Better GPU: The biggest improvement was being able to rent better GPUs and massively expand the dataset—this really boosted the model’s capabilities.
Expanded Context Window: I’ve effectively doubled the model’s “memory,” so it now handles shortcuts with 10-14 actions much more reliably. In testing, it’s done a great job keeping all the steps straight, even for complex workflows.
Fewer Hallucinations: One of the biggest complaints was the model inventing random actions or looping endlessly. After a ton of fine-tuning, that’s mostly resolved. If it still happens, please let me know!
Speed Boost: Probably the biggest change. The model is about 7–8x faster than before, thanks to optimized code and better hardware. No more waiting 15 minutes for a shortcut.
I’m happy to chat or troubleshoot if you notice anything weird!
I also wanted to tell you more about behind-the-scenes on how I got here and where things are going:
The Process: I started out playing around with different LLMs (openai, llama, qwen ......), then spent countless hours and a decent chunk of cash on GPU time to fine-tune a custom model specifically for Apple Shortcuts. I wish it was as simple as just wrapping ChatGPT, I had to really teach it a literally new programming language that is behind Apple Shortcuts.
Challenges: Biggest one was making the model robust enough to handle more complex workflows. It needs to be creative and from a simple prompt generate and functional program that might not correspond to the prompt.
What’s Next:
Customization: I’d love to let people upload their own Shortcuts and ask the tool to expand or refine them.
Speed & Scale: Right now, everything runs on a GPU that’s... not cheap. I know how to speed it up, but it’ll take more funding and user feedback to justify the investment.
More Training Data: I know how to create a bigger dataset so the model gets even smarter.
Feedback: If you run into any glitches or notice something off, please let dm me! You can also like or dislike the generated shortcuts. I’m especially interested in hearing how well it handles more advanced requests.
Thanks again, and I hope you enjoy using the AI Shortcut generator as much as I enjoyed creating it!
So, my app has been broken for almost a year. I realized it was getting huge traffic last weekend when I created a new app. I added it to Simple Analytics and noticed my other app had thousands of views. I checked Google Search Console and saw it had been getting thousands of clicks this whole time! I finally fixed it, but now I need your help.
How should I analyze the traffic and what pages should I create to maintain the traction without scaring Google off? Any tools or recommendations would be greatly appreciated!
Good question. I have to pee frequently. I've been examined by a doctor and it's just that some people have smaller or larger blisters. I think there are a lot of people who suddenly have to use the toilet during a movie. From old people to children.
Ok, I've just checked it out, but to be honest it seems cumbersome and you have to download the app. Some may prefer RunPee, but who knows, maybe my app will find fans too.
Hi, I have created a simple app that allows you to take a quick picture of the menu and provides a history and explanation of complex ingredients. It also provides wine and dessert pairings!
I started creating this app while I was buried in my own study notes at college. I found a way for making question-based study guides and realized how much they were helping me. I thought, wouldn't it be cool if I could get some AI help with this? Enter ChatGPT.
What started as a quick fix for my study problems turned into a fully automated system python script that did all the grunt work - scanning textbooks/hw/exams, cleaning up data, and generating custom guides. And, why stop there? I added a feature to chat with the AI about notes, homework, you name it. This wasn just making studying easier.
Seeing how much it helped me, I decided to share it with the world and it quickly became a serious side project. It's been free for the past one and a half months, and people started using it. But to keep things running smoothly, I needed some help. That's where Stripe came in. Now, I'm just trying to keep the ball rolling and cover operational costs while helping students study smarter, not harder.
Really hope you guys find it as useful as I have, and I'm all ears for any feedback or suggestions you might have!
Hey there! I've changed the page to generate a curl request that will properly call the function. By default it is looking for a POST request (even if data is empty).
Better GPU: The biggest improvement was being able to rent better GPUs and massively expand the dataset—this really boosted the model’s capabilities.
Expanded Context Window: I’ve effectively doubled the model’s “memory,” so it now handles shortcuts with 10-14 actions much more reliably. In testing, it’s done a great job keeping all the steps straight, even for complex workflows.
Fewer Hallucinations: One of the biggest complaints was the model inventing random actions or looping endlessly. After a ton of fine-tuning, that’s mostly resolved. If it still happens, please let me know!
Speed Boost: Probably the biggest change. The model is about 7–8x faster than before, thanks to optimized code and better hardware. No more waiting 15 minutes for a shortcut.
I’m happy to chat or troubleshoot if you notice anything weird!