Leo, blog is really hard to navigate

[PsychedelicEagle]

@Leo Gura 🤣

 

[Girzo]

@Leo Gura Yeah, it's images, not text. They are text-machines, LLMs, and ChatGPT is the worst thing ever, Claude does this type of job flawlessly with academic rigour.

[PsychedelicEagle]

Yep. LLMs excel at coding cause it is symbolic manipulation.

Let's prove this guy wrong, @Girzo! 

[Leo Gura]

There is no coding in a blog. You guys are living in a fantasy.

Edited September 15 by Leo Gura

[Girzo]

But AI tags it, and can give you any output, so you can integrate its output on your own terms. Everything on a computer is coding, you can even script mouse movements and keystrokes so the process goes quicker. You can do it however you want, and I won't believe the task can be anyhow done faster without AI than with it. Even if you want to tag everything by hand, you have to somehow then add that data to posts, and you can script that process = coding. There's always coding.

[PsychedelicEagle]

11 hours ago, Girzo said:

Even if you want to tag everything by hand, you have to somehow then add that data to posts, and you can script that process = coding. There's always coding.

Yes.

But even tagging itself can be framed as a series of coding and machine learning steps:

1. Scrape the blog to download all posts locally (or fetch from the database, since you have access);
2. Cluster the posts to define the tag taxonomy — a set of labels of interest;
3. Classify each post into multiple labels.

The nice thing about this is you could keep the scripts to automatically apply labels to every future post. Or to backfill existing posts if you wish to create a new label.

You could also easily ask the LLM to generate insightful visualizations about the blog.

Edited September 16 by PsychedelicEagle

[PsychedelicEagle]

Asked it to download the last 100 posts and create a scatter plot. Funky cluster names haha (if I had scraped all posts it'd likely be better, but I didn't want to bombard the server for a simple experiment).

I also asked it to create an interactive Marimo notebook where you can hover over and click on posts to visualize their contents below.

I also guided it to extract the contents of images (it used Tesseract) so the quotes (which are images) are transcribed to text. So that the ML model mapping post contents onto the vector space (in which clustering takes place) has the correct information (the actual quote contents).

All this took less than 10 minutes and 4 prompts on my end.

Edited September 16 by PsychedelicEagle