Shiveen Pandita

How I use AI

Let's get a few things straight first.

I dislike AI.

There are many rational reasons I can put forward on why, such as its a stochastic word box with an ability to override our judgment with its confident tone. It's a poker machine of banal ideas mixed in from a corpus of open source books and the internet, skimming the words but not understanding the reason. It has an eccentric bubble of overenthusiastic snake oil sellers that seem to associate with it. Last but not least, this part inherently more dangerous than anything else, is that it is indeed extremely useful in a subset of scenarios and initial conditions, which are hard to predict.

Like any big technological or intellectual leap in human understanding or interfacing of the world, AI feels like a big jump from how we used to interact with the simple oracle like text box. We graduated from the simple HTML text <input/> of Google search days to a sparkling text box that does much the same but with less degree of freedom on discernment.

Alas, this post is not about my dislike.

I did not want to be the only luddite left behind, perfectly encapsulating the grumpy old man wishing the halcyon old days of crunching reading 17 stackoverflow threads to get an answer to a simple question should come back. Rather, I wanted to embrace this new interfacing with the world knowledge.

Committed now, I decided to buy a Claude pro subscription as my more AI-inclined tech friends reckoned it was the more elite highbrow option (I do not understand why, yet). After being a few hundreds of dollars poorer, I decided to put it through its paces so to speak and after a few months of using it, here are the things I use AI for.

This for situations where I know I will follow the hivemind and don't care about confirmation bias.

These are things like "what's a good butter brand in my local supermarkets?" or "What's the expected split for ETF investment for someone my age and country?" I find Claude's research mode extremely useful as it does offer vetted sources, places where I would have checked myself and more or less come to the same conclusion.

I also use it as a first pass builder for things like travel itineraries or recipes. I like cooking but hate parsing through cookbooks for simple stuff. So sometimes I just ask Claude to give me basic steps for things like churning butter in our stand mixer.

Admittedly, these are all tasks I could do myself, perhaps even better. But, as I said, my error margins for these asks or so high that even if I get put in the hallucinating answer part of the error graph, I will be okay.

As a junior dev with photographic memory #

I have used Claude code. I like Claude code. I love the way the text moves in my terminal and makes pretty little graphics while doing so and says silly little whimsical things like "Photosynthesising" or "Discombobulating". But, here's the thing, it's too confident, and it does not think beyond a layer and half of the problem.

Perhaps it's a prompting issue or in general imprecise instructions, but regardless, I am not baby sitting an "AI" to give me a perfect answer. I could build software before, and I can build software now—the only problem is that now, I also have to read while I write, which is not sustainable for me.

So for me, I have settled on the following pattern of working with it:

Most people telling you that AI can do more than this, are perhaps a degree or two separation away from what real software really is or are prompt gods or have an ulterior motive, possibly all three combined.

As a technical assistant #

One of the things I have recently started getting into is homelab'ing. I have a setup at home with a BeeLink CPU and a QNAP NAS. I run portainer to manage containers which range from media management and handling to productivity software.

One of the more cumbersome things of setting up a homelab is that you have to read the documentation of a dozen of different apps and docker containers, figuring out the right folder, networking, permission and metadata structure. This process can get extremely tedious.

I use claude to do the dirty work for me. I heavily use the Claude projects feature here as it lets me compartmentalize the work, and I constantly feed it text and reference documents to keep track of.

This works really well for me, most times. Sometimes it forgets the instructions I just gave it, such as using a particular group id for future docker compose files, or it hallucinates a docker image that just does not exist (and never did).

But more or less it works really well, and I can smash out running services quickly whenever I feel like it without spending a whole evening reading getting started pages or blogs on how to do things.

In the end, I think I get good use from the AI. I'm still not convinced the pro sub was worth it, as there seem to be a few VC-funded AI apps out there including the unfortunately ubiquitous ChatGPT. However, it has given me enough value to not make me completely discount it.

I think AI is a great technological leap if we separate the hype from the truth and use it for what it makes sense -- accelerating low-stakes mechanical pieces of work and using it as a way to lower the inertia to start more critical pieces of work.

↑ Top