How much to offload to AI?

Idk.

Things have quickly changed, but I think I’m becoming a bit more settled in how I want my work to look and feel like going forward.

The activities I enjoy most are learning and building. I never found it intrinsically validating to build something as fast as possible just so I can say I built it. This is probably the equivalent of buying a Lego set that’s already completed. The emotional attachment (joy, and probably corresponding pain) is associated with the building.

The trouble as always is working with others. In one or a few shots an LLM can create a C- solution to a problem faster than any human can. C- here means getting something that works and looks like it solves the problem: this website I’m currently writing on, for instance. If a team values speed over everything, the A- solution that took you two weeks to create will never see the light of day.

The question I’ve been stuck on is: how do I go about reconciling the two?

Maybe they don’t have to be reconciled.

For my own work (typically ML), I think the flow still remains: understand the problem, create a baseline, and start building solutions that try to beat that baseline. I think these steps happen in a loop, though. A lot of the understanding of the problem and designing of a solution happens while struggling through the data and the problem.

That being said, if the baseline is “we use a rolling average plus our instincts” and we need a C- working solution, a 1-2hr LLM session (shared understanding of the problem, then plan, then execute) would get us there.

I think a larger fear of mine is that the more I rely on terminal prompting and not actually writing any code, the further removed I get from actually building. I guess there’s something to be said for being able to build full ML systems (multiple models, feature store, deployment, monitoring) extremely quickly. But the speed doesn’t mean the models do better than a rolling average, and it doesn’t mean the system isn’t built from poorly referenced examples.

I’ll probably have to revisit this, or form some new thoughts, as I start building more complex systems. I’m at peace outsourcing work to AI that I’ve done a bunch of times over before. The thing that’s remained consistent in my journey, though, is that I really enjoy building and learning new things. This probably means a lot more struggle and diving into the details at the sacrifice of speed. I’m at peace with that.