Hacker News Reader

Comment by magnio | Hacker News Reader

For some users I searched, the most similar user is a throwaway account, which is somewhat eye-opening and unnerving.

I was just about to point that out! I tried a friend's account and the second match was a throwaway that I know for a fact is theirs.

That’s a really interesting and unintended use case…

How did you gather the comment histories? Would you mind sharing a copy?

See description at the bottom. We used the Hacker News API to pull data into BigQuery.

From there we ran them through an embedding model and indexed the embeddings in Pinecone.

The actual similarity search is done with Pinecone. (https://www.pinecone.io)

Using Google BigQuery is one way. This comment might be of use:

https://news.ycombinator.com/item?id=25075318

> A reminder that BigQuery (as used in the query in this link) is the best way to play with Hacker News data; don't scrape HN data manually! The `bigquery-public-data.hacker_news.full` table appears to be up to date with the most recent HN data as well (table last updated today). However, I'm not 100% sure the query is correct for unilaterally getting all links, as running the query on the full dataset returns the same results as running it from 2006-2015. And I value my sanity enough to not fuss around with the regex.

Two of my matches were <username> and <username>2.

Huh, checked some of my older accounts and none of them matched each other. So I must be doing something right.

In fact, considering that unknown third-parties freely gather such similarity scores and correlate accounts, across different sites—by now it's a given that one's alt accounts have to adhere to different stylistic choices.

Innit?

Wow, this works seriously good at some level then.

Reply on news.ycombinator.com