While I played with my kids, I set Astra to work on my personal benchmark: building a simple note-taking app that I actually use every day.

How did it do? You can judge that for yourself, but read on for my take.

Isn’t this a trivial test?

You might wonder why I care at all about building a note-taking app when people are demoing things like 3D games and advances in mathematics.

Well for one thing I take a lot of notes. I have peculiar desires for how this tool should work so I used to cycle through different apps and feel constantly frustrated, but now I have exactly the thing I want. It doesn’t make the notes any better, but it makes taking them feel better to me.

More to the point, there is a gulf between making an impressive demo, and making something people actually want to use. Making something I touch a hundred times a day helps ground me in practical progress rather than razzle-dazzle.

The prompts

This has been my go-to test project for new coding models. Claude got good enough to build a useable Mac app for me around Opus 4.5, and I cycled between Opus 4.7 and GPT-5.5 to make the iPhone version.

Here’s what I tried out on Astra.

Ok here's an idea. You are a much smarter model than the one that built the iPhone app. Could you just build a new version from scratch, making it absolutely smooth and fantastic?

Astra one-shot a working iPhone app, but it sucked.

A lot of the subtle details I like (like the buttons sliding off) are missing. Can you take a pass through the old app, and try to deeply endrestand [sic] what I like about it?

There are some parts that could be better (like the look of the home screen especially top buttons) but a lot of the fine details were there on purpose

… and that did it.

After that, I had it rebuild the desktop app using the same technique.

Playing around with both apps, I could immediately feel the improvement. They had been good enough to use, but now they were faster, smoother, and less buggy. The search went from barely useable to great. This is despite having spent much longer building, iterating, and fixing them with previous models.

It’s at a level of quality where I could even imagine someone else wanting to use or hack on it, if they happened to share my unusual note-taking proclivities, so I had Astra set up a personal Mac dev account, set up a landing page, etc.1 while I kicked back. This is the kind of last mile cruft work I would previously never have done for a random tinkering project, but now it was easy.2

Try it yourself

You were promised takes

I have a few.

• Astra is very good. It’s a big step forward and worthy of being called GPT-6. Things are moving very fast now, and the “Chris’s dumb note-taking app” benchmark might be saturated.

• This changes everything about writing software. I have never read the code in this project. When I wanted to know what technologies and libraries were used, I had to ask the robot. This is a new species of software where what used to be code is now effectively machine output, and the real “source code” is the combination of documents, artifacts, and other surrounding context. I bet there are already real companies out there with real products saying “hey, rewrite this thing from the ground up, make it awesome this time, make no mistakes” and that this will become viable and commonplace for even complex, serious software.

• This doesn’t change anything. I’m reminded of this famous comment on the Dropbox launch post on Hacker News.

  

  The hard part about making great software was never just the software. Many valuable things are theoretically replaceable with a small shell script; Substack is “just” some web hosting and subscription software. The gap between a demo and something good is fundamental, and will remain large even as the frontier of what is easy to build moves forward quickly.

• Attention is all you need. As making things gets cheaper, getting anyone to care gets more expensive. Slop happens when you ask others to care about something before you’ve cared about it yourself. Giving a shit, and inspiring others to, is underrated.

• I worry more about Hallucinated Productivity than AI psychosis. Some people do fully lose touch with reality, but rightly or wrongly I don’t feel like that will happen to me. On the other hand, I find it very easy to fall for the illusion that I’m being productive, i.e. making something that will ultimately be of use, when in fact I’m just enjoying the feeling of making something.

  It’s not lost on me that messing around with my note-taking app would be Exhibit A for this, but for the fact that I know the only point is to make a daily task more pleasurable. I’m more concerned about this happening when I try to do actual work. The new tools can be superpowers but they can also be traps. Developing a nose for real progress is a key skill—maybe the key skill—in the era of strong AI.

Worse, the algorithms that govern AI also govern our accounts, which means any human sufficiently engaged online is gradually trained to behave like slop — is slowly transformed into slop. The result is our present, miserable chaos, a daily slop war, to which we are also addicted.

We also get in to our new Pangram integration which Mike is excited for but “the moment the false positives start coming in, I'm going to turn on this like 180 degrees. So just brace yourself for it.”

Pirate Wires

An American media company reporting at the intersection of technology, politics, and culture. Subscribe, or die.

• Against Claudefishing

• Don’t Let Activists or Robots Choose Your Words

Racket News

Regular news and features by award-winning author and investigative reporter, Matt Taibbi.

By Matt Taibbi

• Against Claudefishing

• AI Content Is Everywhere on Social Media, Especially LinkedIn

• How Much of Substack is Actually AI?

User Mag

A tech and online culture publication by Taylor Lorenz, featuring original reporting on how the internet is reshaping politics, business, media, and culture.

