GPT-6 Astra just ended software

1. Sol, Astra
2. SLOPware
3. Amazon
4. Microsoft
5. Claude Code
6. Sonos (before AI)
7. Instagram
8. YouTube
9. Abundance → fatigue
10. Vulnerability curve
11. Tea leakage
12. The virtuoso dillema
13. Ad Astra
14. Do we have time?
15. Hyp(e)rpersonal
16. Per aspera

THE AI BUBBLE WEEKLY 3

OpenAI has released their latest Astra model. With a retro-aesthetic, very human oriented ad. It showed the model operating apps, even the most complex ones like Blender. And context-switching on the go, to order food. This presentation was like the AI from the movies. And right about time!

The internet went wild with announcements that AGI is finally here. OpenAI desperately needed to address the full market, not just devs. But the dev parts of Astra are more interesting here so I'll tell you how it killed software, and how it can try to revive it.

Sol, Astra

Sol was OUR star. Astra is "all the stars". You can't get much better than that. Per aspera, ad astra! To the stars! Even the naming suggests magnitudes!

But don't worry, they can still progress to a galaxy and beyond. (GPT Universe, anyone?).

The model crushes all benchmarks and people share "impressive" 3d games coded in 30 minutes with it. But that's not the reason software is in such a bad state.

In 2026 we're at a point where we got the most apps created ever and the largest disinterest in new software.

The SaaS MRR model is dying as more and more users look for one-time payment alternatives.

In 2026, nearly half of consumers canceled at least one subscription. That's up from just a third in 2024. The average household now manages 11 or more subscriptions and underestimates the true cost by about 2.5 times.

One-time and lifetime purchases rose from 6.4% to 10.3% of app monetization models over two years.

And then there's the quality drop.

I don't mean Sunday vibe coders with this. I mean big, respected brands using AI and delivering slop.

Eventually, even the non-tech-savvy consumer notices the decline.

I'll show you why Astra is a pivotal moment in all of this, but first let's go through some examples of software going wrong this decade.

SLOPware

Software has increasingly been turning into slopware. Little oversight. Accepting the very first thing AI delivers if it seems good enough. Shipping more, but without control, plan or taste.

Amazon

Amazon has had 4 (Sev-1) outages in just one week between February and March of 2026. Their memos showed "AI-assisted" code changes as the primary culprit.

A six-hour outage cost them around $6.3M in orders lost. It was a deploy that went live with no approval or even documentation.

Their internal memos literally said these were "Gen-AI assisted changes" with "high blast radius".

Microsoft

Microsoft wanted Windows to be super-duper-AI-workspace OS. That came with annoying performance issues and a general lack of polish. Vista style.

Windows now feels like a wrapper for some not always necessary AI features rather than an OS. In this case it's not even about AI writing the OS code, it's more about roadmapping the entire release around that AI hype, when most regular users use AI to search for the best cheesecake recipe.

Claude Code

Three overlapping bugs and Claude Code output quality went down for six weeks. Power users canceled. Security researchers wrote papers. People were mad.

Sonos (before AI)

The 2024 Sonos app rewrite was utterly broken when it went live. Took a year of fixes and tens of millions of dollars. They even kicked out the CEO. Many disgruntled users never came back to the brand.

Which shows trust doesn't always come back after its lost with low effort fails.

Instagram

Instagram is a huge mess. AI assisted mess of course. And I don't even mean thousands of lawsuits happening against Meta at this time. Here's a funny hack for you:

For seven weeks hackers exploited Meta's AI Support Assistant to hijack Instagram accounts. The method required no hacking skills. Seven weeks!

Open a chat with the bot, ask it to add a new email address to anyone's account, receive a verification code at your own email, hand the code back to the bot, and it gave you a "Reset Password" button.

That in itself is hilariously bad for the brand. Then add a plethora of UI bugs, annoyances and stability issues. Meta truly is an AI (head)first company. If AI would jump off a cliff, they'd too.

YouTube

Check out how YouTube behaves on desktop. This problem first appeared 6 months ago and still hasn't been fixed. Happens about 50% of the time.

And of course countless startups also ship half-baked and broken, but that was always the norm.

Abundance → fatigue

Vibe coding resulted in an 84% jump in new Apple App store submissions in a single quarter. It was the largest bump in a dacade.

Review times went up to sometimes even 30 days. Apple also started removing "Slop-apps".

Most of the apps from this wave got single-digit downloads. You can build it, but nobody cares.

This reddit screenshot sums it up very well.

