I study electrical engineering at the University of Sydney. You’d think that if I’m still on my phone after midnight, I’d at least be thinking about tomorrow’s lab. Usually I’m not. Usually there’s Juice WRLD in my headphones, something unfinished on my laptop, and I’m telling myself I’ll check X one more time before bed.
That’s how I came across AirJelly.
More specifically, a blog post by its founder. I read it and sent him a message. I thought some of the ideas were interesting and wanted to chat. Somewhere in that conversation, I found out they were exploring AI hardware.
And I thought, hang on. That’s actually relevant to my degree.
Here was a chance to take some of the stuff I’d been studying and turn it into something you could carry around. I hadn’t worked out exactly how it should fit into someone’s life. But I wanted to help build it, so I got involved.
There wasn’t a big cinematic moment. I didn’t close my laptop, look out over Sydney and decide to dedicate myself to the future. I think I just replied to a few messages and realised I was taking the whole thing quite seriously.
Whenever I used AI, I kept having to explain the background first. Who I was, what I was working on, why I was asking. Sometimes I’d get halfway through and find it a bit funny. It seemed to know something about almost everything, but very little about the person sitting in front of it.
To be fair, I wasn’t giving it much. A few messages, a couple of files, a question I’d already tidied up enough to ask.
Everything else was still out there, happening. Whatever I was listening to on the walk to Redfern station. A conversation with a friend that wandered off somewhere unexpected. My decision to get an early night, made after dying far too many times in Elden Ring the night before.
None of that sounds particularly important on its own. But it’s most of what a day actually feels like.
I think that’s when my interest in AI started wandering a bit. I’d think about it while doing uni work, then while listening to music, and especially while gaming. Eventually, all of that found its way back to the hardware I was working on.
Which makes sense, obviously.
You know, I’m a hardware engineer.
1. AI can make the music. We bring the meaning
You know, I’m a hardware engineer, so naturally I spend a lot of time thinking about AI and hip-hop.
The logic is a bit questionable. Luckily, that’s not something I usually check while listening to music.
On the walk from uni to Redfern, I’ll often put on Juice WRLD. Sometimes XXXTENTACION. Depends whether I want my head to quiet down or someone else to make the noise for a bit. Occasionally, one song does both. Pretty efficient.
Some songs take a while to get to you. First listen, you like the melody. Then you hear it again one afternoon and a line suddenly lands. Nothing particularly dramatic has happened. Maybe the lab went badly, there’s a message you don’t know how to answer, or Sydney’s decided to rain after you trusted the sunshine on your way out.
Same song. You’ve filled in the background yourself.
So when I think about AI and music, one of the first things I wonder is how I’d feel hearing a familiar voice sing something that person never actually sang.
Say there’s an AI-generated track that sounds like Juice. The melody works. The emotion feels familiar. Even that little break in a word sounds right. I could genuinely like it. Sometimes your body gets there before your opinion does. You’re still working out how you feel about the whole thing and you’re already nodding along.
Then something catches up with you.
That line that sounds painful never came from a particular night in his life. That breath wasn’t one he actually needed to take. I’m hearing the shape of a familiar expression, but he never made it.
When the voice belongs to someone who’s died, that feels harder to brush past.
Of course you want to hear more. When you love someone’s music, there’s always a part of you hoping you’ll scroll down and find one more track you haven’t heard. But wanting another song doesn’t feel like enough of a reason to decide what that person should say next.
At which point I start wondering whether I’m taking the whole thing too seriously. I don’t investigate the creative history of every song I put on. Sometimes the drums are good, the bass feels right, and I’m walking in time with the beat, briefly convinced that catching a train deserves a camera crew.
And I am interested in making music with AI.
There’s something appealing about having a melody stuck in your head, humming it badly, and being able to say, “Give this a bit more space. Bring the drums in later. Sort of Parramatta Road at night, but I’m actually in a decent mood.”
