Originally published on my Linkedin account as a spontaneous post link
ek chis hai, jo ai ke liye apply hua hai… and humans ke liye bhi apply hota hai
purane zamane mai hand-picked data ke sath feature engineering karke, model train karne ka paradigmn tha, especially in computer vision BUT ab, its all messy.. full internet scale data, openai used CLIP, to train model on image/text pairs over all internet
the data quality is poorer, than crowd-labelled image datasets, but the sheer scale of this dataset is so much higher
its same for humans: clean processed high quality data = roadmaps, courses, polished data seemingly low quality data = random walk.
training methodology: raw messy large scale training followed by finetuning for specific tasks. you give high quality data at second stage
same for humans: raw messy random walks, explore everything you get, jump into random ships, without asking ki yeh padhai mera koi kaam mai ayega kya but, once you have gone through volumes of code, and papers, and books, and online material, and videos and random ponderings then before the interview/attempting some specific task you can see one perfect cheatsheet, and finetune yourself for any specific task
cheers to dopamine-driven random walks!
references: