Weekly Notes 12/2025

by Thejesh GN · March 21, 2025

We returned from Langkawi on Wednesday after a week-long vacation. It was a good trip. I had nearly a thousand unread emails, both work and personal when we returned. Around 80% of them are non-critical - newsletters, alerts from GitHub, issue trackers, etc. But I believe in inbox zero, so I had to read the 20% and mark the rest as read. That's a lot of work. It took me almost a day.

Echo finished his Gotcha Day cake in no time.

Also, this vacation was good for my neck, even though it wasn't planned for that. I see and recognize how being away from the screens has helped recovery. Pain intensity has reduced considerably; it is at zero while wearing a cervical collar and very low when I am not wearing it. I am still wearing a collar to help me recover. I have yet to start on neck exercises, which I plan to do next week after the next consultation with the doctor.
It's Echo's gotcha week. I can't believe it's been five years. He is a fantastic dog and an amazing elder brother to Uma and Pathu. Anju baked a cake to celebrate for both Echo and Path. It took 30 seconds for Echo to finish it.
I have shut down my Mastodon instance and go back to social.lol/@thej
NMG application for 2025 is closed. We have received 100+ applications for 4.5 Lakh rupees this year. In terms of grant money, this is our biggest year. I will start going through the applications on Monday.
An interesting thing I noticed this week on Grammarly is that when it fails (for whatever reason), Instead of showing an error message, it defaults to telling you everything is correct—a misleading false positive.
AWS Bedrock (LLM, etc.) serverless APIs seem like a good place to access running LLM models, especially if you are experimenting and don't have enough money to run a dedicated GPU-based machine to run Ollama+Models. Sadly enough, it doesn't provide OpenAI-compatible API (now defacto API standard for chatting with LLMs on the web). However, Amazon provides an OpenAI-compatible RESTful API translator as a docker image; you can run it or use litellm proxy. Then, add Bedrock to your project or your Open WebUI instance.
LLMs are evolving fast, so I'm trying many tools, formats, and standards. A few have stood out despite the churn—like Ollama, Open WebUI, lilellm proxy, and the OpenAI API format. For me, all the magic happens behind the APIs or downstream. I love standard interfaces, whether it's UI or API. What's the AI toolset that has stayed with you?