As an engineering brief, that would probably get sent back. As a description of a sound, it makes perfect sense to me.
If AI could do something with that, give me something I could actually listen to, I’d have somewhere to go next. Take a sound out. Leave a bit of noise in. Listen again to the part I thought was wrong. Maybe that ends up being the bit I like most.
I like the idea that a half-formed thought could become something audible before I’ve completely worked it out.
There’s a catch, though. If every vague idea gets a polished answer straight away, I wonder whether I’d stop trying to figure out what I actually meant.
Something that sounds good is easy to agree with. Especially when you’re tired, you’ve got class tomorrow, and version three is honestly pretty decent. Fine. Version three.
Meanwhile, the stranger idea I started with―the one I couldn’t quite explain―might have quietly disappeared.
I think I’d want AI to help me stay with that idea a little longer. Make it easier to try things, without rushing me towards a verdict on what sounds good. Give me enough room to say, “Yeah, that works. It’s still not what I’m looking for.”
Then again, when I listen to Juice or X, I’m not always looking for something new.
Sometimes I want one particular track. I know when the next line comes in. I know where the beat drops out. I can still remember a stretch of pavement where I used to play it on repeat. The song’s become attached to a few ordinary days. A cleaner mix or a better vocal wouldn’t necessarily do much for it.
Even if I could generate something perfectly suited to my mood whenever I wanted, I reckon I’d still go back to those old tracks. Put on something that doesn’t quite fit today. Let it take me somewhere else for a few minutes, then come back.
You can keep refreshing the music. People are a bit messier.
I’ve just realised I was meant to be writing about AI, and I’ve spent most of this trying to explain why I listen to music.
Maybe I need to work that out first.
Before I ask AI to give me more of it, I want to understand what I’ve been coming back for.
2. I want AI alongside me. I still want a go myself
You know, I’m a hardware engineer, so naturally I’m interested in what AI could do in Baldur’s Gate and Elden Ring.
Specifically, I want to know whether an AI would die twenty times in a row and decide it should probably have a shower.
I would. I just tend to give it another ten attempts first.
I get back from uni, drop my bag, and load up a game for a bit. “A bit” clearly means something different here than it does when I’m working on a lab report. It’s still light outside when I start. Next time I notice the window, I can see my own reflection in it.
Haven’t made much progress. Have become extremely familiar with one particular patch of floor.
That’s usually when I think it would be nice if AI could help me play.
I’m imagining something that can actually watch the screen, work out what’s happening, try an input, and adjust based on the result. I’ve already read the advice. I understand “wait until the attack finishes before you hit back.” The version of me that understands that sentence and the version holding the controller apparently work in different departments.
Giving AI the controls would be interesting. It would have to turn what it sees into an action, then deal with whatever happens next. A step forward commits it to being there. An early dodge is still an early dodge. Saying “I should be more patient here” is easy. Actually waiting on the next attempt takes something else.
I’d want to see how it learns that. Does it notice the mistake it keeps making? Does it try something different? After a spectacular failure, can it explain what actually happened? Or does it produce a very convincing account that falls apart the moment you watch the replay?
That’s the engineering student in me, apparently. One minute I’m looking for something to suffer on my behalf. The next I’m wondering how to test whether it knows what it’s doing.
But say it gets good.
One night I hand over control, and it dodges the attack I keep getting caught by, finds an opening, and finishes the fight. A boss I’ve been stuck on for days is suddenly done.
My first thought would probably be, finally.
Then: hang on, how did you do that bit? Show me again.
And then I’d probably want another go myself.
That surprises me a little. I thought I just wanted to get past the fight. Once something else can do that for me, I start caring about the move I still haven’t learnt. Maybe the help I want changes as the AI gets more capable. First I want it to take over. Then I want it to practise with me. Eventually, I might just want it to clear its throat when I’m about to get greedy.
I’d probably go for the extra hit anyway.
In Baldur’s Gate, I’d want something quite different.
