Ten years ago I learnt many things - finance, AI, computer science - by failing at them. Slowly. Badly. This week I built a working AI platform in an afternoon.
Things have changed a lot for learning. And I’m not sure if it’s good or bad.
Packed week, and a disorienting one. My face went up on a billboard in Times Square for the H2O AI 100, which is about as far from my usual week as it gets. I sat on a couple of panels and fielded the usual run of calls on gigs and training.
A few things I started this week led to this reflection on learning. First, I started reading a recently released book on AI evals for engineers. Second, I decided to revisit and relearn the architecture and code of a few platforms I used AI to build over the past year, motivated by the recent slew of security issues. Third, I finally started setting up an online course on AI governance and risk management I have been putting off for a while.
Ten years ago, many things in areas like finance, AI, computer science were really hard. And we learnt by reading, doing and failing. Learning these things is still hard now. But we can make it easier. By using LLMs. Or we can skip that totally and just move on to the building part.
So what does it actually mean to learn these days?
How I learnt it in 2015
While I had already done a masters in knowledge engineering (which was kind of AI in the 2000s) and financial engineering by 2015, AI was still a new thing for me then.
And I remember taking Andrew Ng’s Machine Learning course in 2015, free, on Coursera, the old one that made you write everything out. You implemented the thing by hand. You wrote the code, you ran it, and most of the time it did not work, and you tried until you understood why.
After the course I remember signing up to Safari Books Online subscription, the O’Reilly one (which I still have today), and went through the books one at a time. Read a chapter, sit down, type the code out myself, watch it break, go back and work out why it broke. There was no agent to ask. If you wanted it to work you had to understand it.
Three learnings from those years. First patience. Sometimes you only get it at the third try. Second, first principles. When something did not work you tried to understand why. Not just copy and paste code and pray hard. Third, that the fundamentals, the technical and the practical were never separate things. You could not write the code without the maths underneath it, and the maths stayed abstract until you made it run.
That was the thing about learning then. If the code ran well, it ran because you understood it. There was no faking it.
What has changed
That link has broken.
Today you can produce the output without understanding any of it. The agent writes the code. The model writes the analysis. The deck more or less assembles itself. The result often looks identical to the one the person who understood would have made. But producing it no longer proves anything, because the understanding was never required to get there.
I have written before about the three legs of learning something. The fundamentals, which is the why. The technical, which is the what. The practical, which is the how. In 2015 you could not produce a result without having at least some of the three, so a result was proof. Now a tool can hand you the output off a tiny sliver of the practical, a bit of the how and none of the why.
What has not changed
But has it really changed?
A model is still math. There is no magic under the agent, however slick the output looks. Linear algebra, probability, a loss function being pushed downhill. The thing still has to be correct, and correct is still a property you have to check, not a feeling.
The system around the model still decides whether any of it is safe to use. I have spent years making the point that you govern the surrounding system (including the people), not just the model, and nothing about faster code changes that. If anything the surrounding system got more important.
You still have to decide what is good enough. Evaluation did not get easier because the output got prettier. It got harder, because a wrong answer now arrives in confident language. Someone still has to sit there and judge whether it is actually right, backed by understanding.
I wrote a while ago that thinking in AI was never about knowing the latest model. It is about understanding a problem well enough to break it into the right tasks, and then picking the right tool for each one. The models turn over every few months. That skill does not.
And the three legs themselves have not moved. You can lean on the tool for the practical far harder than you ever could, but knock out the fundamentals and technicals and the whole thing still tips over, same as it did in 2015.
And the scammers are still around. The vocabulary changed, that’s all. The people selling a shortcut to mastery, a secret pack of prompts, a way to skip the hard part and get the certificate, they never left. They just learnt the new words. Prompt and pray, prompt and awe, prompt and parrot, dressed up for 2026.
I still remember a class that I sat in on to observe, earlier in the year. The learners went away happy because they managed to type prompts into a tool and get text, images and videos back. The hardest thing they did was uploading documents into a custom GPT. I still wonder why they were happy with the illusion of learning.
Andrew Ng had a line that you cannot learn to ride a bike by watching other people ride. But the issue now is that there is a machine that will ride it for you, smoothly, while you sit on the back and feel the wind and tell yourself you are learning to ride.
So what does learning mean now
I think it has changed in one specific way.
It used to be a by-product of doing the work. You could not do the work without it, so it took care of itself. Now the work can be done without any learning, unless you deliberately slow down and insist on understanding what the tool just handed you. Learning has gone from something that happened to you while you worked, to something you have to choose on purpose.
It is why I spend my time on a side project called Job AQ, a freely available tool that measures whether a person can actually do the work, rather than whether they produced something that looked like they could.
And also why I have been releasing materials on AI governance and risk management at my Quaintitative site - a guide to AI risk management in Singapore, enough materials to fill a full day course on AI governance and risk management for risk and compliance folks, and a half day version for leaders. Also freely available, because I learnt that from Andrew Ng.
Maybe this is just the grumble of someone who learnt the slow way and is a little proud of it. It probably is, partly.
But is forsaking the fundamental, technical and practical stuff because AI makes them easy a good thing? I don’t think so.
#AI #AIFluency #AIGovernance #AILiteracy #Reflections #AIRiskManagement
Reading the article made me think about Michael Faraday.
It may be that some of us, no matter how hard we try, will never realistically be able to achieve the kind of proficiency in building using code. It's a particular kind of skill, I think.
Faraday had little understanding of advanced mathematics, yet he became a pivotal figure in advancing our understanding of physics - I think it's because he had an intuitive understanding of nature. He needed a Maxwell to describe his work mathematically. What if Faraday had an LLM? Would that have made him any less of a towering figure?
I guess, for the rest of us mere mortals, we should learn as much as we can about codes (and mathematics!). But perhaps the value comes from knowing what is worth building and doing it as best as we can.
PS: Usually, I don't comment much. But the article was thought provoking. Thanks Gary.