This is a column about AI. My fiancé works at Anthropic. See my full ethics disclosure here.

I began the year with a flurry of vibe-coding projects: starting by canceling my Squarespace account and building a new personal website, and proceeding to build a suite of tools for creating daily briefings, a journaling companion, and a Platformer archive. The projects were so useful, and so easy to create, that initially I did not spend much time thinking about how unattractive they were. Aside from the website (gorgeous), my new tools were either accessible through the terminal or Markdown files: functional, but fugly.

I’ve continued to use those tools intermittently. Over time, though, their lack of polish came to nag at me. And so I took notice when, in March, Raycast announced Glaze — an all-in-one vibe-coding app for the Mac. Where most coding tools are programming generalists, Glaze was built with the express purpose of building (and sharing) desktop apps. It was not the first to this idea — Wabi, whose founder Eugenia Kuyda I interviewed here last month, does the same thing for mobile apps — but as someone who does most of my work on a laptop, I found Glaze more immediately appealing. Particularly given Raycast’s track record — its launcher app is one of my favorite and most-used pieces of software.

I added my name to the waitlist and forgot about it. Then, at the beginning of June, my invitation showed up in my inbox. 

Last week, Glaze opened up to all users. It includes a free tier with a one-time bundle of credits that will let you build an app or two; after that you'll need a Pro subscription — $20 a month at launch — which refreshes with 200 credits monthly and lets you buy more if you run out.

In my first month, I built three apps that I’ve been using to accomplish various tasks. All of them involve productivity in one sense or another, though only one has any practical utility that extends beyond what was already available in the App Store. 

Still, I have found the process of making Mac apps delightful, even when what I have made is extremely silly. And while ultimately my conclusions about the next generation of vibe coding don’t extend far beyond “it’s fun to make things” … it’s fun to make things! And that has made me modestly more confident that there is a future for hyper-personalized software, at least in the short term.

So with that, here are a few notes on my early efforts to usher in the SaaS-pocalypse.

The Nightwing app

By far the most ridiculous app I have made so far is called Nightwing, after the DC superhero. Recently, in an effort to watch less YouTube, I have begun reading a comic book or two before bed. My original comic-book phase had taken place in middle school; I hadn’t picked up a new one in 30 years. But something about the profusion of AI-generated slop in all my social feeds had made me desperate to see and pay for some handmade human art, and so a couple months ago I bought subscriptions to Marvel and DC’s respective comic-book apps and began digging into what I had missed.

One superhero whose stories I found myself enjoying was Nightwing — Batman’s original Robin, all grown up and leading a team of superheroes in various world-shaking conflicts. Around the time I got access to Glaze I had the amazing, terrible idea of creating a Nightwing-themed to-do app. 

To-do apps are the classic first vibe-coding project, and there is almost never any good reason to make another one. (I can’t imagine anyone meaningfully improving on Todoist.) Still, I find myself changing to-do apps roughly every six months for aesthetic reasons, and for the relief that comes from starting fresh with a blank slate of projects.

Within a few minutes of using Glaze, its advantages over a pure coding app like Claude Code became apparent. (At least, to a non-technical person like myself.) 

Glaze starts with a working Mac app template, which lets you skip lots of prompting and scaffolding work. It compiles and installs the app automatically — another big time saver. It calls image-generation models natively, which Claude Code doesn’t. And perhaps most helpfully, you can edit your app while it’s open — even drawing a circle around specific elements you want to change to guide the coding agent.

You can publish your app to Glaze’s store if you want, or browse other people’s creations to find inspiration. For the most part I have found apps in the store pretty boring — featured suggestions include a PDF reader, a subscription tracker, and a Claude usage monitor.

My Nightwing app is defiantly absurd by comparison. At the click of a button, I can generate an image of a Nightwing-esque character doing any task that’s on my list: sending an email, buying groceries, getting an MRI. When I mark it completed, comic-book-style “boom!” and “pow!” animations fly across my screen.

A widget at the bottom of the app offers a synopsis of a random issue of the Nightwing comic book, while the “issue number” at the top of the app tells me what day of the year it is.

Is any of this remotely necessary? Of course not. Does it represent a copyright violation and an indictment of the safety guardrails of the LLM provider I am using to generate these images? Almost certainly.

In the meantime, though, my little app has succeeded where no previous to-do app has to date: making my chores fun. For me. Until recently, software development was too expensive to pursue something so stupid. Not anymore!

Moreover, all of these features came about through me typing what I wanted into a box, and watching the app update live in front of my eyes. Those live visual updates made vibe coding feel easier and more interactive than it has been so far.

The Platformer Mac app

After Nightwing, I built a better Platformer app. For the first version, I used Claude Code to ingest the publication’s entire archive and allow me to run semantic searches over it. This is useful for reminding me what I’ve said about a particular subject, and when — “summarize my recent coverage of Meta’s Oversight Board,” for example, is more useful than a Google search, particularly on deadline. 

But running those searches has required launching the terminal, launching Claude Code, and then invoking the skill associated with this particular project. It’s not hard, exactly, but I’ve used it less than I might have otherwise due to the friction involved.

