This Is My Next
4 years ago, I left Pocket Casts, the app and company I co-founded in search of the next thing. Four long years I've worked on that new thing as a fairly open secret. But the project started 10 years before that, long before I joined. It's one of the longest development and iteration times I've ever seen. And now...
In 4 short days, it will finally be released.
Skala will be here, come watch the countdown with me.
I Was Wrong About AI
Back in February of 2026 (remember those days, wow, good times) I wrote a few posts about AI which I thought would age really well. I had seen so many hype cycles like this before, and all the people breathlessly marketing them at me. And I had tried to get AI to build me a podcast app, it was beyond a joke. It tried to store my podcasts in one giant JSON file. It made so many basic and fundamental mistakes that I wrote it off as people breathlessly building 70% of an app and calling it a day. These people hadn't yet realised that they would have to live with the consequences of all the weird and wonderful decisions their LLM made, I told myself.
Now it's the end of 2026. Claude Opus 5.5 is here, and if I'm honest it has blown my mind a little. It no longer makes those kind of mistakes (though it's by no means perfect). With enough care, thought and testing, you probably could genuinely get it to build you an entire podcast app with no programming experience. And more. I'll be honest, I don't know what to make of that. I'm not telling you how to feel about it either. Should you embrace it? Should you fight against it? I honestly don't know. But pretending it isn't here doesn't seem like an option either.
I reassure myself, and the future of my career, by telling myself that there's no substitute for taste. For product knowledge. For saying no, that's not quite right one hundred times until you get what you want. But is that actually true forever, or just for now? Can we train an LLM, or put tooling around it, to get to a place where you can automate quality, taste, the thousand nos for every yes? I think we can't, but I'm far from certain.
And now I am finding I can do things I previously wouldn't have attempted. I'm building a game in my spare time, something I've always wanted to do but just never had the time to dive into. I'm still coding some of it, but if I'm honest I'm also just letting Claude run rampant so I can learn things. It's a fun experience, do some play testing, iterate, do some more play testing, more iteration. I now have an experienced developer who I can bounce things off of and architect things with. Is this costing another developer a job? No, I just never would have built the game in the first place. I'm also sketching things in a notebook and getting it to build the assets for me. I'm less easy about this, and I think the gaming community is as well. Art does feel like it should have soul, and an artist. Is me sketching things and coming up with all the ideas enough to make it art, or have I removed the soul? Again, I have no answers. I did play test it on my two younger nephews though and they seemed genuinely excited about all the little details and playing the game. That was always the dream for me, make something fun, bring people joy, something games have done for me for decades now.
I was so right about the crypto bro's building the world's slowest distributed database being a useless endeavour, and it was. So useless. The AI hype cycle had all the same hallmarks, it even re-used their discarded hardware! But this...this is something else.
So What Can The Intern Do Then?
In my previous post, I poked fun at people who don't understand what developers do and want to loudly proclaim that AI has already replaced, or is months away from replacing us. I still think that's completely ridiculous, and honestly it would be amusing if it didn't betray just how little some people know about Software Development.
Now you might have read that post and decided "Old Man Rusty hates AI!". And you'd be wrong. Don't get me wrong I don't love that some Silicon Valley elites hoovered up the worlds data without asking, and then set some power grids alight using it all. I don't love the ethics of these companies. I don't love their attitudes. I don't love just how much money and resources are going into all of this based solely on hype and dreams. I don't love that their answer to all the critiscm is "yeah but in 6 months AGI which will solve all the worlds problems!". Buuuuuuuut. And yes, it's a big buuuuuuuuutttttt. If you want to put all that aside (and mad respect for you if you don't) the annoying thing about it all is that this isn't Crypto 3.0. There is actually some kind of useful tool here for people in my industry. They haven't just invented the worlds slowest, most expensive to run and useless database you can put links to JPEGs on.
