Nothing Was Asked

My favourite tool in my skillbox is asking questions. It's a four layered process.

1. The inventory layer.
"What is even happening here?" List down every possible axis the problem space depends on. You are not going to jump directly into finding a solution instead you will just map the problem space.

2. The levers.
"What are the things affecting our problem?" List them down. Keep "formulation" as a premium lever.

3. The priorities.
"Which of these is a whale?" Find a way to rank analytically when you can't have them measured.

4. The constraints.
"What am I not allowed to do?" "What does moving my current lever cause to other factors?"

Measure. Repeat until you find the right combination of the levers.

This is the combinatorial and exploratory method of finding a solution. Reformulation is not -- that's transformational. Questioning the framing of the problem itself, looking at it through a new lens. The first two can let you learn something totally new in a short period of time, and be moderately competitive under pressure.

A few months ago, I managed to learn enough kernel programming in 16 hours to get 10x the performance of PyTorch for QR decomposition. It was a trial imposed upon me by a few folks who happened to be at OpenAI earlier.

I was told to work with LiteRT as part of a project at IIIT H. I was curious about how NPUs work, as they are not something talked about frequently on the internet. I ended up reverse engineering the Qualcomm NPU compiler, and figured out a few of their internal mechanisms. I spent some time writing an article on it. It was well received, and I later received an email from a senior director of Qualcomm Netherlands that my article had been shared internally in their team.

Both of these, nothing was asked.

If you have to tackle something, reverse engineer it. If you have to get a job, target the places you want to work, reverse engineer their problems, and show that you can solve something of that level. Do what people want before they ask. Be retrospective.

Psychohistory is a thing. You can predict the future by looking at the past.