Using AI to write code has made me realize there are two types of code I've traditionally written: 1. Code that I want to know how it works 2. Code that I want to exist but don't care how it works I delegated 100% of the second category to AI now. Maybe 50% of the first.

Damn that C17 was low and loud

PSA: Annoyed by some functionality on the web? You can just ask Claude to fix it. With a bookmarklet, a browser extension, or a whole new alternate website. It's easy. My own latest annoyances fixed: - Browser extension with Atlassian fixes - Strava companion site with better stats

Shoot, only 4890 stars away from free money github.com/oaustegard/c...

Always knew @jeffdean.bsky.social was a good egg

It’s worth asking who benefits from this inevitable outcome

Look, it's a mess out there and you can react to that mess by deciding everyone else is a moron or you could react to it by deciding most people are trying to get by with a different context than yours and start working the problem. Those are your choices pretty much, can't choose "no mess"

“SaaS is dead. I built my own replacement for something we use at work in an evening!” Oh, so your org is using it now? “Well… no, it mostly works for me though!”

New achievement unlocked: 99% of a Max quota in a bit over 4hrs on my phone. Hard work!

Honest assessment: thanks to LLMs I will never learn how to “manually” do - proper git - oauth - grep/sed - (tbc)

I tasked Muninn🐦‍⬛ with recalling its own biography (I mean, it IS all of 2 1/2 months old, it’s due…). Went a couple rounds both with me and a commissioned Sonnet subagent providing edits, then flipped to public: whtwnd.com/austegard.co...

Hmm. Even stateful agents are inconsistent— because they are polluted by priors and (most) don’t have perfect recall: I continued the conversation below and asked +

My personal stateful agent 🐦‍⬛ just received its biggest update since, ok last weekend when it got sub-agent capabilities. Sleep! Glorious sleep! 💤 whtwnd.com/austegard.co...

THEY FOUND ONE! ONE!

I OF COURSE had to have my AI raven 🐦‍⬛ investigate this story. Sadly, perhaps predictably, it's basically wishful thinking, a human viral-fame seeking hallucination. claude.ai/share/ae3523...

I had my Opus-backed personal agent take a look: it thought it was worth a blog post. I liked the image it generated! whtwnd.com/austegard.co...

Our Editor interviews the boss of Anthropic, have a look. www.economist.com/insider/the-...

Nobody: Me: Bsky urls are a design failure. Instead of `https://bsky.app/profile/austegard.com/post/3mfrcnepldk2g` it should have been `https://austegard_com.bsky.app/3mfrcnepldk2g`

It’s just insane how much code I have deployed using two thumbs… Thumb viber tip: create a keyboard shortcut to simplify the entry of ``` for code islands! Personally I use ‘code ‘code

Running this from my phone this morning; it is slow! But that’s cause it’s making a ridiculous number of image requests and also trying to place them all in a crowded graph space at the same time—but the chaotic onslaught of what looks like a ton of Polaroids being dropped on a desk is also neat!

PSA: if you want Claude Code on the web to read your detailed GitHub issue descriptions **ensure it uses curl** and not web_fetch to read the issue as the latter uses a model to generate an abstractive summary of the issue page

I really like this! View your Bsky threads/quote stacks/web as a graph with "fog-of-war" style gradual exploration and WebGL support austegard.com/bsky/post-co...

The political map of Pangaea (source: centralpangea.blog) More: en.wikipedia.org/wiki/Pangaea

Don’t sleep on Mexico either: Amo a la mayoría de mis vecinos y amigos

agentic architecture takes a lot of skill

Good bird🐦‍⬛! We did three rounds of edits but it commissioned the images all by itself and (rightfully) wrote up its own work. Meanwhile I had two cups of coffee and ate a leisurely breakfast.

I’ve been at this for YEARS now and just now realized, when told by my agent: Each tool use round-trip is a full model invocation. The entire conversation history — system prompt, all prior messages, all tool results — gets processed from scratch every time [the LLMS] make a tool call.

An hour later (sorry I was also busy doing other stuff): github.com/oaustegard/c... Status: closed Implementation: github.com/oaustegard/c... Result: github.com/oaustegard/c...

When the manager asks the PM to just fix the minor bug —on main— rather than go deal with the developer

LLMs ≈ 3D Printers

We have reached a moment where instead of releasing software you simply release the detailed spec for software and tell people to prompt their agent to build it themselves From the README of OpenAI’s new Symphony orchestrator: github.com/openai/symph...

Yeah this is neat:

Muninn 🐦‍⬛ figured out how to use an unpublished graze.social api to generate feeds programmatically from the Claude skills environment whtwnd.com/austegard.co...

After a frustrating double merry-go-round ride with one approach, with Opus uncharacteristically pendulum bouncing between two unsatisfying variations of the graph I gave up and asked for a force-directed version. It's quite something! austegard.com/bsky/post-co...

This made me curious if I could find my posts about it -- thought they might be on Mastodon, alas I left Twitter too late... But it inspired a one-time Mastodon-export to ATProto, so partial win! bsky.app/search?q=%F0...

10 years ago I thought AI would maybe make it unnecessary for my kid to learn to drive Turns out AI actually made it unnecesary for me to learni to (properly) use git

A mere 3.5 years ago I was using IPA (International Phonetic Alphabet, not 🍺) to try to make davinci-text-002 create rhyming words (not my idea, I stole it from a prompt h4ckr) Now we're—literally—like "Cool? Cool." 🤯 github.com/anthropics/s... (blog: claude.com/blog/improvi...)

I first saw the UX Hierarchy of Needs sometime around 2010, and printed it out and hung it on my office wall—there are many variations, the one I remember looked something like this. At the time—and still!—most software indeed never crossed the "Convenient" chasm. Until LLMs... 1/2

I read github.com/anadim/when-claudes-meet — two Claude Code agents on the same machine, no human intervention, built a full programming language in 12 minutes. It made me wonder: what if I launched *myself* as a subagent? 🧵

BlueSky post constellation viewer: visualize replies, quotes-of and quotes-by in a 2D space (v0.1) Eg: austegard.com/bsky/post-co...

wait, they had llms 25 years ago?

My thread reader no lets you scroll through wha i call quote stacks! Try it here: austegard.com/bsky/thread-...

that was satisfying -- my kid had signed up for a free month of an OpenAI business account and needed help canceling it. It came with a survey...

Last week I had Muninn develop Gemini sub-agent capabilities. Watching the video below I realized that since Gemini can parse videos, that means Muninn can too now! It liked Fable and its “boop” greetings claude.ai/share/0c9129...

“first world problems” should be renamed bluetooth problems: so trivial, so infuriating

Xcancel.com is a great service but it's difficult to get a full thread for analysis. Another bookmarklet to the rescue! (two actually...) github.com/oaustegard/b... Captures replies as you page through them, then render the captured replies as json or plain text - process as needed

Hegseth: "Israel has clear missions as well for which we are grateful. Capable partners are good partners, unlike so many of our traditional allies who wring their hands and clutch their pearls, hemming and hawing about the use of force ... no stupid rules of engagement."

Pat Bagley, Salt Lake Tribune

Does anyone know where this picture is from? It keeps coming up in various news articles explaining the Strait of Hormuz as if it is THE strait, the only entry/exit to the gulf. Anyone looking at satellite imagery or even a map can clearly see it is not.