My Thesis on AI's Impact on Software Engineering
Software can only evolve sustainably at the speed that your collective understanding of the code can evolve. This disconnect was always there — outsourced code mills have suffered from it for years — but AI tools have put it on steroids.
Why does collective understanding still matter? AI tools currently operate at a relatively shallow depth in the holistic software engineering picture. Humans are still on the hook for problems that AI can create but can’t solve. That’s a recipe for tension.
If people still need to work on the code, then the hygiene of the code remains important. Performance matters too.
I don’t see much evidence of significant further improvement on this front from the models themselves — at least not from this generation of GPT-style transformer models.
There may be significant improvements possible by simplifying everything else to suit AI — shifting the computing paradigm into something homogenous where the entire end-to-end is accessible to AI models, rather than spread over disparate systems that have been optimised for humans. There are rumblings in this direction, and this will go much further.
It’s difficult to put an exact label on which parts of our profession have been cheapened or commoditised by these tools. It’s deeper than prototyping but much shallower than the whole picture.
At the moment, I feel it’s on a pathway to fundamentally reshape our industry, but not yet completely upend it. Outsourced code mills and companies that work very shallow in the product are definitely at risk.
What would change my mind
The signals I need to see to move from “evolution” to “revolution”:
- Another order of magnitude leap forward in transformer model capability that isn’t mostly marketing, or a new kind of model that provides the same.
- A platform or platforms offering a truly AI-native computing paradigm that allows AI to roam far more freely across the holistic end-to-end of our profession.
- Serious adoption of coding agents in big and complex projects among my peers. Engineers actually handing off big, meaningful features to agents without supervision. I’m seeing glimpses of this, but there is far more noise than signal — too much sales hype, not enough actual traction.