turnsout

Turns out I was onto something

This is super cool… It's interesting to see it build. I can't tell if the agent will run indefinitely, but it's been going for 7 or 8 minutes now, constantly tweaking its composition.

I find this approach to be more appealing than AI models that generate fully baked songs as waveforms. Give me something I can open in Logic and keep tweaking…

Because Opus 4.6 is better than 4.5. So if it's true that Sonnet 5 was so good they gave it the Opus name, does that mean there was an Opus upgrade that didn't pan out? And what is Sonnet 4.6? An upgraded Haiku? Just trying to follow the red yarn in the conspiracy board here.

yes, pretty much

Agreed, this is exciting, and has me thinking about completely different orchestrator patterns. You could begin to approach the solution space much more like a traditional optimization strategy such as CMA-ES. Rather than expect the first answer to be correct, you diverge wildly before converging.

Man, I'm in the exact opposite camp. 1 smart model beats 1000 chaos monkeys any day of the week.

If that's true, it would be surprising; the current Sonnet 4.6 is not in the same league as either Opus 4.5 or 4.6, either anecdotally or on benchmarks.

I think it's simpler than that. AI, like the internet, just makes it easier to communicate boring thoughts.

Boring thoughts always existed, but they generally stayed in your home or community. Then Facebook came along, and we were able to share them worldwide. And now AI makes it possible to quickly make and share your boring tools.

Real creativity is out there, and plenty of people are doing incredibly creative things with AI. But AI is not making people boring—that was a preexisting condition.

More lossy than the current non-English behavior?

I've been pleasantly surprised by the Claude integration with Xcode. Overall, it's a huge downgrade from Claude Code's UX (no way to manually enter plan mode, odd limitations, poor Xcode-specific tool use adherence, frustrating permission model), but in one key way it is absolutely clutch for SwiftUI development: it can render and view SwiftUI previews. Because SwiftUI is component based, it can home in on rendering errors, view them in isolation, and fix them, creating new test cases (#Preview) as needed.

This closes the feedback loop on the visual side. There's still a lot of work to be done on the behavioral side (e.g. it can't easily diagnose gesture conflicts on its own).

Do you think there would be value in a workflow that translates all non-English input to English first, then evaluates it, and translates back as needed?

Difficulty is the only true moat. [Astronaut: always has been]

Current examples: esoteric calculations that are not public knowledge; historical data that you collected and someone else didn't; valuable proprietary data; having good taste; having insider knowledge of a niche industry; making physical things; attracting an audience.

Some things that were recently difficult are now easy, but general perception has not caught up. That means there's arbitrage—you can charge the old prices for creating a web app, but execute it in a day. But this arbitrage will not last forever; we will see downward price pressure on anything that is newly easy. So my advice is: take advantage now.

This is what excited me about Sonnet 4.6. I've been running Opus 4.6, and switched over to Sonnet 4.6 today to see if I could notice a difference. So far, I can't detect much if any difference, but it doesn't hit my usage quota as hard.

It's more about supply going down (fewer people going into trades), which eventually affects demand for that shrinking pool.

At that point, you'd be relying on a bug in curl / Python / sh, not the bot!

Once demand drives electrician salaries up to $200k/year, the influencer grind will lose some of its shine

Absolutely. This is why I'm hesitant to go full "dark software factory" and try to build agent loops that iterate in YOLO mode without my input. I spent a day last week iterating Skills on a project by giving it the same high-level task and then pausing it when it went off the rails, self-reflect, and update its Skill. It almost took me out of the loop, but I still had to be there to clear up some misunderstandings and apply some common sense and judgment.

It seems intuitive that a naive self-generated Skill would be low-value, since the model already knows whatever it's telling itself.

However, I've found them to be useful for capturing instructions on how to use other tools (e.g. hints on how to use command-line tools or APIs). I treat them like mini CLAUDE.mds that are specific only to certain workflows.

When Claude isn't able to use a Skill well, I ask it to reflect on why, and update the Skill to clarify, adding or removing detail as necessary.

With these Skills in place, the agent is able to do things it would really struggle with otherwise, having to consume a lot of tokens failing to use the tools and looking up documentation, etc.

If you've ever tried to build something real with agentic AI, you know that it takes time. You can't (yet) snap your fingers and produce a fully market-viable clone of a SaaS product.

The specifics matter here. If you run a CRM for Dentists, can someone replicate your product in a weekend? I'm going to guess that dentists have some esoteric needs related to their CRM, and it's a little more complicated than an outsider might guess.

So what is the threat model? That a dentist is going to get fed up and try to DIY? I think you should encourage that, so they'll see what goes into it. That a 22 year old chooses "CRM for Dentists" as a thing to vibe-code over a weekend? Again, good luck with that.

I really dislike this SaaSocalypse fear mongering, because it's just not based in reality. Show me five examples of established SaaS companies being wiped out by vibe coding.

Absolutely. I'm glad the linked article exists! Hopefully it can prevent someone who really can't afford it from splurging on expensive cables.

This may be an unpopular opinion, but if people believe they can hear a difference in their $1000 cables, and they enjoy purchasing and testing them, I'm inclined to let them enjoy themselves. I have a basic hi-fi setup with rational cables, and enjoy the cost savings, but to each their own.

I feel the same way about wine. At a certain point, it's not really about objective improvements, it's about vibes and lore.

I love the idea of a living slide deck. This feels like a product that needs to exist!

Interesting—thanks!

With the GLM news yesterday and now this, I'd love to try out one of these models, but I'm pretty tied to my Claude Code workflow. I see there's a workaround for GLM, but how are people utilizing MiniMax, especially for coding?

Downvoted, but fight me on this… It's important to see what it wrote, but what it read?

As a heavy CC user, I appreciate a cleaner console output. If you really need to know which 3 files CC read, AI-assisted coding agents might not be for you.

It needs to be updated to do its calculations in linear light, but it's probably useful for getting an intuitive sense of what the different levers of photography do to an image.

Yes, this is my experience as well. I've found the key is having the AI create and maintain clear documentation from the beginning. It helps me understand what it's building, and it helps the model maintain context when it comes time to add or change something.

You also need a reasonably modular architecture which isn't incredibly interdependent, because that's hard to reason about, even for humans.

You also need lots and lots (and LOTS) of unit tests to prevent regressions.

Yeah… I'm using Claude Code almost all day every day, but it still 100% requires my judgment. If another AI like OpenClaw was just giving the thumbs up to whatever CC was doing, it would not end well (for my projects anyway).

I still use Opus for difficult challenges, but if we're building a web app or creating a few scripts, I default to Haiku. It's so much faster, and obviously doesn't impact your usage as much.