With Glaze, I quickly built a Platformer app that now lives in my dock. Like the Claude Code app, it can run queries over the entire archive using my Anthropic API key. But it also extracted the key topics and people I return to over and over again, and I can browse my coverage of them by clicking on the chips that sit on the app’s home page. I also added a feed of my most recent 10 columns to the home page, along with randomly selected quotes from the past six years.

If my Nightwing app is silly, this one is quite serious. It’s a cheap, elegant research assistant that no one else would ever have built for me, and I put it together in about a day. It’s the sort of app that I can imagine being useful to writers of all sorts, whether for journalists keeping track of beat coverage or fiction writers trying to stay on top of their lore. And it can be adapted to the writer’s exact needs, whatever those might be, and however they might change over time.

My final project is still half-baked. It’s a contacts app that I’m trying to adapt for my journalism called Source Code. I’ve never been great at keeping track of who I talk to and when, and with Source Code I’ve combined the contacts in my phone with my contacts on LinkedIn to begin to build a more visual graph of my network. 

What I really want is to create a CRM-like system for updating my network’s contact information automatically as their jobs change — a kind of private, locally stored LinkedIn where I can add notes and documents, and perhaps run semantic queries as well. (“Who do I know who used to work at Google?”)

For the moment, I’m thinking through issues related to privacy and security. I’m also not sure whether I even like my current design, or whether I want to pay for a CRM service.

Still, I feel a sense of excitement and possibility that I never previously felt while updating a contacts manager. So much of the historic experience of using software is having to put up with the things that suck about it. The new vibe-coding tools promise a world where whatever sucks about the software you use can be changed — instantly.

If it’s boring, make it funny. If it’s ugly, make it pretty. For too long, software has expected us to bend to its will. Glaze and other vibe-coding tools are turning that logic on its head — and the revenge feels very sweet indeed. 

A MESSAGE FROM OUR SPONSOR

Protect your identity and finances

Scammers use AI and deepfakes so convincing you can’t spot scams anymore. One moment of trust, and they steal your identity or drain your accounts. Coveron offers what prevention can’t - identity theft recovery (up to $2M), fraud insurance (up to $10K), and expert support when you need it most.

Get 70% OFF

Following

China weighs restricting AI

What happened: This month, Chinese authorities met with AI developers to discuss potentially restricting international access to China’s most advanced models, Reuters reports. Companies including Alibaba, ByteDance, and Z.ai were involved in the talks.

In the past few years, Chinese AI companies have become known for their open models — LLMs close to the frontier that developers can download and customize. China’s Ministry of Commerce is now in discussions with those companies to restrict these open models, as well as closed ones. Officials talked about making any leak or theft of proprietary AI tech an offense under China’s national security law.

Chinese authorities are very worried about Anthropic’s cyber-capable Mythos model, Reuters reports. As a result, officials are weighing policies similar to those the Trump administration adopted this month. (US officials have asked OpenAI and Anthropic to limit their most powerful models to a small group of approved organizations, and only in the past week lifted export restrictions on Anthropic’s Mythos and Fable models.

Why we’re following: Mythos — and the Trump administration’s strong reaction to it — are drawing more attention to the national security implications of AI systems.

If China's proposed restrictions go into effect, startups and researchers around the world that rely on open models to perform tasks cheaply and customizably would have fewer, and worse, options.

Offering open models that remain close to the frontier has helped speed adoption for Chinese AI developers like Alibaba and DeepSeek. But it’s less clear what China stands to gain from continuing to proliferate open models: Alibaba has been struggling to monetize its Qwen line despite high performance and a large user base, and powerful open models may ultimately benefit China's adversaries just as much as they benefit domestic interests.

What people are saying: On X, Brookings research fellow Kyle Chan wrote that he’s not sure “whether Beijing would put a hard limit on open-source model capabilities, given how central open source has been to China’s entire AI strategy.”

On Bluesky, Wharton economics professor Ethan Mollick thought such limits might happen: “This is a key reason I don’t expect the flow of frontier open weights models to continue indefinitely, or even for very much longer. The gap between open and closed capabilities may soon start to grow, not shrink.”

—Ella Markianos

Australia’s social media ban stumbles

What happened: 

Australia’s pioneering under-16 social media ban is failing at step one as platforms are failing to verify users’ ages, a new study found. Out of 50 test accounts where users claimed to be 16, platforms asked for proof of age on none of them, researchers told Reuters.

The oversight comes from the initial vetting stage that estimates a user’s age range based on their online activity, the researchers said, which doesn’t appear to be flagging young users for further checks. 

While none of the platforms let users sign up if they said they were under 16, only one — streaming platform Kick — refused to let users create an account without proof of age. (A Kick spokesperson said the platform could not rely on age estimation because it did not have enough data to guess a user’s age.)

A Meta spokesperson said the test accounts stated they were over the minimum age and it was unclear if the accounts had “posted content or engaged in a way a true under-16-year-old user would.”

Why we’re following: The ban has indeed been ineffective to a significant extent — research published last month found that 80 percent of under-16s are still on social media three months after the ban, as we covered in last week’s Platformer column. 

Australia’s government said the country will double the maximum fines against companies that break the law in response to the study and accused the platforms of setting the ban up to fail. Its eSafety Commissioner is actively investigating Meta, Snap, TikTok and YouTube.

Despite the ineffectiveness i