A lot has changed since I last made a note in this diary, more than three years back. LLMs have turned the whole outlook of the economy upside down.
I remember discussing with a colleague the possibility of AI-based, self-improving applications as the future of software engineering. Now LLMs are doing that, in a way, using matrices on steroids. Not good architecturally, but viable from an economics point of view. And they get things done seemingly better than the existing systems, which takes the focus away from other machine-learning options. To make matters more out of hand, bad software engineering practices have compounded.
However, it raises another question: do the existing software engineering principles matter, now that they have taken the human out of the coding part? A good thing is that things now centre around domain expertise rather than code quality. This puts my assumptions for AI-assisted model-driven development in the right place. It also places me on my sub-conscious path towards self-adapting systems.
I am striving for a future where at least my software systems are less dependent, or not dependent, on LLMs. This is contrary to the current trends. Ironically, I am commissioning LLM-based agents to do so. Why? The economics says it all. When I can spend Rs 5K on an AI agent that gives 10x output, why spend at least Rs 25K on a human engineer for 1x output? But the dependency scares me. It also bugs me internally that it is not the best solution to the problem at hand, despite the fact that it works wonders on almost any soft-skill task.
Now the IT industry is moving towards systems built around LLMs, just as the web centred around React, Vue.js and WordPress. It is hard to move away from them. However, I must confirm that things are very uncertain. No one knows for sure where things will go in the next two to five years. It makes LLMs seem more like a discovery than an invention.
Drafted in September 2026
Updated for site in October 2026