Well, I celebrated my birthday. :P
I worked on a functionality end to end using AI. It was a simple feature and AI wrote everything correctly in a single pass. I used Codex. After that, I had to ask it to adjust some edge cases, add more validations, and improve error messages. I also asked it to write an end to end test script, and now I have a script to test the whole thing.
After raising the PR, my boss suggested some changes. So I went back to Codex, asked it to make those changes, and it did. I again had to adjust some edge cases, but the tests were already there, so it was easy to verify everything. One thing I noticed was that the code had access guardrails, but it did not reason about determinism. For example, it did not check whether a variable could only have two possible values and still adjusted for a third condition. I think context is still a bottleneck, or at least a difficult problem.
I also did some systems thinking :P and helped a few colleagues with their issues.
I completed a service migration end to end using AI. Did it make mistakes? Sure. Did it correct those mistakes? Yes. I had to provide some logs and data, and it fixed the code. It was honestly life saving. It reduced a lot of stress and cognitive load for me.
As the data in the database increases, you start facing different kinds of problems. I think a heavily used service is a good problem to have, but you also need multiple guardrails and self healing mechanisms just to maintain consistency.
AI has taken a lot of load off my back. Otherwise I would have spent multiple hours on code migrations. Now I can spend that time looking at the system and thinking about what can be improved. If you are a software engineer and want to sleep well at night, you need self healing mechanisms and ways to guarantee consistency. But nothing is perfect, so there is always some uncertainty.
I also designed the full flow for the functionalities I built and made some changes to offload rate limit healing to Celery. I also fixed some very small bugs.
What else... hmm.
I set up my projects on a completely new server because we were doing tech stack migrations. It was fun. I had to take care of many things, but then I realized the project is almost fully self contained. Just add the environment variables and it works.
I think keeping things simple will become even more important now that AI is going to do a lot of the coding. You have to explicitly ask it to keep things simple. Otherwise when things fall apart, you will not be able to fix them easily. And at some point, things will fall apart.
I submitted all my self review documents. Let's see if I get a pay bump and a good bonus. I hope so.
I am trying to become a research engineer now. I am interested in inference infrastructure, and I think moving toward systems engineering would be a worthwhile pivot. With AI, we will have more time to experiment, and I plan to do exactly that. I am not very interested in training LLMs because I do not have that level of mathematical background, but I think I can engineer better systems. Making inference cheaper can add a lot of value to a company. If a company could build its own nuclear power plant, its electricity costs would drop significantly and margins would improve. More profit is always better, right?
Am I anxious about my career? Yes. Do I want to work on cool things? Yes. Am I qualified to work on cool things? Who cares. I will learn. I think it is time to take risks and maybe even cold email Sam Altman :P
I do not have much more to write. March was a bit slow. There were many production issues and we were focused on streamlining processes. This blog probably does not make me stand out as engineering talent, but I was mostly working on making things more stable and slightly faster. I did make one API faster though, and that felt good.
