Kernelbench Agent
Can we get AI to optimize kernels agentically?
Can we get AI to optimize kernels agentically?
What does Jack read?
It says "3352 AI TOPS" on the tin, can we actually get there?
I try out claude code. It's pretty good!
No more nice editable json for settings, the AI world demands (checks notes) a JSON blob shoved in a cell in an SQL table ๐คฆ
An in-depth guide to leveraging Cursor's Agent Mode for test-driven development, with powerful workflows that maximize AI assistance without sacrificing code quality.
A comprehensive guide to getting the most out of Cursor and other AI-assisted development environments with practical workflows and hidden features.
A stunningly-good CoT model, with even better distilled children.
A quick trial of Claude's new computer use feature.
My thoughts after working on AI for a few months
How to make an Optional<T> into T when we're sure
A detailed look into Jack's personal workflow and productivity tools.
"Hey please generate a json with these fields" never ever works. Can we do better?
No test facilities? Edge cases that are hard to reproduce but most apps run into? Just another day on Apple's wild ride.
Compiled Docc + Blocking file calls = ๐คฎ
I like Bloaty! I also like cool treemaps. What if I could combine the two?
uFeatures offer lower compile times, better modularity, better testability, stronger isolation, and more! Why wouldn't I use them?
SPM modules are laggy and stuff. Are there better options?
It would be nice to catch fatal assertions, and magically turn all functions into throwing functions. Is it possible?
Porting an app to a uFeature architecture leads to a doubling in binary size. Why? How can we fix it?
Trying to make realtime and offline cool audio effects without ffmpeg on iOS. Did someone say undocumented C APIs? I wish they didn't!
What if you want an app with different downloaded weights for different devices?
I never ended up writing up the stable diffusion port and immensely regretted it, so this time y'alls get to hear about this odyssey of a port.
I wanna use llama, but it doesn't work on Apple Silicon! Can we fix it?
Trying to get those nice JS testing semantics in Swift
If you record all of the actions in an app, assert all of the states at every step, and record all of the side effects, you can replay the recording in a test and verify that the app behaves the same way and investigate why. This is a post about how I built a test recorder for TCA.
A post for my wandering mind during the couple frozen minutes after any file change in Xcode.
A workaround for non-nominal types that donโt conform to Equatable, but really should
When you want to inject a normal C function, things get tricky
Hard to non-destructive trim!
The alternative to 'Speaker' is, apparently, 'Receiver'.
Maybe, one day, Apple will switch to using Swift and we'll see a reduction in SDK bugs.
The curse of dimensionality meets (ai generated:) the curse of copilot
Using runtime declarations as a feature in an otherwise strongly-typed language is not a great idea.
Hopefully this is the beginning of the end of writing Python scripts in Swift code.
A -> B -> C, but a metal file in B needs to include something from A?
A used in B used in C, but A relies on state owned by C.
An abandoned attempt to analyze Victoria 2 savegames and do cool website stuff with it. Perhaps I'll pick it up later?