Good Day ya’ll!
So this was the festive week over here in India with Diwali over the weekend (the reason I am writing this on a random monday morning). It is just beautiful to be wandering around in the streets at this time of the year, with the cold air brushing past and decorations all around. The best time to visit India in my opinion.
But here’s the recap.
Readings
I like collecting books. Reading? Not so much. You see, I am a master procrastinator, I can put off of doing work until the last survivable moment and its the books that take the toll.
Whenever I come across a book that seems nice, or if I recall someone suggesting it, I just hand over my money. I end up reading about 10 pages until something important comes up and the book gets up getting tossed up somewhere in the shelf. Then I go out, see a book, and the cycle continues.
But, I am changing things from now. This week I got a ~120 page book which I came across from twitter: “Steal Like An Artist” by Austin Kleon. I am only half-way through, but it is an amazing book so far. Glad I picked it up (tbh, I ordered from Amazon rather than a store).
Projects
There was also OpenAI DevDay this week, and I got to experiment with a bunch of stuff.
1. Video Inference
With GPT-4 being multimodal, I followed a guide online to try to bring to life a 5 year old project of mine: Human Activity Recognition.
Back when I was a sophomore student, I worked as a research intern at a university in Noida, India, during my university’s winter break. The aim was to categorise short videos into different categories (eg: walking, combing). At that time we tried to use a “Video → Frames → Imagenet → RNN” sorta model, don’t really remember the specifics, but it performed poorly, <60% accuracy I guess. It seems that “Video→Frames→ GPT-4” model would blow the SoTA away now.
With this guide, I modified it a lil bit and tried to replicate it on a short game-clip from FIFA. Well, I got ratelimited lmao. I even got rate-limited on the original video clip the blog used. Seems like I need to spend more money before my quota increases, until then I’ll try to just use the original HMDB-51 dataset just to keep my sanity.
2. GPTs
Also, GPT4 has a 128k context now, enough to handle a ~200 page book at once. I think I can make a startup to let anyone upload a book and ask questions from it.
Surprise! OpenAI already has that, customGPT as they call it. So I downloaded the epub of “Almanack of Naval Ravikant” and created a custom GPT with it. It sucks. While it does work, and it does answer the questions, It is too slow. It uses code interpreter to parse the text EVERY TIME SOMEONE ASKS A QUESTION. Too inefficient. I think theres an opportunity here. Maybe just fine-tuning will work. Lets see.
Here is the link if you want to try the customGPT though: NavalAlmanacGPT
Games
I started playing Starfield today, the space nerd in me was just too excited not to try it. I suck at the controls though. And the UI. And the Navigation. And flight controls.
Cyberpunk also released the Phantom Liberty expansion a couple of weeks ago, but the game keeps crashing whenever I go to a ripperdoc to get the Mantis Blades. Man, what a drag.
And FIFA. I am addicted to this terrible game, but it is the only one where I can just tune my brain out to think about all other things while my muscle memory autoplays the game, so it kinda still helps me to clear out my brain at night lol.
Heres my team:
Can you tell that I am a Man Utd fan yet? lol.
Up Next
Not really sure how this week shapes up tbh, I am still busy with organising docs for grad school. I’ll try to ship the pending changes to wanderfusion this week though, and maybe spend some dollars on advertising to gain users. I don’t expect much clicks though, I just wanna try this route once. It would still be up as a side project, I would want to use it to plan my trips exclusively. There must be a few people who like to plan like me, maybe eventually I’ll find my niche.
Until Next Week!
Adarsh