I finally plugged GPT-3 into @pastelapp to generate color palettes from text prompts, and it just does exactly what I hoped it might do.
-
Prompts are taken as-is, so "a gallery of sample palettes to include in an art app” turns into all of these with no other code. Names provided by the generator
While I don’t intend to ship GPT-3 as a user-facing feature in Pastel, I can certainly use it on the development side to build collections of palettes to populate a gallery feature I was toying with. It will make it very easy to pre-fill with content based on themes
-
While I don’t intend to ship GPT-3 as a user-facing feature in Pastel, I can certainly use it on the development side to build collections of palettes to populate a gallery feature I was toying with. It will make it very easy to pre-fill with content based on themes
Today's task in @pastelapp is to rewrite the color wheel to move it off third-party ObjC code and bring it in-house in Swift. Almost done — it’s so much smoother to interact with now, even if there are still a couple bugs to work out. Pastel is still a ways off being Swift-only, about 6,000 lines to go, but it’s nice to be able to get major improvements as I do-over each class
-
Today's task in @pastelapp is to rewrite the color wheel to move it off third-party ObjC code and bring it in-house in Swift. Almost done — it’s so much smoother to interact with now, even if there are still a couple bugs to work out. Pastel is still a ways off being Swift-only, about 6,000 lines to go, but it’s nice to be able to get major improvements as I do-over each class
Next up for visionOS: @pastelapp. A few dozen build errors to start with…
-
Next up for visionOS: @pastelapp. A few dozen build errors to start with…
…well that's certainly a look
Time to dive in
-
…well that's certainly a look
Time to dive inBetter, but I can see I'm going to have plenty of elements to rewrite ahead of me
-
Better, but I can see I'm going to have plenty of elements to rewrite ahead of me
Some of this UI does look fun, but it's gonna be one hell of an OS to window-manage
-
Some of this UI does look fun, but it's gonna be one hell of an OS to window-manage
I think the color picker makes sense as an ornament, but it is pretty hard to gauge some of this stuff from a simulator
-
I think the color picker makes sense as an ornament, but it is pretty hard to gauge some of this stuff from a simulator
Going in another direction for @pastelapp. I will have to rewrite the sidebar as a collection view (it’s still UITableView, of all things), and I'm glad I spent time this week working on a new color picker tab bar because I'm gonna need it. Overall I think it's going to require a lot more effort than Broadcasts, but it's mostly stuff I needed/wanted to do anyway
-
Going in another direction for @pastelapp. I will have to rewrite the sidebar as a collection view (it’s still UITableView, of all things), and I'm glad I spent time this week working on a new color picker tab bar because I'm gonna need it. Overall I think it's going to require a lot more effort than Broadcasts, but it's mostly stuff I needed/wanted to do anyway
Starting to make preparations to branch @pastelapp to v3.0, so I can work on some larger changes coming down the pipe. I think that will include bumping the minimum OS version, which is currently macOS 11 & iOS 15. I will be dropping support for macOS 11 in this release, for certain, and maybe push even higher depending on timing and how development plays out. This is probably a 2024 update rather than for the iPhone launch this year
-
Starting to make preparations to branch @pastelapp to v3.0, so I can work on some larger changes coming down the pipe. I think that will include bumping the minimum OS version, which is currently macOS 11 & iOS 15. I will be dropping support for macOS 11 in this release, for certain, and maybe push even higher depending on timing and how development plays out. This is probably a 2024 update rather than for the iPhone launch this year
This is the time post-WWDC where I try to get all my projects building to a shippable state again with the new Xcode after working on experimental tangents and rabbit holes for months
I think everything in my dev version of @pastelapp works again now, but I can't remember what I broke.I'm doing a pass at a bunch of little things, shuffling some bits around, updating iconography etc
-
This is the time post-WWDC where I try to get all my projects building to a shippable state again with the new Xcode after working on experimental tangents and rabbit holes for months
I think everything in my dev version of @pastelapp works again now, but I can't remember what I broke.I'm doing a pass at a bunch of little things, shuffling some bits around, updating iconography etc
There is no ‘iOS 17 update' for Pastel, as the new system features don't really touch upon anything in the app. Kinda gives me some space to think up new features to build as I work towards v3.0. The app is in a great place, I think, so I don't feel like there's any pressing need to dig in right now. Some of my other projects need the attention a lot more
-
There is no ‘iOS 17 update' for Pastel, as the new system features don't really touch upon anything in the app. Kinda gives me some space to think up new features to build as I work towards v3.0. The app is in a great place, I think, so I don't feel like there's any pressing need to dig in right now. Some of my other projects need the attention a lot more