My total users is roughly my wife and one guy in Finland who I suspect downloaded it by accident.

This is overabundance meeting an overstimulation.

Vulnerability curve

Georgia Tech's Vibe Security Radar noticed software vulnerabilities directly caused by gen-AI coding went from 6 in January to 15 in February to 35 in March 2026.

All that shipping leads to even more holes in everything.

Tea leakage

The jokes write themselves!

Remember Tea? The women-only safety app? Their backend was just a public Firebase database with no auth.

70,000+ images including passport scans got leaked. It became the prime example of "vibe coding = not secure". Current models got better at this, but you know what? You know it. I know it.

We're in a bubble of our own. It will take years to convince the general population, like the Tea app users to trust a "vibed" product again.

Once regular people can pattern-match the slop, they pattern-match more things as slop. Sometimes wrongly.

The virtuoso dillema

There's also the perception of effort.

Imagine Les Claypool playing the bass. He's probably the best bassist in the world (Sorry Davie504!) and when you see him play you know why.

To regular people, what he does is mastery. And at the same time, to regular people a skilled DJ is "just playing from vinyl".

It's not that the DJ is not skilled. He can be a Les Claypool equivalent when it comes to musical talent and creativity.

But just the amount of "technology" that goes into his setup changes how most people see and compare those two musicians.

This is happening to coding and design now. Art has already been taken over and slopified. Real people start to see slop. They see low effort, sometimes even when it's not there.

So is software over?

Or are we just reaching the peak of the enshitiffication curve?

Ad Astra

Models are getting better. Skilled people are doing great things. And no, I don't mean GTA 6 clones vibe-coded in 30 minutes. A great game won't be built that way.

But Astra went a step further. It tests front-end and does QA checks itself. Which technically means less broken UI in apps. Higher quality of both backend and frontend. Better security. Impressive code quality. Even DHH is impressed by what agents can do now, and he knows code.

But you know what's tragic about this?

Mythos and Astra may be a fix that comes slightly too late. Like replacing a bad actor in a play with an amazing one, but after the audience has already left. With one star reviews.

Ok. Ok. I'm being dramatic so you keep reading. You got me!

Sorry. Let's be serious again!

There's an audience. We, tech bros are the audience that's not giving up. But that is not enough.

Frontier AI labs already see that and they're annoyed Susan, an 80 year old grandma is not vibe coding her own apps and burning lots of tokens. Why is she just researching cheesecake recipes?

Regaining trust of non-tech people will take time.

Do we have time?

No idea. AI is eating up SaaS already, maybe there isn't time to fix SaaS before that happens.

I'm anti and pro AI at the same time

Don't get me wrong. AI coding is mesmerizing if you know software architecture. You definitely CAN build things. Either for yourself, or for others.

The problem is that SaaS, especially monthly priced has also reached a low. People try to look for cheaper or single-payment alternatives to everything.

Hyp(e)rpersonal

I write (what you're reading too) on a 12" Macbook from 2017. I installed Omarchy linux on it and the laptop has gained a new life. It's also "agentic" now, with AI deeply embedded into the OS.

Why does it matter? the Macbook has a large trackpad very close to the keyboard. Accidental swipes moved my cursor while typing, and I ended up typing in the middle of a previous sentence.

Super annoying! So I asked the agent to FIX THE OS. And it did. No more accidental swipes. I also made a couple of plugins for the OS to make my workflow with it even better.

See? AI is not all bad. Society is just focusing on the bad examples because it is what we humans remember more.

I have built stuff with AI myself, and I'm either lucky or my design background has helped me actually make all the things I've built meet their goals. From hundreds to thousands of users.

But that's NOT the norm. I am in a privileged position also because of my reach.

Per aspera

Will Astra save software? Or maybe the next step, the Galaxy will?

We don't know. Things accelerate fast. Why everything happens so fast now?

To an extent, because the AI labs are running out of money.

The GPUs they have are getting old and will need replacing soon.

Investors are weary of circular financing and the "bubble popping".

But right here, right now, we have more software than ever and most of it is just plain bad. Even from respected brands.

It may be too late for the models to reverse the distrust trend anymore. People still need software, but the biggest change is not in the code.

It's in public perception of taste, effort put into making things and attention to detail. And I'm not sure we can put an equal sign between a great Bass player and a DJ yet.

In the minds of the public of course. They will consider craftsmanship as worth paying for. Vibe coding is not associated with craftsmanship. Maybe the models will figure that part out.

So far they haven't.