Ian Lurie
banner
ianlurie.bsky.social
Ian Lurie
@ianlurie.bsky.social
1.2K followers 420 following 570 posts
Cynical idealist. Digital marketer. Dungeons & Dragons player. Writes over at ianlurie.com. he/him
Posts Media Videos Starter Packs
And I say that from a place of love. We've been gaming together for 5 years and I wouldn't change a thing. Except a few NPC names.
That was a single 15-minute period, if I recall. Our games are like a clown trying to referee four trench coats filled with raccoons.
Again: Crying myself to sleep.
Lots of people don’t intend. And yet here you are, not apologizing. “I don’t mean it” is like “I’m sorry you misinterpreted”
I’m Jewish, and I’m adamantly opposed to Israel’s actions in Gaza. But you’re implying that because I’m Jewish, I support the genocide. You’re antisemitic.
Figured that out about 20 seconds ago. Did you know there are these websites that'll let you search the entire internet for a photo?!!!! It's amazing!!!
It looks like a cross between a tire lever and a lego tool...
Not the Gorg!!!! I need to build something on a more scalable platform. A few sites I work on are too big for Sheets. That’s not a flex - they need help. For now, though, Sheets are easy.
Me, putting Claude Code to good use.
Aaannnddd I just realized I fell into the SEO Trap of calling a "topic" a "keyword."

Which is how I started all this confusion in the first place.

I'm comparing document embeddings to embeddings for TOPICS. TOPICS.

Sorry. My bad.
...and named a model after a Sesame Street character.

But that gets way beyond my skill level. So I'm still just extracting embeddings to find near-duplicate content.
...but who figures out what "relevant" is? If only someone had built a big computer program that figured out the relevance of documents compared to a query...
To compare page embeddings to topic embeddings I'd need to hand the topic to the same embeddings model. I've tried passing everything to OpenAI's embeddings extraction but I haven't been thrilled with the results. I suspect I'd have to find other "relevant" documents and compare...
I should've been more clear. I'm not using Claude to do the comparison. I don't trust Claude to do long division...
I'm calculating cosine similarity using embeddings from the site only, not across multiple sites. And I'm using a consistent method to extract the embeddings. Assuming the embeddings are based on counting the number of times the letter "A" appears on the page, that should work, right?
Not sure we're talking about the same thing. I'm a lowly History major, but I thought if you generate embeddings across an entire site, then use cosine similarity to compare those pages, you'd get a decent predictor of similar pages.
If I'd taken the "my kids are AI" approach 20 years ago it would've saved me a lot of aggravation...
I’m just experimenting with comparing embeddings to keywords. I really use the embeddings to find similar content that’s not the typical near-duplicate.
AI is an assistant. Not a practitioner.

All the feckless AI "thought leadership" that has major corporations firing exceptional talent start with one wrongheaded idea: That AI can DO things.

It can't DO things. It can help US do things.

Please, Alfred implores you. Figure it out.