Simulation is still deterministic on CPU but now I write directly into GPU buffers and moved interpolation, culling and other compute all into GPU.
Now renders 3M particles on iPad @ 60fps and uses 1/3 RAM (~400MB for 3M).
But why run 3M on iPad, when I can run 6M on M1 Max ;)
Now I finally have plenty of headroom for collisions! I think I've optimised enough for this week.
Pixl Particles – 3D deterministic particle system.
100% Swift, TRULY cross platform and its FAST!
Super excited to share a screenshot of Pixl Particles – a small part of a game engine I've been working on.
Deterministic here means every simulation using the same seed runs IDENTICALLY on every run AND every platform! Its truly deterministic and as a result supports scrubbing, playback, reversing, out of the box.
I wrote the editor for my son who's 9. He's making the effects for our games so it needed to be on iPad first and easy for him to use :)
The particle system itself runs on ANY Swift supported platform including Windows, Linux, WASM.
Here you can see 1 million particles simulated and rendered @ 60fps – you'll never need that many particles but I needed to push the benchmark. The system is single-threaded with rendering and other compute done on GPU. Data oriented, SIMD, and all that lovely stuff.
The engine can actually render closer to 2M particles but 1M is comfortable. The editor is SwiftUI ofc but I Pixl is also a larger game engine which has a PixlUI library that feels like SwiftUI so I'll be porting the UI over to that and writing it in-engine so it will work on any supported platform as well.
The engine is a BEAST but I'll be sharing visuals and games written with that in the future. This is a custom engine written 100% by me (with some AI assistance ofc) and may eventually be available under license but otherwise closed source for now.
Couple additional stats for this run of 1M:
~350MB RAM
~30-40% CPU (single core)
sʜᴘs retweeted
🎁 Kimi K2.7 and GLM 5.2 are now available on @tryGenerous alongside the most recent @AnthropicAI models.
Code with LLMs on your GitHub repos in seconds with these amazing open-weight models ✨
If we’re no longer free to choose the right tool for the job without feeling the need to justify ourselves, then we’ve lost something far more valuable than efficiency. We’ve lost trust.
I wrote a full essay you can read here:
open.substack.com/pub/shaps/…
I’d genuinely love to know where you stand on this.
Have you ever chosen not to use AI because you believed it was the better decision?
Did you feel comfortable making that choice, or did you feel the need to defend it?
And perhaps the bigger question is this:
At what point does a useful tool stop being a choice and start becoming an expectation?
RetroCalc is a fun calculator with an LCD display, soft buttons, and one very important feature: it helps you write words upside down.
Why? I wanted to share with my kids, one of my favourite old pastimes: typing numbers, flipping the calculator upside down, and discovering words hiding in plain sight.
I could not find one using a proper LCD-style font, so I made one. They loved it.
Its a personal, silly side project with my kids, made to show them that writing code, games, and apps can be playful. Also, calculators should let you discover that 5318008 is not the only joke in the drawer 😉
It is still a working calculator, obviously. It does sums, handles the usual operators, and then quietly encourages nonsense when nobody is looking.
Sharing for those of you who also want to be 12 years old again.
github.com/shaps80/RetroCalc
SwiftUIBackports is in active-mode with all he amazing WWDC announcements in SwiftUI.
So I decided to write a skill that also makes it trivial for agents to use and get a more consistent codebase, less custom code.
Don't forget to follow for updates, backports are coming!
AI Made Me Rethink How Good Decisions Get Made
What if the biggest impact of AI isn’t on implementation, but on judgment?
open.substack.com/pub/shaps/…
When AI can write software an order of magnitude faster, its tempting to be revolutionary instead of evolutionary.
#agentic #softwareengineering
shaps.substack.com/p/innovat…
Can't wait to try this! As always great work from Natalie!
We're about to launch @useSerialDev - an agentic platform that's a mix of GitHub Codespaces, Cursor IDE, Lovable. It’s faster, multiplayer, and more scalable.
If your company could use a multiplayer VS Code IDE in the browser with agents, I'd love to chat.
Desperately looking for a new iOS role. Contract or full time. If you are looking or know someone looking please get in touch. #iosdev #swiftlang #mobile #ios
sʜᴘs retweeted
Mojo 🔥 1.0 is in beta! Beta 1 marks the first step towards finalizing 1.0 later this year, which will bring a new level of language stability. The beta lands safe closures with a new capturing syntax, conditional trait conformance, and major variadic improvements. Plus, Mojo has its own home at mojolang.org.
What I can do with my experience and amazing tooling has no limits. AI isn’t the tool that enabled this for me. My experience and perseverance was all I ever needed and it’s true now more than ever.
If you’re looking for an engineering mindset that’s not limited by tooling, get in touch. I can help you transform your team, tooling and hopefully your mindset.
Open to full time or consultancy roles.
The Vampire Effect — anyone else feel the energy drain after coding-by-prompt day in, day out?
There is real value in using AI for engineering work. I’m not questioning that.
But over the past few months, I’ve had a lot of conversations with engineers who are starting to describe similar side effects.
It starts simply enough. The first few prompts of the day can feel almost magical.
You get that hit of excitement when something works. You’ve built something that might previously have taken days or weeks. Maybe something you couldn’t have built alone.
Then the second phase begins.
Now you have to review it.
Not just check whether the code compiles, but understand the assumptions, the trade-offs, the edge cases, the architecture, the tests, the things it changed that you didn’t expect it to change.
You’re getting “so much” done, but you’re also carrying more context, more diffs, and more uncertainty.
As the day goes on, the context window fills up. The agent starts drifting. The back-and-forth turns from exciting to frustrating.
By the end of the day, you can feel more drained than if you had written the code yourself.
I think part of it is that the sense of achievement changes.
You did create something. Maybe more than you thought possible. But the work can feel less like building and more like supervising, correcting, validating, and absorbing complexity.
That’s a very different cognitive load.
I don’t think this is just an individual productivity issue. I think it’s becoming a team design issue.
How do we structure engineering work so AI helps us move faster without quietly draining attention, ownership, and judgement?
I’m still finding my way through it.
Curious how others are handling this. Are you changing your workflow, your review habits, or your team norms around AI-assisted development?
