AI will partially replace software engineers posted on 05 March 2024

I disagree with NVIDIA’s CEO Jensen Huang’s statement that AI will fully replace software engineers (at least not for the next 10+ years) and that there is no need to learn to code.

AI will change software engineering and my take is that the following will happen:

  • Software engineers will become significant more efficient with AI. The number of software engineers jobs will significantly decrease as AI becomes better at writing code
  • Only those who know how to use AI will stay relevant in the industry

For those who argue that AI won’t be able to do the work because your code base is too complex or your APIs are undocumented, keep in mind that:

  • AI for software (e.g. copilot, codewhisperer, code etc.) is today at its infancy – it will only get better. I’ve personally mostly used Copilot, but as models become tailored per company, I believe they will drastically improve.
  • AI is better at processing large amounts of code than we are. It doesn’t necessarily need to understand a system at a high level, it can just cognitively keep every detail in memory – so it will outperform humans in many scenarios, e.g. understanding/explaining/updating a legacy code base that no one knows
  • AI can discover hidden patterns/relations in the code that are unknown to people
  • AI can solve mathematical problems on part with a human Olympiad gold-medalist (see AlphaGeometry)

I believe software engineers will still be needed for a few use cases:

  • AI will need to be given well written/scoped prompts and its output to be reviewed and/or corrected. This means that the ability to write code won’t be the main value of software engineers anymore – their ability to translate business requirements to high level solutions and review the code will be their most important asset.
  • Some domains (e.g. health systems, government fields) require correctness, and this will be achieved only through human reviews (even if humans perform worse than AI, it will take a long time before we fully trust AI for these cases – similarly to self driving cars being better than humans is not enough)

Have you jumped on the AI bandwagon already? I have at Databricks and while AI may spit out garbage every now and then, I’ve already found a few use cases where AI makes me significantly more efficient.

LinkedIn post