I can imagine having an AI in the party, actually taking part in the mess. Helping decide whether we should go through that door, who should go first, whether that offer sounds suspiciously like a trap. It would need to remember what we’d already done. It would also need to cope with me abandoning a perfectly sensible plan because a stupid idea had just occurred to me.
It finishes explaining the risks, and I go, “What if we nick that?”
I’d like it to ask why, at least.
Immediately turning my impulse into an efficient action plan seems unnecessary. I’m already quite capable of making a mess.
If it knew I couldn’t leave a suspicious chest alone, or that I’d take a difficult fight to keep a particular character alive, it could carry that into its decisions. Same with knowing that “I’m taking this run seriously” still leaves plenty of room for wandering down every side path.
It could have opinions of its own, too. Suggest something I hadn’t considered. Push for a choice I wasn’t sure about. We might have to talk it through. Eventually, I might turn out to be right, which I would remember for a very long time.
If it turned out to be right, I’d probably just say, “Yeah, fair. Let’s keep going.”
Memory between gaming mates has never been entirely balanced.
Of course, an AI that always knows the best move could make the whole experience a bit strange. I could end up checking with it before every conversation, every fight, every turn in the path. Always making sure I wasn’t missing something.
By the end, I’d have been everywhere without spending much time feeling unsure.
And the bits I end up talking about with friends are often the uncertain ones. An answer that sounded wrong the moment I said it. Someone acting before we’d agreed on a plan. A plan we did agree on, then collectively forgot about three minutes later.
So I think an AI gaming mate would need some sense of when I actually want help and when I’m happily getting myself into trouble.
If I ask, “Can I make that jump?”, it should probably know that sometimes I’m asking a question, and sometimes I’ve already pressed the button.
Games start to look like a pretty useful place to work all this out. Has the AI understood the environment? Does it remember what just happened? Can it turn a suggestion into an action? Can it pick up on how I actually feel like playing tonight?
A few good answers won’t tell me all of that. The game keeps moving. We have to find out what happens next.
I’d also get a better sense of what I’m happy to hand over. Getting from one place to another, probably. Repetitive bits, sure. When I’m stuck, we can work on it together. Some decisions I’d still want to make myself, even if I regretted them straight away.
I’m not necessarily loading up a game because I want to get something done efficiently.
Sometimes the lab’s finished, my head’s still going, and I want to spend a while somewhere else. Having something there that can follow what’s happening, help occasionally, and occasionally make the whole situation more complicated sounds pretty good.
As for whether it can beat that boss for me―
If it can, don’t finish it off just yet.
Leave a bit of health. I want another go.
3. To understand me better, AI needs context beyond the chat
You know, I’m a hardware engineer, so naturally I’m interested in the relationship between AI and hardware.
Finally. Something relevant to the degree.
After all that time in my playlists and the Lands Between, I’m back at my desk. Same unfinished coffee. Same loose cables. Something open on my laptop that I’d rather deal with later. A mate texts to ask where I am.
“Nearly there.”
I haven’t found my keys yet.
If an AI asked whether I needed help at that moment, I’d probably say no.
There is evidence to the contrary.
The longer I use AI, the more I notice how much work happens before I ask it anything. I have to work out what I need, remember the relevant bits, and put them into words. Sometimes I have to explain why I care in the first place.
“Help me reply to this message,” for instance.
Sounds straightforward. But who sent it? How did our last conversation go? Are they carrying on a joke? Why have I typed a reply three times and deleted it?
All of that is sitting just outside the chat box.
What I eventually send is a fairly compressed version of myself.
As an electrical engineering student, I should probably be alert to information loss. Usually, though, I just can’t be bothered typing another two paragraphs of background.
Once I got involved with AirJelly, I started thinking more seriously about giving that background another way in.