So if I don't think it can replace me, and is nowhere near replacing me, is it still useful? In my experience, yes! I use a combination of Codex and Gemini CLI regularly to help me get work done. Before you ask I can't afford Claude, I am but a poor humble small time developer and those two are just cheaper (yes I know, heavily subsidised, but again this post isn't about the negatives). What it's great at might seem a bit confusing at first. It's great at the very obscure, detail orientated very small things. For example, you have an unsymbolicated C++ or Swift crash log. I don't care how much industry experience you have, 95% of us have NFI how to decipher one of those, and the other 5% are liars. We often struggle even just to symbolicate them. LLMs don't. They excel at language tasks, and this dear reader, is another kind of language. They can read it very well. They can interpret it, and they can then tell you what probably caused the crash and where it is. That is, to put it bluntly, insanely useful. What might have taken me hours, is now minutes. Great! Or you want to convert Objective-C code to Swift code, guess what, great at that too because it's another language based task.
Now ironically, it's also very good at the big wishy washy things. You might think that's counter-intuitive but it has so much training data that creating something which looks kind of like what you want isn't hard. Again, it's other great skill is imitation. You show it a picture of an app, it can make something like that very quickly, ignoring the details and rough edges of course. A few months ago we wanted to build a vocoder at Bjango, but we had no experience in doing that. We wanted to know if it was even possible, and if so how. So we prompted some LLMs and then later that day we had a working prototype. Working is very generous here, it didn't actually vocode, but it showed us what we wanted was indeed possible and we just had to work through it. Marc and I gave a presentation about it if you want to learn more. We ended up spending 3 months re-writing everything the AI did, but it was insanely useful to have the way forward illuminated.
The other example is more recent. I know nothing about neural nets and machine learning, but there was a project we are considering doing that needs me to learn this. So again, I turned to my LLM friends who've hoovered up the worlds data about these processes and got them to get me 70% of the way there. They explained about TensorFlow and PyTorch and Epochs and Losses and training data and so much more. I could have found all that on my own, but to have that knowledge to draw on and get me close enough to a working prototype for me to do the rest on my own was invaluable.
It's also useful if, like me, you work alone and you don't have any other humans to do code reviews or talk through problems. You can get an LLM to do that and it imitates a developer well enough to get you to realise where you might have made a mistake or where you might have gone wrong. They are also trained to be insanely upbeat, which while sometimes annoying can be a nice change from dealing with real humans.
It's also useful for what I call "fun stuff". Things that you will only use personally once or twice and you don't care about the code quality. Or things where you want to prompt an LLM into building you a fully working prototype of an idea before you go and replace everything it did with your own code. Move this up. Change this screen. No lets try linking this to this. Ok that's no good give the image a gradient fade. And so on until you get what you want. You'd be surprised how easy production ready code is to write once you have taken the time to actually prototype something and work through all your ideas.
Finally and more soberingly I know some people out there are actually using it to do their job, managing agents, scheduling tasks, performing reviews and getting it to submit code to production for other people to use. To those people I want to say two things. One is I see you, you've found a new toy and you want to see just how fast it goes, I get it. In the opinion of this humble author though, one day you're going to slam that shiny toy right into a tree. Might not be today, might not be tomorrow, but one day it's going to come back to bite you. If you're shipping code at a velocity where you don't understand it and you're not actively working on it don't forget that ultimately you're the one responsible for the outcomes. When someone can't get part of your app to work because an LLM decided JSON was an amazing storage format for indexed, sortable data. When you leak all your customers personal details to the internet. When things break and the bots just can't fix them without breaking more things...well...that's all on you. And that, dear reader, is why the title of this post is what it is. Treat everything that comes from an LLM as if the intern that just started this week did it. Double check it. Triple check. Don't assume anything it told you is true, because that's the actual job you're paid to do.
AI Developer Challenge
Ok, I've had enough. The hype train around AI is just writing cheques the technology simply cannot cash. I'm not saying it doesn't have its uses, it does. Please don't read this post and hear things I'm not saying. I'm just fed up with the "AI is inevitable, developers are out of jobs in 2026" crowd. I didn't jump up and down saying all lawyers were out of jobs when an LLM passed the bar exam. I didn't say journalists and writers were doomed when an LLM could write something that soundly vaguely like a journalist would. Why? Because I recognised there was a massive delta between imitating a journalist/writer, and actually being a one. That's called nuance, I'm hoping the Internet learns about it one day.
