Same system that powers docs.datadoghq.com/ddsql_refere...
09.03.2026 02:29
👍 1
🔁 0
💬 1
📌 0
intuition I’ve developed recently (mostly based on Opus):
- answering something from model weights? still some risk of hallucination
- answering something based on info in context window? almost certainly did not hallucinate
wrote up some lessons from what I’ve been doing at work: www.datadoghq.com/blog/enginee...
summary: lol
in case you were curious what the "thought process" was:
genuinely think that Thorsten Ball is one of the most interesting+thoughtful people writing about AI but also every now and then I get the impression his brain is being cooked on Twitter registerspill.thorstenball.com/p/joy-and-cu...
Been finding this quite useful this morning; using it to build+check my understanding of a big rewrite of a personal project github.com/DrCatHicks/l...
love seeing that Rust brown in a GitHub link preview
tried it when it was piloted here in 2024 and I was disappointed - crispy onions and sauce were both meh. hopefully better in the second coming www.mcdonalds.com/ca/en-ca/new...
did you see this? shame he deleted them bsky.app/profile/segy...
it's genuinely tough which is why I keep trying to explain it lol
and I think "coding" is slowly turning into "most things people do on a computer"
yeah, I should probably have Claude add systemd timer functionality to github.com/rgwood/syste...
like, the bare minimum is that it's easy to:
1. see past runs (including their stdout/stderr)
2. run a scheduled task ad-hoc (for testing and/or recovery)
3. see/configure which user a task runs as
perhaps this is a skill issue but I've always found scheduling things to be a pain on Linux; cron has too many weird gotchas, systemd timers are better but require a lot of boilerplate, neither has a decent GUI or TUI
Greg Egan's Permutation City has been on my mind lately. Touches on a lot of "what happens when computation is scarce?" questions that feel relevant today, it's a good book!
immediately thought of several horrifying ways to solve this and realized that people will almost certainly try them out
I think so. Gonna bite me eventually but there is *so* much friction with permission approval+management
that was a fun time!
not sure how I feel about this trend of company stock price ripping face after announcing massive layoffs
ssh is an obscure but widely-deployed command. It stands for Secure Snake Home and was made in the 90s to securely play snake online
I made a massively multiplayer backend for it with support for thousands of concurrent snake players
ssh snakes.run to join!
Over on dead-Twitter, @geoffreylitt.com asked the following question last week:
"I desperately need a Matt Levine style explanation of how OAuth works. What is the historical cascade of requirements that got us to this place?"
Here's my attempt at an answer: leaflet.pub/p/did:plc:3v...
the reason I know this word is that I once had to cover up a zit before a first date and then I found out that the makeup I used causes zits!
my luke warm take is that we tend to over-extrapolate the speed that things diffuse into widespread existence
I actually kind of wonder if there's just a mediumterm world where only ~10% of people actually care and deploy these tools to any useful effect, and they become a wizard class of employee
tools.rmv.fyi
I am especially in love with all the color tools
We'd love any moves toward standardization for tool search and client flexibility w/ tool results (ex: write to disk or pipe to jq instead of sticking the whole result in context window).
(also, hi! we shared the stage at MCP Night but I didn't get a chance to introduce myself in the green room)
👋 I work on the official MCP server for a big observability platform. Biggest issue we face is fragmentation across clients; Claude Code+Cursor do clever things w/ large numbers of tools and large tool results but we can't assume that all clients will; constrains how we design our tool surface.
seems like their APIs are more open than most other chat platforms, was considering it for some home automation agent stuff
otherwise: no, not that I'm aware of
it's a *lot* better than it used to be. but if you do run into any issues: bsky.app/profile/reil...
I foresee a lot of projects adopting SQLite’s private test suite approach
