The Simplicity Advantage: Crafting Software That Lasts

Personally, I always strive for simplicity and elegance in software. AI makes that preference more important because both humans and models have a well-understood cognitive bias toward adding rather than subtracting.

• People systematically overlook subtractive solutions 

• Models have an identical bias and sycophancy makes it worse.

• The models know it:  Claude Opus will always overengineer with extra files, unnecessary abstractions and unrequested flexibility

The danger is not necessarily one bad generation. It is an accumulation.

• 623 million changes found cross-file reuse down 35% and refactoring down 70% from 2022.

• Copy / paste (copy-pasta) rose 41%, duplication 81%, and two-week churn 15%.

• AI adoption is associated with a 7.2% decline in delivery stability

• "AI slop is a tragedy of the commons" -  generation is cheap; review and maintenance are expensive.

• “AI-Generated Code Smells” increase bloat and coupling; not even detailed prompting could prevent it!

• There may be hope!  These new LLM chatterbots are amplifiers.  Some LLM-generated code is simpler than human reference code in every tested language except Python.  a generation can be simple while a thousand additions make it complex.

Everyone should add subtraction into the development loop.

The technical co-founder of OpenAI and co-creator of chatgpt says that:

• unchecked assumptions, 

• sycophancy, 

• overcomplicated APIs, bloated abstractions, and 

• dead code

. . . are the most-offensive recurring problems in AI-generated code. He goes on to say:

Coding agents also really like to overcomplicate code and APIs, they bloat abstractions, they don't clean up dead code after themselves, etc. They will implement an inefficient, bloated, brittle construction over 1000 lines of code and it's up to you to be like "umm couldn't you just do this instead?" and they will be like "of course!" and immediately cut it down to 100 lines.

And Kaparthy is not the only one who recognizes this problem. And newer, simpler approaches are not better because human judges add their own "always add complexity" bias when the models are trained.

Therefore before asking AI what to add, we must always ask the model what to delete! Do you remember your college curriculum on Software Design?

• Software Design - Complexity is the central design problem!
• No Silver Bullet - Accidental vs Essential complexity
• Simple made easy - Simplicity is deceptively difficult!

And finally, as I keep repeating in these posts, when you are trying to defend your code from vulnerabilities, exploits, etc:

There is no attack surface as good as no attack surface.

Less code means fewer vulnerabilities.