A lot of people seem to think that if the code compiles that the job is 90% done. The truth is, writing code that can compile and pass a test is literally only maybe 10-20% of my job. If the goal was simply to write something that could compile and pass a test, yes, I would indeed hand in my badge and go find a beach to retire on. If the job of a laywer was to pass a bar exam, they too would all retire. If the job of a journalist was to put the right amount of paragraphs into articles, they would too. That simply isn't how development, lawyering or writing works. Allow me to explain via two examples.
Let's start simple. There was a crash in Robot Rocket that a customer sent in. It was an unsymbolicated C++ crash so I turned to Gemini CLI to try and help me diagnose it. I told it where to find the crash ips file, what the issue was and that I wanted to know which part of my code might be responsible. In about a minute it figured out the exact line, and offerred a fix. It wanted to add a null check in front of a null variable to fix the crash. And dear reader, that would indeed have fixed the crash. Job done, right? Wrong. That's literally what I'd expect out of a developer fresh out of the education system and in their very first day at work. Because as any half decent developer knows when a crash occurs your job isn't actually to fix it. It's to determine why it happened, and what the correct way to fix it is. The code crashed because you didn't anticipate something. In this case I wrongly assumed that a method called "initialize" in an API would be called before other methods, and I was wrong. In other words, just adding a null check was ignoring the fact that the method it crashed in was setting up some configuration on an object. If you just add a null check, guess what, no configuration and suddenly you have introduced a very subtle bug that will take you much longer to find. This would lead to someone saving some important config for say a live gig with your app, opening it at the gig, and not realising half their config might not have been loaded properly. Now could I have sat there prompting an LLM into fixing it properly, sure, but I bring this example up for a reason. The LLM isn't a developer. It doesn't think like one. It doesn't have enough context to reason properly through complex projects and what the implications of making changes truly are. It's only good at imitating what a developer does.
But enough simple examples, let's go concrete. A few years ago I wanted to build my own simple podcast app, for just me. I wanted to be able to add podcasts, have their episodes sorted nicely in a list and be able to play them easily at any time. I wanted to support streaming and download. I wanted to support CarPlay. I wanted it to be so simple and bullet proof that I'd never have to think about using it. So I sat down and over the space of a week or two I coded it up. I've used it ever since and it has indeed been bullet proof. Maybe once or twice a year I might clean something up or tweak something but after those initial 2 weeks I haven't significantly changed any part of this app. It's just worked the way I wanted it to.
And isn't this exactly what the AI hype train is promising? That the AI is now good enough to do exactly the above. To replace me, to be a developer. So if that's true, here's my challenge: I want you to build an app like this for yourself. I want you to prompt (no coding allowed) your way to a fully featured podcast app. I'll even give you mine as a template:
It's a super simple 2 tab design. The first tab shows all the podcasts you've added:
</p>
The second tab shows your episodes:

And finally there's the CarPlay integration, again, not complicated:

And if you manage to get to a fully featured, bug free app that does all of the above and you love using each and every day then I will be forced to admit that LLMs are far more capable than I am giving them credit for. Again, don't hear what I'm not saying. I don't care if you get 70% of the way there and it looks kind of like my app. That's not impressive, that's an LLM imitating what I did, not actually doing it. I want a 100% complete and working app. And I want it to be built by a non developer. Someone who doesn't really know how to code, but knows how a podcast app should work. Because that kind of person would be able to hire a developer and be able to reproduce the above. So if the LLM is now good enough to be a developer, and replace one, then it should be able to do this without any issues.
If you've managed to do that, please reach out to me on Mastodon and claim your trophy. After all, perhaps I'm just holding it wrong?
Robot Rocket
We just released our Vocoder plugin, Robot Rocket to the world. TL;DR: you can download it here.
For a slightly lengthier introduction about what we did and why, Marc Edwards and I gave a talk at Dev World 2025 about it, which you can watch here on YouTube.
If you're at all into audio plugins or just want to support us, please, check it out and consider buying it. It's a once off purchase for something you can keep in your audio arsenal forever.
Also, it looks pretty damn cool:

