Pick an industry where you can stop learning new things once you're out of school, because the computer industry is not one of them and never has been.
My fucking job used to be writing 6502 assembly language on an Apple ][, and I loved it, but I hope I've forgotten enough of it to have room to learn new things. If only I could forget all those hex I/O and peripheral addresses from $C000-$CFFF and the Monitor ROM routines from $F800-$FFFF, without forgetting how brilliantly beautiful Woz and and Allen Baum's code is.
https://6502disassembly.com/a2-rom/OrigF8ROM.html
Forgetting old stuff to make room for new stuff is one of the most valuable skills you can have in this industry.
The next most important skill is persistence ;) -- writing stuff down before you forget it, in a way that won't make future-you hate present-you when you need to learn it again.
I saw this idea from elsewhere that the things AI helps with was never the "profit bottleneck" of companies. 10x engineering productivity gain does not translate to 10x more revenue if your profit bottleneck is customer acquisition and retention. In other words, your profit is limited by how many people are willing to give you money for your services and AI can't really affect that.
And even the productivity gains per individual is a generous assumption. An individual can only prompt (and check! You guys check right?) so much. Sooner or later the pendulum is gonna swing in the other direction and it will be cheaper to build a team than equip individuals with AI _to deliver the same value_. I'm also assuming that AI is still in the VC-subsidized pricing stage.
Hence why I think the job in the future is gonna be pretty similar to the job four years ago.
> The next most important skill is persistence ;) -- writing stuff down before you forget it, in a way that won't make future-you hate present-you when you need to learn it again.
Huge amen. Another thing I've found Claude very useful for is writing documentation for legacy systems and cleaning up my own notes on it. I only need Claude to be 70-80% correct because from there I can take it. That error margin is no different from moderately-outdated-but-still-useful documentation.
When that pendulum swings, I will have a documentation binder that I will print money with.
I am very convinced that today's SOTA level will be accessible 3 years from now, but you might be right about the SOTA accessibility at that point in time.
Still with today's SOTA you need to know much less details to be able to write code than without it.
People neglecting this side will inevitably become less proficient at their job and become rather helpless without AI, that's all I'm saying.
You put a number on it at the top of the thread: three years until people relying on agents have forgotten how to do the job. Copilot's technical preview was 2021, ChatGPT was November 2022, and agents that write the majority of the code are a 2024-2025 thing. So the window you're predicting is roughly the entire span in which these tools have existed, and the atrophy hasn't shown up in it.
For scale, the last ARM I wrote was for a Sony Clié and the last PowerPC for a Mac, both more than two decades ago. If I needed to write ARM for my Mac now, I could brush up about as fast as I could on the 6502 I mentioned upthread, most of whose opcodes, $C000 I/O addresses and monitor ROM entry points I genuinely cannot recall any more. Two to four decades of not touching an instruction set didn't cost me the fundamentals, and it hasn't made me less proficient in Python and TypeScript or helpless without AI. Three years of using a good tool is not going to do it either.
There's also a distinction buried in your last line, because "people neglecting this side will inevitably become less proficient" bundles two different things. Losing a skill you had is not the same as never acquiring one, and only the second is an argument about tools.
So take acquisition. I've been learning PDP-7 assembly in order to recover, OCR, and analyze 128 pages of PIXIE assembly source and octal machine code that Heinz Lemke wrote at Cambridge in the late 1960s, and get it running in an emulator. A language I never knew, with custom instructions for bespoke networking hardware, no Stack Overflow, and almost nobody left to ask except Heinz himself. AI has been extremely helpful for exactly that, and the understanding is still mine to build: I have to predict what the machine will do, and catch the explanations that are wrong.
https://www.youtube.com/watch?v=jDrqR9XssJI
On feeling like a fraud: the test isn't whether you typed it, it's whether you can predict what it does and fix it when it breaks. Plenty of hand-typed code fails that test and plenty of carefully reviewed generated code passes it. Your reasons 1 and 3 are good reasons to hand-code, since problem solving is genuinely fun and nobody is waiting on your free time, but reason 2 is measurable, so measure it instead of trusting the feeling.
The atrophy you're worried about is real. It just isn't caused by the tool. It happens to people who stopped wanting to learn, and that predates AI by decades, even millennia. I watched it happen back when much of the industry ran on assembly language.
If you're not using AI to teach yourself new things, you're not holding it right.
Personally the tools have simply gotten too good to ignore so I have been thinking about how to write skills and prompts in a way that enable me to keep learning, overcoming friction and challenges and not simply using the tool for hand holding and jumping straight to the end result.
For me personally, what matters is the process of learning and mastery, my self esteem depends on it, and enriching that with AI instead of AI eradicating that process for the sake of efficiency is my current challenge.