Realizing, i have discussed the topic of llm short cuts many times, but have not mentioned how prescient the 2012 article from erik dietrich on this was. on the rise of the expert beginner, Diedtrich argues that there ore many people who fall into mediocrity because they have achieved expert status and there’s noth hing left to learn. And I suspect llms can also hold back your learning while feeding you enough positive feedback to keep you happy where you are. They are holding you back by stunting your deep learning.
And for coding in particular, people code agentic harnessing their way to freedom maybe similarly as erik’s bowling onology, will not know how stunted they are the more proliphic the harnesses become.
(As an aside there might also be people aware they are plateaued, and okay with that, not seeing any benefit of skill acquisition if companies no longer care whether you use coding harnesses or not. For this group of people, lets consider them separately later as well).
Just review the code its fine
There is an interesting psychological effect that happens when something complicated is demonstrated to us where we are almost confident we understand what’s going on. But when feynman asks you to explain it in your own words, you cannot. This also has the corollary in complexity theory, where there are a class of problems that take linear time to verify but nonlinear maybe polynomial time to solve. Like verifying sudoku say will be proportional to the number of squares more or less , but solving sudoku takes more time. And basically all of programming falls into this category too. Reviewing and verifying code will be faster than writing it from scratch. Running a program and seeing and understanding the input and output still takes time yes but not as much as thoroughly understanding the problem space, taking time to think through solutions, try one out and maybe even run into o wall and go back to the drawing board, foil oggain and eventually figure out some subproblem, that was in n the way, then solve the subproblem, gobock, to the original problem, and now solve that . Thats not exactly linear 馃槄.
Sweet spot
I got onto this topic listening here [1], on the goldilocks zone for LLM use wrt to writing, coding, note taking and anything else. Carl and Casey discuss that it is difficult to determine be the sweet spot especially since after all, history said writing and books will stunt minds at one point in the past.
So maybe if there is a level of challenge, and you dont feel bored, you feel you are challenging yourself in the right way, learning transferable problem solving skills and critical thinking, then thats all fine?
One Way Door?
Moybe i am writing about this because shallow understanding disguised as real understanding is also a story thing. Is it a dangerous one way path? I like Cory Doctorow’s example [3] of perhops what it feels like to use a natural language only interface on top of a database. It feels empowering perhaps, he theorizes, to ask questions of your own Snowflake data, without hoving to sit down and learn the underlying query language. (Side note, thats probably a good point also since Snowflake I recall added their own clauses like QUALIFY on top of ANSI SQL ). I describe a bit of my thoughts around a MCP SQL natural wrapper workshop I attended too [4], wrt the ironies of needing to add back complexity because of authorization after bare-bonesing it.
Side note on skill acquisition
Rereading [2] on Dreyfus, model, where an expert understands everything. Well, one, famous experts still admit to so much that isnnot known.
In any case, rereading his article also reminds me of the concept of needing to throw out a lot of what you know because you learned a lot of hacks and now -you hove to painfully unlearn them before moving on. (See Chopin Etudes).
I think even without llms, being stuck in a local maximum is easy but llms make it easier.
The industry effect?
Also right now the whole of software as an industry is hoving a kind of preCOVID moment. lets token our way out of all of the problems. Lrts choose only the problems tokens can solve? Leave the hard problems alone? What could go wrong?
Stuck on purpose
Will the music stop? What if LLM use stops being subsidized and now critical thinking and solving problems by hand becomes important again?
references
- philosophy, programming and prompts virtue ethics episode, https://youtu.be/xhgREGlAnPU
- Erik dietrich, rise of expert beginner, https://daedtech.com/how-developers-stop-learning-rise-of-the-expert-beginner/
- https://pluralistic.net/2026/08/01/dare-snot/#:~:text=Take%20that%20Cortex,they%20understand%20it.
- https://michal.piekarczyk.xyz/note/2026-05-01-odsc-closing-notes/#mcp-travel-agency-workshop