I’m building AI-native tooling in public.
https://t.co/m6Zs3j5Loi - shipping code with agents
https://t.co/SwaoJIm5b6 - agent memory and durable context
For longer form deep dives, https://t.co/CMKTdrMyDH
I've been on the record for ages about poor return on code reviews. Meaningful bugs are seldom caught by humans, the overwhelming majority is people saying "I like this style instead" or "make this reusable" and then later "don't prematurely design/optimize".
All it does is crush velocity and kill morale.
AI, unlike humans, will actually find bugs. I've been following the AI review + merge since Opus 4.5.
Yes, humans do sometimes (but very rarely) find meaningful bugs. And yes, humans sometimes (but very rarely) extract meaningful context about what teammates are doing. And lastly, humans do learn new ways to do things (if they are open to it) with code. The problem with that is, we no longer need to learn those things anymore.
Ship the goddamned code. Test in production. If you don't know what that truly means, fucking search it up (since you probably don't trust AI).
Today, a human presses merge. Within a few months I expect: automated reviewer LGTMs -> another Claude with a fresh context window judges if it’s simple and low blast-radius -> auto-merge
Was that 7 years ago.
People did code before 2010, just saying.
When Java got hot, people were trampling over each other to get out of their shit c/cpp stack and into Java.
Java can be terrible and correctly avoided today and yet still have been universally respected as the best language st some point in it's history.
Quite possibly the dumbest tech take I've seen on Twitter.
Most C coders in the product space said Java is better than C. It's not close. You may as well have said the sky is dark brown.
I saw brokerage firms and stock exchanges rewrite their stacks in Java in the early 2000s.
I know Javs is not liked today, but that's not how it was received then.
It's not just the smartest people. By far they are the most desired employer in tech. By a country mile.
https://t.co/WwcSH6TZWl
For me they are always pushing the envelope. They invented the first useful coding agent. They gave their model a personality that GPT has never had. They introduced subagents and more recently ultra code.
Codex is excellent now, especially the desktop app. But for most tasks Claude is just more enjoyable to use.
It's not all about getting shit done. A Honda will reliably get you from point A to point B. But BMW still has a nice business, right?
@ziwenxu_ 32 principles ~= 0
At the very least you should be able to count them on 2 hands, but 1 is preferable :-).
Lastly the very first principle...maybe I'm getting old but haven't all of the big hits of the century been free plans?
@GergelyOrosz@antirez@matteocollina really bums me out that all our eggs are in that basket...or China's. And it's not that socialist high taxes bullshit, Sweden has kicked out a ton of tech taxing us at 57%
Pleased to announce that today, Sneka (https://t.co/PApte4VRAp) has soft-launched Diffinite (https://t.co/ugkrrchunX) in private preview. To learn more, check out my post at https://t.co/m41BHgAwBO.
@danshipper hmmm why aren't the 'loops' automated from /loop or /goal sessions then...i mean if you've got hundreds of agents and bottomless token capacity...can't you just built that for us :-)
@usr_bin_roygbiv Totally relate, i have an infinite thread with Claude for dictating ideas that get organized into my Obsidian vault, then various agents resurface them via independent workflows.
Tbh pen and paper probably equally effective lol