspectives in Tech
Do you ever remember random moments in your life with an incredible amount of detail? I’ve read somewhere that your subconscious may store those moments in your memory because it wants you to learn something from them. But it could also be the other way around: you only remember those moments precisely because you learned something from them. Whether the egg or the chicken came first, there are two such random moments in my memory while studying science and working in tech. If my life were a comic book, I’d appear in these comic strips with my eyes wide open and a lightbulb in a balloon coming out of my head.
The first one.
Stanford, 1997. I am an engineering master's student and I am doing homework with one of my classmates in the Fourier Transforms1 class (Don’t worry if you don’t know what those are, it is not relevant to the story and, honestly, I do not remember either). “Doing homework” meant comparing the exercises we had previously and separately resolved. The first exercise gave us a mathematical function and asked us to calculate its Fourier Transform. My classmate was quite impressed when I showed him my two pages full of equations, where I deduced the result by applying a step-by-step mathematical integration method. But I was even more impressed when he showed me his approach: he got to the solution in one line, with no single operation! We had arrived at the same conclusion. It just took me two pages while he needed one line. “How could you know?”, I asked, baffled. “Duh! I just looked at the table in the book that shows the transform for this specific type of function”, he answered.
This was an Aha! moment for my physicist mind. “You don’t need to understand how to solve the equations, just read the answers in the book!”, I thought. And I became so much more practical after that. When building technology you do not need to understand the formulas; you can trust that somebody has proved them and just work with them, saving lots of time.
The second one.
Madrid, 2017. I was the CEO of the start-up I had founded a few years prior, and I had not practiced real math for a long time. I had been working with the CTO of the company on a methodology to measure the incremental impact of advertising campaigns, and I was extremely proud to have come up with a formula to calculate the “incremental factor” over the weekend. I used the Internet; the equivalent to the book 20 years later. Of course, we hired a data scientist to advise us before implementing any methodology. Once he understood our problem, the data scientist delivered a proposed solution based on a computer simulation, without a single formula. The incremental factor would simply be extracted from the observed results. “But is my formula right?” I asked, a bit puzzled. “It probably is. I don’t know, but it doesn’t really matter”, he replied.
This was another Aha! moment for my physics/engineering mind. “Now you don’t need a mathematical function to predict an outcome based on a certain input. You can simply use computing power to simulate many inputs and observe what happens!”, I thought. And when you have data scientists and available servers in the cloud, that is the easiest way to go, even if maybe not the cheapest.
I used to tell this story because it helped me understand the importance of mixing different perspectives. And this was, by itself, an important learning.
But today, I thought about this story when reading one of the many articles in my inbox: an application based on generative AI launched a notification with fake news and made the additional mistake of attributing it to a specific reputable media company. I imagined generative AI as the latest shiny technology that enables the third approach in my story: provide input and observe the outcome without knowing what happens in between. And this made me wonder whether we are taking this approach too far, both in our personal and professional lives.
You could argue that in 2017 we should have used the formula to save some energy, but that was not what companies around us did: the data science approach was in fashion. Today, using generative AI to launch notifications based on news may not only consume a lot more energy but also sacrifice something that is core to the product: accuracy. At least as long as the problem of hallucinations is not solved. Isn’t there already existing tech that can do this more efficiently?
There are many clear examples where generative AI solves problems that other existing technology could not have solved (we have not yet discovered the “formula”) or where it solves these problems so much faster: early disease detection in medical imaging, environmental modeling for climate scientists, protein folding... But examples like dealing with already known facts, generating research reports, or performing other tasks where 100% accuracy is required… Well, they don’t seem so clear.
Are we adopting generative AI in some use cases just because it's revolutionary, cutting edge, the fashionable thing to do, or simply because it's available? Just like I adopted the advisor’s approach in 2017… Or is it the best available technology to solve a given problem? This is a good question to ask ourselves before we go along with the hype. Understanding when and how to apply the most advanced tools - and recognizing when they should not be applied at all - is another great perspective.
https://en.wikipedia.org/wiki/Fourier_transform