Some ideas on successful ways of building and shipping software. 1. Static language. Go, Rust, Scala are today's kings. 2. Monorepo for code, monorepo for infra. 3. GCP > AWS for most things. 4. Use K8s. The custom solution you made and like is just a subset of it.

My take is getting more and more bearish the more I see and hear LLMs actually used. "Eh it's useful I guess" is my broad take.

Join millions who have switched to Grok.

still valid, no notes.

Finished moving DNS registrar - my former one was yoinked by a PE firm a few years ago and technical quality went down. :-/ . Now I'm on @Porkbun . Back to higher level of technical competence.

RT @hasek_dominik: This is such a huge shame! This is how the biggest criminal of the 21st century, responsible for the death and maiming o….

RT @EllaMcHawk: William De Morgan was the Arts & Crafts movement’s most brilliant and prolific ceramic artist. This biscuit (cookie) set i….

Working through the service pattern approach where service has xyz features. Those can be exposed via rest or grpc, then the canonical access is $service-ctl which uses the client lib and provides a full access to features. Its somewhat inspired by kubectl, vault, docker, a few.

A few notes on 'solo founder' and AI (& other productivity tools). I've been building mostly solo for 2 years now on the side. Code, ops, product research, all of it. Just to give some clarity on my starting point. The underlying product has to be quality. Its a security.

1989? Yeah. Man. I remember when Mom and Dad had their own rocket ships. We vacationed on Mars- once, Neptune. Dad was so poor we had got 40 acres and a 13,370 foot house. Mom made apple pies every few days. It was idyllic. Man. Life sucks now.

RT @bohannon_bot: Behold! The first practical monostable tetrahedron!. Why? Because things like moon landers fail by tipping over onto the….

RT @dissproportion: One of the craziest parts of old radio and tv shows is how people apparently would go to their friends' houses multiple….

Interesting.

The junior dev will learn and keep learning, and be able to apply judgement. Use the LLM in a different capacity- one that aligns with its strengths as a mathematical tool.

Sure, a skeleton works, but you spend lots of time tweaking and fixing it for the use case. Just know what you're doing and go for it. Stop with the templates. LLM - same problem. You can diarrhea code. It's not going to be what you need. Hire a junior dev.

It might be OK writing technical docs based on code. I need to prove that. It might be OK to generate templates that can be modified. What it can't do is reliably generate professional quality code at scale. It reminds me of Wizards doing codegen in the 90s.

LLM AI seems to do well on associative work. It will explain code to some degree with some reliability. It seems to do OK when pattern matching (unit test generation). It seems to do OK when estimating errors (Same reliability as forum posts, which is fine).

Generative code writing AI is NOT doing what it's hype squad claims. Repeatedly. I can make it work, but it's not faster or better than I am- as a rule. I am formulating notions of what it's good at.

Yesterday I asked Jules - googles async agentic AI to implement a grpc interface. Protos written. REST already works. Tons of docs, specs, etc. Turns in something that doesn't work. Jules is supposed to know Rust. Man. There's just no juice to the squeeze.

Not a bad idea, particularly if you instruct it to be blunt and harsh. Syncophantic behavior misleads leaders badly. It's good to have external opinion analysis.

Interesting thread. I like the Kinkade signified. But I can't abide his paintings. I'll have to think about why. Whoever this person is touches on it but I'm not persuaded it's a full accounting.

I play a certain amount of online chess. It's remarkable how many games are determined early on by one misplacement. Not even "big mistake". Just. Not perfect. Not sure there's a generalizable lesson there. It's just interesting how things unfold.