Thread by @ItsTwoBags

Twitter released the source code for the Recommendation Algorithm.

In total, there are 1440 files and 186K lines of code to review.

I this series, I use ChatGPT (GPT-4) to analyse the entire codebase and produce best practices we can follow to improve engagement.

Let's go! 🧵

Firstly I'd like to share challenges using ChatGPT and how to overcome them so you can do this yourself!

1) No, ChatGPT cannot access the web, so you cannot give it the URL of the source code and say "read it all"

2) ChatGPT isn't great at remembering full conversation history

Fixing 1)

You act as "access to the web" and provide it with:

A) Directory listings
B) Copy and paste individual files

This is as good as ChatGPT having access to the web, just slower and more painful than saying "read it all" (which today, AI thinks is April Fools joke 🤣)

Fixing 2)

Prompt-engineering basically means "asking a good question to get the best answers".

I designed a prompt that I use every time I reply to ChatGPT to keep it focused. It could be improved, but it works. Here it is.

Replace the last line with whatever you are giving it

What did we find?

First, I gave GPT-4 the overall README file which gave it context on which components to look at first.

Summary of the main algorithms to understand:

1) Networking/who you connect with

2) Your "tweepcred" (reputation)

3) Your tweet content

GPT-4 wanted to look at "heavy-ranking", which is the AI/machine learning component.

This isn't too useful, as Twitter haven't released the AI model with the source code.

Hopefully they will do, it's been requested:
github.com/twitter/the-al…

Next GPT-4 wanted to look at "tweepcred". This algorithm calculates Twitter User Reputation, and will be the focus on this thread

I provided the directory listing and the README for tweepcred

It described the code files, but read continue the thread for a summary of the code👇

I gave it the code for each file, and here's what we learned about behaviour that optimises (or damages) your tweepcred.

1) Follower-to-following ratio actually matters!

- If you're following more than 2500 people and have a low number of followers, this hurts your reputation.

2) Quality over quantity is a recurring theme.

- Follow and engage with people who have a good reputation.

- Create less, higher quality tweets over many lower quality tweets.

3) Give and you shall receive

- Engaging with others who have high tweepcred - mentioning them, retweeting their content, replying to their tweets etc, improves your tweetcred

- Being consistent and active gives the algorithm more iterations and chances to discover your content

4) Verify. Don't get suspended.

- Verification does add to user credibility, and will have a much bigger impact in the future when, as @elonmusk informed, the "For You" page only includes verified accounts (April 15th)

https://twitter.com/elonmusk/status/1640502698549075972

@elonmusk 5) Things might change.

- Your reputation can fluctuate daily depending on your activities

- This algorithm will change over time, we should see the source code change on the GitHub repository

@elonmusk Using GPT-4 for this analysis produces far, far better results, but GPT-4 is limited to 25 messages every 3 hours.

So, in another 3 hours I'll do the same analysis but look into optimising tweet content.

If you're interested in this, follow @ItsTwoBags so you don't miss it!

@elonmusk That's all for "Tweepcred" and profile reputation.

If you found this useful please retweet the original tweet below 🔁👇

And follow @ItsTwoBags for the continuation to this series!

https://twitter.com/ItsTwoBags/status/1642183667157192704