By Taylor Lorenz

Thank you , , , , , and many others for tuning in.

Get more from Chris Best in the Substack app

Available for iOS and Android

Get the app

’s Text is king

“Article Armageddon”

The impeachment of President Plump

Mills’ sartorial quest to distance himself from his children, and ’s review thereof.

I didn’t go looking for a fight with Britain’s Online Safety Act. But as the CEO of Substack, an online publishing platform for writers, journalists, and creators of all kinds, figuring out how our company would adhere to the new regulatory measures became part of my job.

Substack has a strong and growing presence in the UK, with independent journalists and major cultural and political figures publishing their work and communicating with their audiences through our platform. So when the OSA came into effect, we set out to comply.

What I’ve discovered as we’ve implemented these rules shocks me.

In a climate of genuine anxiety about children’s exposure to harmful content, the Online Safety Act can sound like a careful, commonsense response. But what I’ve learned is that in practice, it pushes toward something much darker: a system of mass political censorship unlike anywhere else in the western world.

What does it actually mean to “comply” with the Online Safety Act? It does not mean hiring a few extra moderators or adding a warning label. It means platforms must build systems that continuously classify and censor speech at scale, deciding—often in advance of any complaint—what a regulator might deem unsuitable for children. Armies of human moderators or AI must be employed to scan essays, journalism, satires, photography, and every type of comment and discussion thread for potential triggers. Notably, these systems are not only seeking out illegal materials; they are trying to predict regulatory risk of lawful, mainstream comment in the face of stiff penalties.

Once something is classified as potentially sensitive, the next step is age gating. Readers—who in our case are overwhelmingly adults reading lawful material—often must be asked to prove their age through third-party checks that may involve facial scanning, providing identification documents, or financial verification. These measures don’t technically block the content, but they gate it behind steps that prove a hassle at best, and an invasion of privacy at worst. Readers who hoped to engage with the material are deterred from doing so; writers and creators, some of whose livelihoods depend upon getting their work in front of potential subscribers, bear the penalties. The result is that vast swathes of legitimate cultural discourse are swept up, bogged down, and discouraged.

Substack is building an economic engine that supports authorial independence, and as such, we are strong defenders of the freedom of the press, which we believe is essential to a free society.

For me, this belief began early. Growing up, my parents allowed an exception to bedtime if you were reading a book, and I took full advantage. Those books gave me a window into a world far larger than my own Canadian suburb. As a teenager, I explored the wild internet of the 90s, encountering both bad ideas and transformative ones. I found perspectives and cultures that eventually led to my career as a technologist and my life in San Francisco. The opportunity to read and explore widely was essential to fostering curiosity and independent thought, and it’s something I encourage in my own children today.

None of this is to say that there aren’t serious problems that manifest online, or that the government has no role to play in protecting children on the internet. Fighting those dangers with effective state action is important. When it comes to crimes against children, governments have a responsibility to respond with every tool at their disposal, especially vigorous investigations and prosecution of perpetrators.

What the OSA does is something different.

It focuses on speech on online platforms, forcing companies like Substack to build systems that decide when an essay crosses the line from satire to threat, or when photojournalism is too graphic—or tries to predict when a regulator might deem it so.

This would be concerning anywhere, but the UK already has expansive policing of speech. Recent reports suggest that police make over ten thousand arrests each year for online communications offenses, while only a small fraction result in conviction. Now with the OSA, there will be pervasive classification and identify verification, too.

This is how you end up with “papers, please” for the internet.

Substack is bound by the law and will continue to comply with it. We have taken care to implement these measures in the most transparent way possible, and we continually refine our approach to maximize freedom of expression within the bounds of the law. Smaller publishers and platforms, however, may not have the resources to do the same. In those cases, writers, publishers, and readers all suffer.

The Online Safety Act does something different than what it says on the tin. It is not the most effective way to keep kids safe, and it hurts a free society.

If this model spreads, it won’t just block content for children. It will determine whether adults can read, write, and argue freely without first submitting to surveillance. Companies like ours will respect the laws you choose, so please be sure you choose the future you want.

Their power to silence depends on creating fear: We don’t take political positions at Substack, but we do sometimes talk about our principles so that people know what to expect from us. and I wrote about how today’s threats to free speech depend scaring us into thinking they are inevitable — and why that won’t work. (This one is an essay so read it.)

a16z is Moving to Substack: I caught up with , , and about the move, free speech, and the attention economy. (Substack, YouTube)

Complex Systems: a.k.a. patio11 stopped by the Substack office and we took our recording studio for a test drive, talking about epistemic infrastructure, Substack defender, and more. (YouTube, web)

Bonus: we made shiny new badges to set the subscribers apart from the Non-Paying Customers.