@cloudofsienna Seems OK --> Wow they are loud though!

Notes on my blog at

Today (well, recently) I've been test driving Claude's built-in metaprompt, and even taking the output to other models. It's zero effort, huge lift!

If you drop one of these into your webapp, you could offer users: - low-latency validation - suggested tags - error message help which I claim are better UX without a remote server round trip

You read that right - I registered just for this 🙃 instead of something more sensible like submitting to cursor dot directory

Today I test drove @stackblitz 's bolt dot new to make this silly turkey tap page. It works amazingly well and was able to fix it's own errors. I used almost 200k tokens, which is most of the daily free tokens. https://turkeytap .netlify.app

@KaichaoYou Boring post since the tooling makes things os easy!

@karthikkalyan90 @NaderKarayanni I didn't fully grasp it. Best of all, I understand now that lowering temperature amplifies probability! So you can't exactly look at a probability near 1 and conclude that the model is confident but you can compare it to other probabilities.

@NaderKarayanni Write up Inspired by

@karthikkalyan90 Still experimenting. I tried "What are the top selling blue products for the top 10 customers by revenue" which is a bit more complex because it requires an inner query. 1.5B failed that, but 32B succeeded. (Maybe I should seed this database and actually run the responses 😀)

Here's a brief write up

Today I'm test driving hybrid full-text search with (again) sqlite-vec. Example below with 3 different approaches. I enjoy it being pure SQL, although it is a bit (um) lengthy!

Today I'm test driving sqlite-lembed, a SQLite extension that adds functions for calling a local embedding model. It does what it says ...but it leaves me wanting to dig in to implement MLX functionality to make it faster on my machine.