Forever junior: Skills AI can't develop for you

My work is more data science related, but remember those same struggles where I spent hours, days, weeks, months, even years crashing my head against a problem. A lot of daily hours researching, talking to colleagues, implementing tangential papers that lead to nothing. Most of the time I arrived to a satisfactory solution, but others just have to gave up.

Maybe because I experienced all those struggles I'm confident to delegate a relevant portion of my work to some LLM, being able to validate the results, knowing what to ask and detect easily when something was bad implemented or where I gave ambiguous instructions.

The catch is that I'm starting to do things where I haven't experienced those struggles in prior times, for example, web app development for my data work. I don't feel confident that I'm doing a good work, I'm just vibing, don't know the implications of some decisions. It doesn't makes me feel comfortable shipping things I don't really understand, but also is nice to be able to do things that previously required years of studying and practice.

We are on strange times.

I have a tendency to prefer writing my own library for something instead of relying on someone else's. Not always, obviously, but quite often existing libraries fall short or are too complex to use. And solving the problem myself is almost always an educational experience.

For example, I'm currently writing my own graph layout library because I'm not happy with Dagre and ELK.

Nobody would pay you to do something like this today. You'd be expected to punch that ticket in a couple days and move on.

You would not have time because you'd be fired before you finished. That's what's stopping people from doing projects like this today.

My side projects are all hand coded for this reason. There's no pressure to ship fast, so why not take the time?

Terry Tao has a term for this I liked quite a bit, "productive struggle"

I want to say that this isn't a new thing.

Somewhere a while back (maybe it was in a Ruby CVE) I heard the phrase of "beneath and before" relating to the layer you're working in.

You need to know what goes on beneath the layer you are working on. If you're writing Java with Spring, you should understand how the annotations work and are wired in and how Spring does its magic. If you've got methods that call a database layer, you should understand the beneath of how the database is structured.

The before part is understanding what is calling the code you are writing. What is it providing you? How is the data structured? Or if you're responsible for that, how should the data be structured so that it is easier and more reasonable for both the code you're writing and the calling code.

However, that curiosity about the other parts is largely gone. I've seen this in 2026... but I also saw it in 2010 and to some extent even before that (though in the naughties I was at a tech company and surrounded by people who were genuinely interested and proficient at their craft).

Today, I'm working with people who seem to have had their meat-ware understanding frozen at Java 7 and unable or unwilling to progress... and they've been frozen at that for the past decade.

I'm not blaming AI for that - it's been a problem since the before times. The pipeline has been collapsing for a decade with the forever junior. AI has made the forever junior more "productive" and may have accelerated the problem. I lay this firmly at the feet of the people who haven't any action to better themselves and try to learn more of the beneath and before.

I feel like this has been a growing problem even before AI, I've seen so many people who have spent more years at better universities than I have, who during their junior years could barely do what I was doing when I got out of community college with only an associates, and this was back 10 years or so. I also have a few new graduates who are very solid, I don't know if it's just that their specific university prepares them better, or they took it more seriously. I remember how notoriously hard the testing was at UCF for the CS program, I hear they've made it 'easier' and I believe it was and still is pen and paper, this was for the CS Foundation course, which if you failed (and many did/have) you couldn't take upper level CS courses.

I think the reality is that schools need to stop pretending like they're producing good students just because they're doing the classwork, nobody wants to hire incompetent people.

I think the reason Anthropic thinks programming will be dead in "6 months" or whatever random nonsense they spout is because you might not need intimate knowledge, but they forget how over expensive all their tech is, by the time they can produce an always perfect software developer, will anyone be able to afford it?

Heck they can't stop people from jailbreaking the models, or from having it commit felonies.

Pangram says 25% AI but there's a lot of lightly edited mic drops that AI likes to do along with Not A but B frames (like "As a Junior fresh out of school, “still learning what I’m doing” isn’t a caveat — it’s the job description" doesn't get flagged as AI even though it reads like something fresh off Claude or distilled Claude)

ironically, the best way to learn how to write is to actually write. the other choice is to forever remain a mediocre writer, limiting your reach and audience

Yeah, it's just exhausting to read that even if the message might have been interesting.

A paragraph below she uses it twice in a row:

"But somewhere in that process I noticed the real gap wasn’t in my agent’s behavior — it was in mine. So the focus of this article isn’t really about the skills I wrote. It’s about the ones I had to develop in myself to keep up with the ones I was developing with the agent"

Saw this comment and it got me wondering too, because there are at least some LLM tells as well as some places where it diverges into more natural writing.

There is a byline with a fairly google-able name (reasonably uncommon), and it appears to represent a real person who was formerly an intern at Criteo. English may not be their first language, I guess.

But yes — I think it is a little unfortunate in the context.

Yep, I was just about to comment about this but you beat me to it. You know something's amiss when you have to use incognito mode in order for the site to look the way it's supposed to.

I'm getting:

  uBlock Origin has prevented the following page from loading:

  https://tech.criteo.com/blog/human-skills-ai-cant-develop-junior-engineers/

  This happened because of the following filter:

  ||criteo.com^

Umm, okay? The entire domain is blacklisted?

I was using a lot of dashes before AI. What do I do now? -- Change?

AI has gotten to the point where there really is no telling.

Don't believe this about yourself. Everyone is capable of becoming exceptional in their own way.