I’m helping design a portable device that can capture bits of everyday life through photos and audio. The idea starts with something fairly ordinary: people we meet, conversations we have, things we notice. Some of that could help AI understand us better, but a lot of it has already slipped away by the time we open a chat.
Say you leave campus and head into Newtown with a mate to find something to eat. You talk the whole way without anything happening that obviously belongs in a calendar. They mention an interview next week that they’re nervous about. You notice a gig poster and say it looks good. Then the conversation moves on.
A few days later, you have this vague feeling there was something you meant to ask them about.
Things become important after the moment when you could have thought to save them.
That’s a big part of what interests me about the device. A photo or a bit of audio could leave something to come back to. It could also give AI a little more to go on when I need help later. Maybe I wouldn’t have to start quite so many conversations with, “Hang on, let me explain this from the beginning.”
That’s the direction I want to explore. There’s plenty to work out between capturing parts of a day and making them useful.
For a start, having a recording doesn’t mean you understand what happened.
I’m smiling in a photo. Someone might just have told a terrible joke. I say, “Whatever’s fine.” Maybe I mean it. Maybe I’m too tired to make another decision about dinner. The same few words can mean quite different things on different days.
I also get myself wrong. Something feels urgent at midnight, then much less urgent after some sleep. I play a track all week and suddenly don’t want to hear it. An AI with a longer record of my life could become increasingly confident about an opinion that’s no longer true.
I’d want it to leave some room for that.
“You seemed quite keen on this before. Still interested?”
I’d be happy to answer that.
“Based on your past behaviour, this is the sort of person you are” would make me want to switch it off. I barely recognise the bio I wrote for myself last year.
We talk about wanting AI to remember us. But being remembered can feel very different depending on how someone does it. A friend remembering something you dislike is thoughtful. A friend defining you by one thing you said three months ago gets old pretty quickly.
I think a device needs to give you room to correct things, explain yourself, and change your mind. Some moments are worth keeping. Others I’d be happy to let pass. That should be an easy decision to act on. I shouldn’t have to dig through several menus to tell something I carry around, “Don’t keep this bit.”
And everyday life involves other people.
A mate sitting across from me at dinner hasn’t automatically volunteered to become material for my project. If I’m building something to help capture personal context, I need to think about how it feels to everyone around it. Can they tell what it’s doing? Can they comfortably say no? What gets kept afterwards?
Those questions affect whether I’d take it out with me at all. However clever it is, if I have to give a lengthy explanation every time I sit down with friends, I’ll probably start leaving it at home.
This is where the hardware gets very real.
Do I actually want to carry it every day? Does using it feel natural? Does it make a perfectly normal conversation slightly awkward? Something that works on the bench still has to make it through an ordinary Thursday.
I might have just left a lab, still thinking about a problem I couldn’t sort out, and be hurrying to the station. At that point, whether the thing is easy to use matters a lot more than how impressive it sounded on a product page.
That’s part of what I like about building hardware. Eventually, you have to put it into an actual day. Put it in your bag. Walk somewhere. Sit down with a friend. See whether you still want it with you. A lot of ideas only really start answering back once you do that.
With music, I kept coming back to what sits behind a voice. With games, I wanted to know whether AI could follow what was happening and understand why I wanted to play a particular way.
Working on this device brings those questions into my own life. A conversation today. A thought I didn’t finish yesterday. Someone I see all the time who’s never appeared in any of my chats with AI. Could those details help it make fewer guesses when I need something? Could they save me a bit of explaining?
I don’t know how far we’ll get yet. But it’s something I want to spend time working on.
When I first found that blog post on X and heard they were exploring AI hardware, my reaction was basically: that fits my degree.
What’s kept me thinking about it since has been a collection of much less academic things. Why a song gets to me on a particular day. Why I still want to finish a fight myself. Why something a friend mentioned in passing comes back to me later.
Around this point, my phone would probably go again.
“Have you actually left yet?”
If AI eventually knows me well enough, I hope it hears “nearly there” and reminds me to find my keys.
