I’m probably not the first person to write about the insane leverage that LLMs confer to engineers, but Stack Overflow’s 28% layoff really got me thinking about the future of human-generated data, especially in the context of a potential model collapse (whereby “models forget the true underlying data distribution” once they are trained on machine-generated data). I explore whether we are at “peak data” in terms of both the quality and percentage of human-generated data on the internet, how this might affect the efficacy of future AI models, and potential solutions/product opportunities that exist.
For me, in the pre-ChatGPT days, my workflow as an engineer in big tech when tackling a brand new project would be reading a lot of docs, Googling/Stack Overflow’ing when something inevitably breaks, and eventually getting something serviceable working. Now, ChatGPT is my tireless companion/engineer who consistently generates well-thought out solutions (and occasionally snarky responses).
I’ll illustrate via an example:
I’m currently using Google Firebase (and Firestore) for handling a lot of our backend business logic/data storage. Because Firestore is a document-based database, I asked it to generate an entire “data model” of our app (I’m migrating from vanilla Postgres). And it generated the full data model on the first try, even reminding me about the need for data denormalization.
It then proceed to help me generate rules for field level access.
To continue reading this article, click here.