James Kelly
@IjkTech
CTO @passio_ai | Astrophysics PhD | Working on the front lines of AI innovation | Keen on solving real-world problems through technology #AI #ML
🚀 I’m building a series of LLMs optimized for compute- and memory-constrained devices—perfect for mobile and embedded systems. The first model is live on 🤗—link in the comments 👇 This is just the start. I’ll be using GPRO distillation to push these models even further 💪
You know the app layer will win because instead of saying <model did it>, users will say <app did it> and what they will recommend is the app, not the model. The app will take extra special care to solve their problem.
Looks like Lambda Labs is my new go-to GPU platform. Running ByteGPT training on a GH200 for 1/3 the price of a slower TPU, and the machine was up and running in just three clicks.
Are TPUs worth the hassle? I set one up to train ByteGPT, wired everything up with torch_xla in Cursor, and it should be running smoothly… but startup is slow, and once it gets going, it’s barely faster than an A6000. Am I missing something?
github.com/groundlight/r1… - interesting. Training with RL is going to unleash smaller models that work on device which will be good for privacy on mobile devices and will be huge for robotics
What’s everyone using for cloud notebooks with GPUs? Been on @HelloPaperspace, but A100s are often unavailable. Colab is OK, but setting up envs and Drive is a hassle. Looking for something simple—no fighting GCP quotas or IAM policies. Any recommendations?
I've been thinking about ways to further optimize ByteGPT for Mobile: if we assume sequence length < embedding dimension, we can unlock efficiency gains beyond the standard linearizations of the q @ k.T dot product. Excited about what this means for mobile LLMs! 🚀📱
🚀 ByteGPT-r1 is almost here! I'm optimizing a powerful r1-type LLM for mobile and updating the Android SDK, but first, I need to tackle some memory allocation battles. ⚔️ Soon you'll have a seamless way to add a useful, on-device LLM to your apps. Stay tuned!
Super-excited about this launch! Our @passio_ai Nutrition-AI is powering @dr_ericberg new Junk Food Meter app design to show you what’s inside your food and help everyone make better food choices!
Introducing my Junk Food Meter app! Scan barcodes to instantly reveal hidden sugars, starches & seed oils. To download, click this link: drbrg.co/3Cwxq5A
A quick weekend project to get refreshed on the JS side of things: a minimum viable chatbot using FastAPI, bootstrap css, and vanilla JS. Hopefully it's a useful starting point for others. Would be great to hear feedback and ideas. github.com/kells1986/mini… #FastAPI #JS
I just finished The Old Man and the Sea. It’s an incredible allegory for startup life: the hubris of fishing too far out, the struggle to bring the catch home, and the nobility of striving despite coming back empty-handed, and battle scarred wisdom that informs the next iteration
Rebranding all Python dictionaries in my codebase to serverless vector databases for funding purposes
United States Trendy
- 1. Good Sunday 48.3K posts
- 2. #HealingFromMozambique 14.7K posts
- 3. Discussing Web3 N/A
- 4. #sundayvibes 4,263 posts
- 5. Auburn 47.6K posts
- 6. Wordle 1,576 X N/A
- 7. MACROHARD 8,665 posts
- 8. Gilligan 6,746 posts
- 9. #SEVENTEEN_NEW_IN_TACOMA 40.3K posts
- 10. QUICK TRADE 2,211 posts
- 11. #SVT_TOUR_NEW_ 32.1K posts
- 12. FDV 5min 2,237 posts
- 13. Utah 25.6K posts
- 14. Market Cap Surges N/A
- 15. Dissidia 7,437 posts
- 16. The CDC 32.4K posts
- 17. Whale - Buy 1,845 posts
- 18. Boots 52K posts
- 19. Kirby Smart 9,098 posts
- 20. Vergil 3,481 posts
Something went wrong.
Something went wrong.