As a software engineer you need to be prepared to no longer write any code.
I sat down with my team's and asked them the question, what would it take for you to feel confident to not read the code?
Process, documentation and testing ranked in the top 3 changes, what would you need to do?
Reading through the replies of this AI thought experiment, the main points a number of you raised was:
- Tests and specs should be independent of the generated code, otherwise green tests can just lock in the wrong behaviours.
- Guardrails and structure matter, but only if the input/spec is actually correct.
- Accountability doesnβt go away. Bugs still slip through (same as when humans write every line). Building a no-blame culture is the practical answer.
- In some domains (especially long-lived or no easy rollback) the bar stays higher. Internal tools are already closer to this.
The original post has so much more great points of view, go take a look.
As a software engineer you need to be prepared to no longer write any code.
I sat down with my team's and asked them the question, what would it take for you to feel confident to not read the code?
Process, documentation and testing ranked in the top 3 changes, what would you need to do?
@DeniCodes Really enjoyed it so far. I'm about 75% through and it's already one I would recommend. I'm building a little library for my team in the office, a good reason to buy some physical books.
@ReyaneRemache 100% it's the same problem we are seeing. I love knowing the code and being able to navigate it easily just based on the last 10 years of operating in it.
As a software engineer you need to be prepared to no longer write any code.
I sat down with my team's and asked them the question, what would it take for you to feel confident to not read the code?
Process, documentation and testing ranked in the top 3 changes, what would you need to do?
@stas_sorokin_ This is really interesting, security around AI is a big topic at the moment.
Something will need to be done in this space as you raised.
I still read a lot of the code but I do find this year I hardly needed to change the code the model produce.
For internal tools this is pretty much how we run already, if the output is good we are good to go. Production we have a far higher bar but this could be down to the industry we serve (no rollback option and 1 codebase could be deployed for 10+ years).
100% we are already seeing this demand increase just down to the output rate or development teams can run at. This was some of the reason we ran this thought experiment.
Domain knowledge is also vital, developers need to understand this (even today) to really write the correct things.
@mdnadeemai All very valid points for the implementation part. I think the rules and guide rails are incredibly important but this all only works if the input is correct.
So a validation loop for the specification could also go through points 1 to 3
@David_d65029 Testing needs to become a solid wall but if the spec is not correct the test may also let the "bug" (or is it a feature) through. Interestingly all of these points matter AI or not and even with the best reviewer in the world.
π interestingly this is exactly the industry we work in.
Management are also onboard as it is not that we are going to stop reading the code but we are moving to a world where less of it will be read due to more being generated.
I prefer to run these thought experiments with the team thinking about the extreme, that way we can focus down on the key changes and put in actions to move in the right direction for the future.
I like this idea of reproducible generation. I have not seen this talked about but that makes complete sense. (I will be taking that back to the team)
For me the best word in this tweet though has to be "comprehensive".
Too often specifications are vague and the bug is not in the code but in the implementation.
This accountability point is very valid and I know quite a few people are feeling it.
First thing we did was to try and build a no blame culture and just understand that bugs happen and they slip through. We then learn from these and try and improve the process.
I see it a bit like on this Friday evening. We needed to jump on the PHP vulnerability, no one was accountable for the CVE in the PHP libraries (who goes that deep into library code). Things should be expected to come up, written by your best developers or some AI process.
@nipunn1313 Those are the people that need to be most worried about "AI taking out jobs"
If you are not willing to do the learning you will get left behind.