DSLs

Are there spaces for DSL anymore? With AI you can do the grunt work and also if you just wrote your own it’s probably not in the training data.

I’ll take the example of my intern project at Google: Meterstick. Way back in 2016 we had the problem of having write a lot of analysis code in pandas which sucked. In particular adding slicing, comparisons, and confidence intervals to analyses were terrible.

In comes the pimply-faced intern: me! Over the summer I write a package which distills previously 100-line monstrosities into¹

Ratio("clicks", "impressions", name="CTR") |\
    PercentChange("experiment", "control") |\
    Jackknife() | compute_on(data, ["country"])

So much easier to write and review! I was rather naive at the time so I thought the reviewing benefit was just icing on the cake. Now after having managed my own hectic team: it is the most important thing! Knowing with near certainty that the result is what it purports to be at a glance is even more important the faster the results are to write (and hallucinate).

Now the tricky bit is getting it into the training data. I suspect the vast majority of meterstick usage is at Google so it’s not going to be scooped up anytime soon. Similarly any new DSL is going to be slow to diffuse. I suppose I could copy and paste the documentation into my chat: but at that point it’s just easier to write it myself.