Am I understanding it right that for each new text (tweet) you generate its embedding first, try to match across existing vector embeddings for all other text (full text or bag of words), and then send the text to the LLM for tag classification only if no match is found or otherwise classify it to the same tag for which a match was found.

Will it be any better if you sent a list of existing tags with each new text to the LLM, and asked it to classify to one of them or generate a new tag? Possibly even skipping embeddings and vector search altogether.

Yeah. But I think one problem could be if this list is very large and overflows the context.

I was thinking giving the LLM a tool `(query: string) => string[]` to retrieve a list of matching labels to check if they already exist.

But the above approach sounds similar to OP, where they use embeddings to achieve that.

I think your understanding is correct.

I actually built a project for tagging posts exactly the way you described.

I'm trying to like Bluesky. X has a large tech community and some deep thoughts/discussions on tech and philosophy matters, but I can't find similar people or groups on Bluesky yet.

On both Twitter and Bluesky, my best results at feed building have come from finding one person I find valuable, then looking at and following people they interact with and follow. A brief look at bios and recent posts helps. Then repeating this for additional interesting people.

I am fast to follow but also fast to unfollow if a person turns out to be a dud. An example of a dud is a person who is a cybersecurity expert but almost all their posts were about their travel: which hotels treated them well, complaints about flights, etc.

Bluesky also has a notion of lists, which folks seem to use. For example someone curates a list of people active in local politics and it’s a quick way for me to plug into what local activists and politicians are talking about. Again, I found it via one person I found interesting.

Search for "<topic of interest> starter pack"

It really depends what you are looking for. Try searching for some starter packs to seed your follow graph. The below site works well but the search is a bit dumb/literal and the packs are all user created so you'll want to check some of the accounts in the packs that come up before you blindly follow them.

https://blueskydirectory.com/starter-packs

Same with Mastodon for me. It's good, and the app is well done, but there's just no reason for me to use it.

Landing pages can still save time to validate an idea/concept, build a leads list, and with cold outreach even.

Dev tools or B2C: landing page (collect email) with app/product tour video or mockup screenshots

B2B: landing page with video or mockup screenshots, and 'book a demo' link/form (collect as much info)

Here's a good summary of the study: https://x.com/RubenHssd/status/1934968317627998534

In short, workers don't want full automation. They want partnership.

From the article: " The authors are currently running a new study to deepen the understanding of the dog-human relationship. Dog owners across the world are invited to answer the following questionnaire: https://tally.so/r/nPXKPb "

did the survey. very badly designed. asked wrong questions. ambiguous questions. and double negatives

Haven't noticed any billing issue. Maybe try reaching out to them at https://x.com/CloudflareHelp

If Stripe Radar (has a small fee per transaction) is enabled in an account, then an early warning (before chargeback) is sent via the `radar.early_fraud_warning.created` webhook event, which can be used for a pre-emptive manual review, or an automatic refund and cancellation. No dispute fee on a refunded transaction!

There are services like Chargeblast, ByeDispute etc. which also help avoid disputes and chargebacks.

Chargeback early warning ends this month. It won't be a thing going forward. (Might just be applicable to Visa)

Really! Is there an official mention anywhere by Stripe?