On AI Relacing Software Engineers
There have been many discussions recently about us heading toward a future where every job can be done by AI.
Even software engineers are feeling anxious after demo of Devin because it can build apps using different tech stacks, debug and fix errors faced on the way, and deploy apps.
When it comes to evaluating Devin, SWE-Bench was used (a benchmark that checks how well language models solve GitHub issues), and Devin successfully resolved 79 of the 570 issues, which is 13.86% success rate.
That’s impressive result for language models. And the percentage of issues it can solve will go up over time for sure.
The question is - when will Devin reach that level where it can work on complex projects end to end without human supervision?
I don’t believe it will happen anytime soon.
Instead, I believe AI agents like Devin will evolve as an extension to human software engineers, freeing them from routine tasks so they can focus on more interesting and complex ones.
Because software enigneering is not solely about coding, it’s also about understanding problem domain, identifying and analysing problem, cross-functional collaboration.
The role of software engineers will evolve too. SWEs will need to:
- Be good at prompting engineering, because the ones who master it can be faster at shipping code, writing design docs, etc.
- Be able to set up and manage infrastructure around models built by ML engineers and researchers.
The jobless future for software engineers seems to be far off. To get there, someone at least needs to build a reliable AI that can do any task humans can do. And they need to prevent regulators from creating barriers to the widespread deployment of such AI systems.