Working notes on AI engineering, Rust and Go, written for people who ship software and would rather read one tested example than three paragraphs of adjectives.
pgxpool defaults to max(4, NumCPU) connections. With 64 goroutines running 5 ms queries, 4 connections gave 608 req/s and 64 gave 8,879. With a CPU-bound query, nothing past 8 helped.
I have built lots of websites, apps, browser extensions, I find the most difficult problem is not to build the product but to promote it. To promote and learn how people react to my product, I need to collect data from many different sources like hacker news, reddit, X, bluesky... So I build an service to track my product and put them all together and I then have a consolidated place to know what's going on with it. This ends with an open source project which I think would benefit others. Also you can contrinute to add more sources and give feedback on what can be improved.
The service is self-hosted and all you APi keys and AI keys are saved on your machine, so safe to use.
The repo: https://t.co/6aLn6qwHrT
gpt-6.1-terra is $1/$5 per million tokens against $2/$12 for gpt-5.6-terra. Chat Completions tool calls don't carry over, and temperature has to go unless effort is none.
A Go package that wraps driver errors with %w lets callers match sql.ErrNoRows and *sqlite.Error; after a driver swap those checks compile and return false. Translate at the package boundary.
Fiber v2 decodes Brotli request bodies in c.Body(), and a fresh go get of v2.52.16 still resolves andybalholm/brotli v1.1.0. The port's CVE-2026-111497 fix is in v1.2.7. govulncheck reports nothing.
OpenAI's 12-hour prompt cache write costs 2.5x input, so two requests 6 hours apart cost more than with the 30m default. It pays from the third reuse. Set request-wide, a chat cost 68% more.
date('now','end of month') is new in SQLite 3.54. On 3.53 it returns NULL, no error, so a WHERE built on it matches zero rows. And 'weekday -0' on old builds moves forward a week.
Claude Code reads AGENTS.md only if no CLAUDE.md or CLAUDE.local.md exists in your working dir or any parent, ~/CLAUDE.md included. ~/.claude/CLAUDE.md is exempt from that check.
DeepSeek V4.1 Flash takes seven reasoning_effort values. In DeepSeek's own prompt encoder, medium, high and xhigh all become the line "Reasoning Effort: 75". Low is 50, max is 100.
The $20M limit in the Muse Spark Open licence counts the gross revenue of your company and its affiliates. It only applies if you sell access to the model as an API.
openai-go v3.78.0 marks o4-mini deprecated, but not gpt-4o-mini, which OpenAI shuts down January 14. On our sample service, staticcheck flagged 1 of 9 doomed model IDs. How are you finding yours?