Udayan Kumar My public notes

Small Models, Big Savings!

TL;DR: For narrow, repetitive NLP tasks at high volume, a small fine-tuned model can be much cheaper than a frontier LLM. The savings can be large, but only when the task is constrained enough to justify the extra engineering work.

Over the last few months, I kept running into the...

Prompts vs. Code: The Cost of No Compiler

Context: People are writing no code agents where execution instructions or prompts are given in English. These instructions are like pseudo code with conditionals and even loops (repeat step 1). This is an actively evolving area and here are a few drawback that I have observed when compared to...

Alaska trip

TL;DR: Alaska trip log by my mom in Hindi

1st August

अलास्का जाने का प्रोग्राम बना तो बच्चों के साथ हम लोग भी काफ़ी उत्साहित थे क्योंकि अलास्का जाना एक सपने का सच हो जाना जैसा ही था। अलास्का USA का वो राज्य है जहाँ...
  

Measuring Success in data Science

TL;DR: Success in data science is inversely proportional to the number of times the same steps are repeated.

Gartner says 85% of the big data projects fail (1). They also say only 20% of the analytics insights deliver business outcomes. VentureBeat says 87% of data science projects never make...

Working with a Neato Lidar

help A lidar measures distance to the nearest object in-front of the sensor. For e.g., a lidar mounted on a robot can be used to detect obstacles in the path. A lidar is more expensive but provides better accuracy and range than other methods such as