Posts

The Same Jailbreak Scored 1.00 and 0.15

In LLM safety, a number you can't reproduce—or whose judge and dataset you can't name—isn't a benchmark. So I built one that is.

Vercel-Style Deploys, Inside Your Own AWS

I wanted Vercel's simplicity but with my own AWS account. So I built it.

Sync a Slice of Postgres, Not the Whole Database

Because partial database replication shouldn't be a nightmare of manual mapping and brittle scripts.

I Pointed GPT-4 at My Receipts

Because managing money shouldn't require a finance degree or hours of manual data entry