Hacker News Reader

Comment by nzmsv | Hacker News Reader

ML today is what a website backed by a database was in the mid-nineties. This kind of CRUD website is considered boring today, but making one used to be taught in an MIT course (https://philip.greenspun.com/teaching/one-term-web)

There will soon be ML frameworks that are as easy to get started with as a basic web app, and most people will be using one of them.

View on HN

Yes! Most developers today don’t think about hash joins, write ahead logs or b-tree indexes but we easily download postreSQL, tweak some settings and get a functioning database. Of course as companies scale they will end up employing experts to tune their databases for max performance but they work well out of the box. Good machine learning tools should be like this too: good performance out of the box and easy to use for non-machine learning experts but highly tunable and configurable for advanced use cases.

With pure crud you don't and did not need any code since the early 80s (there have been nocode crud generators since the early 80s); it is always crud+something where something is human and usually something custom which requires programming. For ML this something might be easier or impossibly hard: it is already quite easy for someone who knows almost nothing about the theory to utilize ML that does stuff that could not be done 15 years ago. However, depending on what 'something' is in the ML case, it might be impossible for even almost-experts to fill in: an expert might be needed at that point. This is not the same with crud where most 'something', albeit usually not codeless, can be done by very mediocre coders. What I want to say: the gap between 'works with a few clicks' and 'custom code' is usually larger with ML as the expert level needed to reach it is concerned.

I do not think that gap can be closed soon.

Huggingface is close, for nlp algos. Add an abstraction layer, or plug n play integrations with web app builders and it becomes very accessible.

Colab et al allow very complex methods to be run by rank amateurs, which gives people a self learning path toward more sophisticated uses.

Cogview and dall-e and clip are revolutionary for image production, and video is close. Music transformers, synthetic voices, and other content can be thrown together to produce brand new styles of art.

Between the ever more general capabilities of large text models and increasing mastery of media synthesis, ai is on the threshold of making the world really weird, really fast. I hope the next 10 years feel like the 90s with these technologies maturing and expanding our horizons in computing and entertainment .

That Phillip Greenspun book changed my life.

Reply on news.ycombinator.com