Technology Update Oct 16 2026

I have been working on a bunch of pending technology things this month.
Some of this is new, some old.

Use a dumb llm model. Yagni

I finally purchased the $20 OpenAI subscription. My only other LLM subscription that I pay for is the google-antigravity one; I pay for it because it comes with storage that I get to share with the family.

Using antigravity was fun, but I hated the fact that I could not use opencode or something like that. The model is not too smart and I loved it for that matter.

The gemini models are not the best (at the time of writing this) but then again, I am rarely doing something that needs the SOTA models.

In fact I enjoy using a model that is less capable because this means I get to look at its output and get frustrated. This makes me more vigilant during the planning phase. Cunningham's Law is real. My current workflow has a plan phase where I sit with the LLM and chart out a plan in great detail and the second phase where I let the LLM build things. This is strictly copied from the culture at Cloudflare where the accepted model of working is to write a spec first before writing even a single line of code. This is from the pre-LLM days.

I start with a small description of what I want and then I watch the LLM over-engineer, over-complicate and over-architect this. Then I spend time manually cutting the scope. This is where the magic happens. After a lot of this is not needed and please do not overcomplicate I arrive at a final draft of the plan.

Now another cheaper model takes this plan and starts building. Once done I look at the code and validate. :)

I spend time refining the plan. Then I spend time validating the output of the plan.

This means that I am on top of the planning as well as the execution. I like it this way. More importantly, as an SRE, this is the expectation I have from myself.

I have been using LLMs to learn new stuff as well. LLMs make a great sounding board for things that I am struggling with. Again none of this is SOTA. I just need something to talk to about a weird syntax that I am witnessing for the first time. The LLMs are also great at explaining stuff; though independent verification of something remotely esoteric is often needed. For this I use Google Gemini and ChatGPT side by side. I am happy to report that this setup is working well for me.

I can focus on what I am learning and not get lost in the vastness that is out there; though sometimes I miss getting lost in the beautiful vastness.

It is also doing a lot of the grunt work. For example, my Linux laptop keeps crashing from time to time. Something about the AMD GPU and Xorg, etc. I have a script that I use to collect logs and piece together something. I have done this multiple times and now the LLM does this. It is able to comb through the piles of logs and piece together something meaningful which I get to see sometime later.

Nix package manager

I fired up the Raspberry-pi. It took me a while to do this. I was needlessly procrastinating on this. All it took was one evening of work.

I decided that I want to manage the Raspberry Pi with the Nix package manager. I used an LLM to cook up the config, but I have not yet executed this. The reason being that I do not understand the system well enough to execute it on the Pi. It will take me a few days to get up to speed before I feel comfortable with the Nix setup.

I plan to host some stuff on this setup.
For now I am running tailscale and a Silverbullet server on this.

Silverbullet is a note-taking application and it has a PWA which gives you full offline capabilities. This makes it great for using it remotely on my phone.

lazygit

At work most code is LLM generated and I would use VS Code to look at the diffs and review things.

VS Code was having a hard time keeping up with the large-ish codebases and would often just slow down. I do not know if this was VS Code or just the Git extension, etc. Anyway I chose not to debug this.

Since my primary reason for using VS Code apart from writing notes is to look at Git diffs, I decided to do some research and came across this beautiful tool called lazygit. This brings me back to those Magit days from when I was on Emacs.

So I had the LLM set up lazygit on my nvim and computer and I am set now. I really love how easy the whole experience of looking at code is now.

I still sometimes go back to VS Code when I am navigating a large codebase because I cannot live without the click-to-jump feature and the setup on nvim is clunky for me.

Herdr and Remote Control

I set up a Herdr on my laptop and I am using Tailscale to tunnel the Herdr web interface to my mobile phone. The Herdr interface on the mobile phone has a PWA which makes it easier for me to interact with the Herdr that is running on my computer from my mobile phone.

I want to move this to the Raspberry Pi as well because I do not keep my computer running all the time and I rarely do work that needs constant remote guidance.

I am keeping an open mind and finding things that I can do with this setup but I rarely come across something interesting.

Okay!!

Bye!!

That's it!!