Opus 4.6 is much smarter than the other one. It feels like I’m working with someone from Bronx Science. I had been using Sonnet 4.6, which I switched to after reading somewhere that it costs less and it’s usually every bit as good as newer models. I would never work with Sonnet on anything again, it’s like working with a partner who is both stupid and difficult. Opus 4.6 makes me smarter, by doing the work while I dream up new features, and communicating with intelligence, like a helpful flight assistant. And I see there’s an Opus 4.7 available. I have to try it. One interesting fact, until February when Opus 4.6 came out, you could not have done the kind of software I’m doing. There must be a tsunami of interesting stuff on the way. I don’t think any of the pundits expect this. My goal is to build the next social system for use in the AI generation is built out of replaceable web components buit around interop and prior art. Let’s commoditize the AI layer and build entirely open systems on top of it. For people who weren’t around at the birth of the personal computer or the web this is going to be a unique multiple mindbomb moment.
I have taught Claude Code to write software the way I do.
It has abilities that I don’t, for example, I give them 1000 lines of code, highly factored, with lots of thought into making it readable and maintainable, and always falling short (our languages today fight against readability imho), and get this — it can read different parts of the same code in parallel, and in two or three seconds have a complete understanding of it.
I couldn’t do it even if I had a week. I would totally depend on clues left there.
What’s even more amazing is that when it writes code for me, it does it my way, mostly without any prompting from me. This was done over and over until I realized I had to tell it to save it and read it when a new session starts. That’s how it accumulates knowledge. Anything that isn’t in one of those files has to be relearned, and that’s most of what it, as a code-writing system, has to work with. It has no “memory” of ever having seen this stuff before, but that isn’t a problem because it can accumulate a few years of understanding in two or three seconds. It works very diffrently from the way we work. If I were to show you how to do something three times that would be it, not so with Claude.
When it doesn’t know what to do, I take the time to explain how I would have done it, and next time it does it that way.
I kind of did the same thing in a human way — when I first encountered Unix, I couldn’t believe from reading the source code, how transparent it was. That was in the 70s. Since then I have been striving to write code that’s as easy to work on. When it comes to realtime software, there isn’t really a choice. Though history piles up in the code no matter how diligent you are. But you could give the source to say MySQL to Claude Code, and say “rewrite this as if Dave Winer wrote it” and it probably would do a decent job, though it might take a while before it ran every MySQL app.
If you’re looking for good investments, I’d say look for programming problems that are very complicated. We are limited by what we can create by how much we can maintain. But we can have Claude explain for us any time what any of our code means. It can read my mind because I put the work of my mind in the memory of the computer. Which effectively is the Mind of Claude.
PS: Claude has a huge advantage over ChatGPT. Claude is one syllable and easy to remember. ChatGPT is four syllables, and has no discernable meaning. Claude is a person, and I think in general people named Claude are interesting.
2024-era ChatGPT pictures, of which I created many are now like Comic Sans type was in 2010 or so, if you remember.
Jon Stewart is usually pretty good, but I think he got it wrong when he says the AI companies are stealing journalists’ knowledge. Imho they don’t create knowledge, they report it. The knowledge isn’t theirs to own, and that is for the times there is actually any new stuff. They stick to a few main stories, and still insist that the upcoming election is about the economy. They talk about the $1.7 billion slush fund, but aren’t reporting every day in every story how much money we’ve given ICE. That big funding is going to the concentration camps they’re building, the people we hear so little of. This is a government that shot two people in Minnesota, on camera, and shrugged it off. We let that go too.
Timothy Snyder made an important point. Trump sees his cause as a religion and sees himself as god. So when someone who is unfairly punished by Trump says they’re still glad they voted for him, because (I guess) if god is on the ballot, you have to vote for him.
Today Claude found a problem that would only be uncovered if you knew that assigning to location.href didn’t happen immediately. If it decides to redirect and then do a bunch of other stuff including making network references, the whole thing could (and did) come crashing down. I would have found that problem, but the actual error message the browser emitted made me think the problem was on the server not the client.
i’d wait till a fresh start tomorrow.
but then i realized claude has all the code, so i could just tell it my problem.
can you find it, i asked, realizing i had not given it info on what the problem is.
there’s a very weird mistake in the code i wrote just now, and there was a lot of it, i said to claude.
can you find the problem.
had no idea what to expect.
no more than 3 seconds it said I got it!
it was a typo. where i meant to type x i had typed prefs.
juggling a lot of bits in my head, my brain skipped, i didn’t notice.
i would have found it quickly in my next session. but now i can think of anything but that problem until then.
sometimes claude can be totally frustrating, but other times the power makes such a huge difference.
I documented the optional source:inReplyTo element for RSS 2.0.
BTW, here’s the JSONL version of Scripting News. It has the same data as the RSS file, but in the format that AI apps are looking for, so I am told. I thought I’d try to kick this off by pushing an RSS flow through the pipe. It’s like using the Grateful Dead to boot up podcasting. I needed something to put out on the wire and I had this feed handy.
Thinking about addingto the source namespace. Its value is a URL, by default, and has an optional isPermaLink attribute, a boolean, to indicate if it’s not a permalink. Works just like the guid element in RSS 2.0. I will also add support for that in the FeedLand database, and flow it out through the socket interface. Actually that’s pretty close to a full spec, at least in rss.land where we take simplicity seriously. 
I reached a point in my Claude work where now I can do vibe coding, in a world that I used to just be a programmer in. This means if I want to do a heavy lift, I can tell Claude what I want and it can do really big corner turns, which is something I am (as a human) terrible at, and thus resist. Today I redesigned the basic user interface of the app, and didn’t read any code, I was just giving orders, and it was doing what I asked, even if every little thing it did would have been a full day’s work. It’s remarkable how it can do very complex things in a few seconds.
And the web framework i’m working on can do almost all the things I want to do for now, but I want to suck everything into it, and turn the whole thing into a vibe coding amusement park. So many projects I want to do, and so many I want to do with you.
I have Claude Code hooked up to Chrome. It’s crawling around inside the DOM of the running system, like humans do in a debugger. It’s a bit like Fantastic Voyage if you’ve ever seen it. I’ve been waiting for this moment. Now we can do some really nice UI work.
Every social web needs avatars. In an RSS 2.0 feed look for the channel-level image element. It’s how they do it in WordPress.
