Baptism by Fire

When it was time to sign up for electives, I was late to the game again.

Coming back to school after years away meant that even ordinary things felt different. Everything was digital, and there was also something I had not fully appreciated before joining UCLA: I was not limited to the EMBA classroom. I could take an elective outside the programme, cross campus into public policy or another school, and learn something simply because it interested me.

I ended up taking a public policy course too — more on that later.

But one of the electives I found was Todd Holloway's Applying AI to Operations.

I have always believed that professors with substantial experience outside academia can bring something unique into the learning process. When I looked at Todd's background, it was a no-brainer.

For someone new to AI, the course became a baptism by fire.

I do not want to turn this into a story about the course itself. What mattered was the breadth and depth of exposure it gave me in a very short period of time. We were looking at AI in business operations, different models, digital productivity, automation, agents, cases, news, podcasts, and the risks and limitations of the technology.

One part of the course structure stayed with me. For the case assignments, we were expected to use an AI copilot to explore the case — and then the paper was not simply the AI's answer. We had to critique the answer. Was it accurate? Was it complete? Could we use it directly? What did it miss? What should we do differently next time?

That mattered because it exposed me very early to the strange combination that still defines my experience with AI: it can be remarkably useful and confidently wrong at the same time.

One analogy that stayed with me came from Ted Chiang's essay, ChatGPT Is a Blurry JPEG of the Web. The comparison was not meant as a literal technical description. It was a way of thinking about an LLM as a lossy representation of patterns in the information it had been trained on — capable of reconstructing something impressively useful, but also capable of inventing details when the underlying signal was weak or missing.

That was one of the first explanations that made hallucination intuitive to me.

It also helped me understand why prompting and context mattered. If the system was generating from learned patterns rather than simply opening a database and retrieving the correct answer, the way I framed a question, the context I supplied, and how I checked the response all mattered.

By the end of those weeks, I had stopped seeing AI as just a chatbot.

I was beginning to understand the models, their capabilities, their weaknesses, and how quickly they could be connected to the way businesses actually operate.

But another question was also starting to bother me.

If all of this intelligence felt weightless on a screen, what was happening behind it physically?