Making progress on @pastelapp today. I've walked back some of the more-complex changes I was making for visionOS, and taken a pass at rounding out all the sharp edges. There is still a lot to do, but I think it's turned a corner now
-
Making progress on @pastelapp today. I've walked back some of the more-complex changes I was making for visionOS, and taken a pass at rounding out all the sharp edges. There is still a lot to do, but I think it's turned a corner now
An awful lot of bringup for an existing iOS view controller on visionOS involves three things:
• Change your background color
• Round out your shapes/set your button configurations to the system rounded style
• Add roundrect hover stylesHere's a before/after on my wallpaper generator; just a handful of one-line changes
-
An awful lot of bringup for an existing iOS view controller on visionOS involves three things:
• Change your background color
• Round out your shapes/set your button configurations to the system rounded style
• Add roundrect hover stylesHere's a before/after on my wallpaper generator; just a handful of one-line changes
TL;DR much of this stuff looks incredibly daunting when you first run your iOS app on visionOS, where everything has the wrong color and style, and feels like a mountain of work ahead of you. But actually, it's not as bad as it seems
-
TL;DR much of this stuff looks incredibly daunting when you first run your iOS app on visionOS, where everything has the wrong color and style, and feels like a mountain of work ahead of you. But actually, it's not as bad as it seems
Here’s @pastelapp with a few more passes around the app. Honestly? I could probably publish this build without changing a whole lot else
-
Here’s @pastelapp with a few more passes around the app. Honestly? I could probably publish this build without changing a whole lot else
I built this palette-generation feature two years ago for @pastelapp, hoping that Apple would ship LLM APIs I could use instead of GPT-3. Womp womp.
Here's hoping tomorrow brings some change here…
Steve Troughton-Smith (@stroughtonsmith@mastodon.social)
Attached: 1 image I finally plugged GPT-3 into @pastelapp to generate color palettes from text prompts, and it just does exactly what I hoped it might do. How I wish I could do this with an on-device Siri API call; I don't intend to ship with a dependency on OpenAI, so for now this is a 'maybe someday’ feature
Mastodon (mastodon.social)
-
I built this palette-generation feature two years ago for @pastelapp, hoping that Apple would ship LLM APIs I could use instead of GPT-3. Womp womp.
Here's hoping tomorrow brings some change here…
Steve Troughton-Smith (@stroughtonsmith@mastodon.social)
Attached: 1 image I finally plugged GPT-3 into @pastelapp to generate color palettes from text prompts, and it just does exactly what I hoped it might do. How I wish I could do this with an on-device Siri API call; I don't intend to ship with a dependency on OpenAI, so for now this is a 'maybe someday’ feature
Mastodon (mastodon.social)
Finally trying Apple's Foundation Models for palette generation, side by side with what I generated with ChatGPT.
The first image is the prompt "Planets of the Solar System" with Apple's on-device LLM, the second image is the same prompt using GPT-3.5 as of 2 years ago.
It can certainly do what I need it to do; whether the results are any good is an open question. I think it's inoffensive, and harmless, at the very least, so maybe I can move forward with it
-
Finally trying Apple's Foundation Models for palette generation, side by side with what I generated with ChatGPT.
The first image is the prompt "Planets of the Solar System" with Apple's on-device LLM, the second image is the same prompt using GPT-3.5 as of 2 years ago.
It can certainly do what I need it to do; whether the results are any good is an open question. I think it's inoffensive, and harmless, at the very least, so maybe I can move forward with it
As a proof of concept, it confirms that the Foundation Models can indeed be applied to any kind of silly little task, and it’s so easy to get structured, deterministic output from it using the new @-Generable macro. No JSON, no parsing, just perfectly formed structs as output
-
As a proof of concept, it confirms that the Foundation Models can indeed be applied to any kind of silly little task, and it’s so easy to get structured, deterministic output from it using the new @-Generable macro. No JSON, no parsing, just perfectly formed structs as output
Even though the concept works, the performance is a little worrying. It takes an M1 iPad Air about 30 seconds to generate a handful of color palettes — which are just some hex codes with a name, nothing complex. Streaming the results in one by one becomes a necessity with those kinds of processing times
-
Even though the concept works, the performance is a little worrying. It takes an M1 iPad Air about 30 seconds to generate a handful of color palettes — which are just some hex codes with a name, nothing complex. Streaming the results in one by one becomes a necessity with those kinds of processing times
Enabling streaming, you can see that the Foundation Models API is slow — I would say alarmingly slow — on M1 hardware. I'm not even sure what kind of optimization I might be able to do to speed this up, as it's seemingly purely on the token generation side
It's going to take more than a little care and attention before you start sprinkling this all over your apps, that's for sure. Good to know
