Expertise in the Age of AI

Brian Kihoon Lee
Essays

Expertise in the Age of AI
2026-05-12
Tagged: llms

Does it make sense to hire junior engineers in the age of coding
agents?
Junior engineers are expensive, both in salary and seniors engineers’
time. This cost was partially recouped through code contributions, but
today, it’s more effective to directly maximize the output of your
senior engineers. The hiring market reflects this trend: senior
engineers have an easy time finding jobs, while fresh CS grads are
having their worst years ever. And yet, OpenAI, Anthropic, and many top
companies continue to compete fiercely for junior talent. What’s going
on?
In this essay, I’ll explore the changing nature of expertise in the
age of AI.
Math as an analogy
I think it helps to think about the impact of AI in terms of math,
which had its AI moment half a century ago.
There used to be a job called “calculator”, which was a human who
could do math calculations accurately and quickly. These people balanced
books, calculated artillery firing angles based on distance and wind
adjustments, calculated optimal hull shapes for ships and aircraft
bodies, and so on. This job doesn’t exist anymore, and the last serious
use of abaci and slide rules was in the 1970s, due to the invention of
the scientific calculator. Calculators have only become more
sophisticated over time, with today’s numerical modeling software
running full scale physics and engineering simulations. (For the purpose
of this essay, I’ll use “calculator” to mean everything from basic
calculators to modeling software.)
Despite the existence of calculators, we teach and expect people to
learn algebra, geometry, and calculus in high school. Continuing into
the college level, we expect STEM majors to learn multivariable
calculus, ODEs, PDEs, statistics, and linear algebra. Upon graduation,
the vast majority of them use calculators every day and forget how to do
all but the most basic mental math.
There are two basic explanations for this discrepancy:

(Signaling hypothesis) The STEM degree filters the set of people who
can both learn and persist through four years of difficult math.
(Skills hypothesis) Struggling through math classes imparts some
hard-to-quantify mathematical intuition that is valuable for operating
today’s calculators.

As a formerly strong believer of the signaling hypothesis, I am now
increasingly buying the skills hypothesis (let’s say ~50% attribution to
each cause). It’s clear that senior engineers today are far more capable
of using coding agents than their junior counterparts, and a large
portion of this is due to having struggled through 5+ years of writing
code manually.
A job market in flux
Currently, the level of computing intuition needed to additively
prompt the coding agents sits at roughly 5 years’ experience level.
Today’s seniors were lucky enough to get paid to build their computing
intuition, but the gap grows as coding agents continue to improve.
In between coding agent improvements and natural variation in
learning aptitude, maybe 50% of new CS graduates will not be able to
catch up, ever. Some senior engineers will also eventually fall behind
the curve despite their head start.
To answer the opening question of the essay: only some junior
engineers are worth hiring, specifically, the ones who are good enough
to reach some useful threshold of “coding intuition” within ~2-3 years
of having graduated. Since there are not very many of these graduates, a
small number of elite companies compete fiercely for this talent.
The second-class tier of software consultants will continue growing,
expanding the total size of the job market, but I don’t anticipate that
their salaries will grow anywhere close to as rapidly as today’s senior
engineers.
Everyone should learn some
coding
Even as the bar to get into software engineering rises, I still think
everyone should learn some coding. Too often, I see people treat
computers as appliances - capable of doing what they were built to do,
but nothing more. If you don’t think of computers as scriptable or
programmable, then you won’t ever think to ask AI to automate something
for you! The same is also true for many other fields, too! Math, law,
taxes, medicine, DIY home repair, etc… Abundant and cheap expertise is
now available for just $20/month, if only you know how to ask.
I would say that the major unlocks are at:

1-2 weeks: Basic understanding of what the field is about and what
general words to use when asking the AI to do something.
1-2 months: Basic understanding of how and when to ask the AI
something.
4-6 months: Ability to check the output for correctness (using
external sources as needed).

If you’re already a software engineer, you might consider dabbling in
data science, frontend, backend, security, and performance
optimization/profiling – all of which are distinct skillsets.
Here’s a data science example of a “how + when + correctness”: A
coworker was running some correlational analysis on a dataset and found
it difficult to understand what was going on. I suggested he literally
ask Claude to “make it prettier using NMF”
– and all of a sudden, useful clusters started appearing.
(The expanded version of this prompt: NMF on the pairwise distance
matrix gives k cluster centroids and cluster membership scores.
Reordering the original distance matrix according to argmax(cluster
score) highlights the clusters. The “how” here is knowing the keyword
“NMF”; the “when” is “clustering on distance matrices”, and the
“correctness” is knowing the preconditions for using it.)
Conclusion
Do your homework! One weirdly common and nihilistic take on AI is
that you should stop trying so hard, and just use AI to speedrun your
classes. I think this is probably the worst possible response. Doing the
work is the best way to build mastery, and just like you weren’t allowed
to use a calculator on your middle school math classes, you should hold
off on using AI to do your classwork. The calculator advice sounded
condescending when I was a kid, and this AI advice probably sounds the
same – but I really do believe it’s for your own good. This advice
continues to hold after you graduate, too. Don’t use AI until you’ve
done it by hand at least once.

Follow me via RSS or via the newsletter.

You can also share on Hacker News.

The essay explores the evolving nature of expertise in the age of artificial intelligence, using the historical development of mathematics and computing as an analogy to frame the current shifts in the technology landscape. The author draws a parallel between the historical transition from human calculation methods to sophisticated numerical modeling software, suggesting that the understanding of mathematics and computing skills is now foundational for interacting effectively with AI systems.

The author considers two primary explanations for why individuals often struggle with high-level mathematics despite formal education. These are the signaling hypothesis, which posits that STEM degrees filter for people capable of persisting through difficult mathematical study, and the skills hypothesis, which suggests that struggling through these classes imparts valuable, though hard-to-quantify, mathematical intuition necessary for operating modern computation tools. The author leans more heavily toward the skills hypothesis, arguing that senior engineers possess a greater capability in utilizing coding agents, partly due to their extensive experience writing code manually.

The essay analyzes the current state of the job market, noting that the level of computing intuition required to effectively prompt coding agents is roughly equivalent to five years of experience. The author observes that this gap is widening as coding agents continue to improve, potentially meaning that only a small subset of new computer science graduates will catch up, though some senior engineers may also eventually fall behind. Consequently, the author concludes that only a select group of junior engineers are worth hiring, specifically those who can develop a useful threshold of coding intuition within two to three years of graduation, leading to fierce competition among elite companies for this talent.

Despite the rising bar for entry into software engineering, the author maintains that everyone should learn some coding. This perspective emphasizes that treating computers as programmable entities is essential for understanding how to leverage AI for automation across various fields, including math, law, and medicine. To unlock this capability, the author outlines a tiered approach to learning how to interact with AI effectively, suggesting that the major unlocks involve gaining a basic understanding of the field and appropriate prompting techniques, followed by learning the methods of asking the AI, and finally, developing the ability to verify the output for correctness.

The author provides an example of this iterative process in data science, demonstrating how specific knowledge of methods, context, and preconditions—such as knowing the keyword for a technique, the context of the problem, and the necessary conditions for application—is crucial for yielding useful results. Ultimately, the essay advises against the nihilistic approach of relying on AI to speedrun coursework, emphasizing that true mastery is achieved through hands-on work. The author insists that one must engage in the necessary effort, similar to how early mathematicians were required to perform foundational calculations, before attempting to delegate complex tasks to AI.