Journey into Coding with AI [1/4]: Running Back to Code
(Part 2: Shifting Gears)
I love running, but I also love coding. I can still code for hours a day. But there’s a limit: I could only get a few hours of quality work in before feeling exhausted. That changed a month ago when I started experimenting with coding AI. Since then, I’ve had to consciously cut down my coding hours — it’s so much fun (and addictive) that I simply can’t stop! The continuous joy and dopamine hit from a sense of accomplishment just keeps me going and going!
A Few Surprising Examples
Here are a few examples that genuinely surprised me:
- Coding AI caught a bug in my refactor that had disrupted gradient flow — a hidden problem that would’ve haunted model training down the road.
- After realizing I needed a “hook” pattern in Python, I asked which of two approaches would be more efficient. Coding AI wrote a precise benchmark on the spot, ran it, and told me exactly how many milliseconds one beat the other.
- To replace a home-grown implementation, I asked coding AI to find a reputable OSS alternative with a commercial-friendly license. It surfaced a few, recommended one, and coded up the integration with some trial-and-error (the API wasn’t properly documented). Voilà — I had a faster, more reliable implementation ready to use!
- After a big refactor, I typed “update the docs,” and coding AI got it done while I grabbed a coffee.
- I needed a web-based demo for a project, so I spent a couple of hours with coding AI and implemented a pretty polished Streamlit app with streaming debug info.
It’s like I suddenly assembled a small team of servant interns: super-smart at what they do, driven, and very efficient. All I need to do is give clear, thoughtful instructions, provide the right amount of supervision, and suddenly nothing feels impossible. It’s exhilarating!
Programming with Concepts, Not Code
This is much more than what the term “vibe coding” suggests. It feels like the next big evolutionary step in programming. In the ancient days people coded in 0s and 1s. Then came assembly, macro assembly, interpreted languages, compiled languages, procedural languages, object-oriented languages, and so on — a steady progression of abstractions. I think we’re at the cusp of the next step: programming with concepts instead of code. To implement a solution, you lay out a clear conceptual plan and instructions, and AI compiles that into implementation. The efficiency gain is analogous to how the auto industry uses robotics to mass-produce cars. No one in their right mind now would suggest going back to building cars by hand (well, except for a few).
With this excitement, I’d like to jot down in this (long) post my experience and a few observations and projections. Hopefully I can clear up my thoughts a bit in the process — and also keep you, my audience, reasonably entertained! 🙂
(on to Part 2)
Continue the series: Part 2: Shifting Gears → · Part 3: Decision-Bound Programming →
Originally posted on LinkedIn.