Even bigger and more expensive models aren't necessarily better for your use case. And how do you rigorously test prompt changes?

With evals! I've added evals to my open source project, The Archive. I wrote about doing so here:
https://www.abramjackson.com/artificial-intelligence/the-archive-pt-3-dont-hack-away-on-vibes-alone/

Everyone is talking about Cursor, the AI programming assistant. Why is it so good that users claim to be 50% more productive using it?

Yes, the interface is excellent, but the real reason is that language models are incredible at programming. Learn why in my article:
https://www.abramjackson.com/artificial-intelligence/why-programming-benefits-the-most-from-ai-assistance/

🚀 Does the moon need GPS? 🌕 I think we can do better than a $100B satellite constellation!

DARPA asked the research community to suggest breakthrough approaches to lunar navigation. I believe future lunar colonists could navigate on the moon with just their smartphone cameras. 📱

🌞 The sun, earth, and stars are more visible from the moon’s surface
🌍 Entirely distributed, and thus resilient, scalable, and flexible
🎯 Much more accurate than inertial navigation!

What would be better than the Humane Pin or the Rabbit R1? If it ran locally and was rock-solid every time. - and maybe if it was a cheap app instead of a $1k device. And used the latest techniques like MCTS?

Introducing The Archive! My goal is to combine Wikipedia and an LLM into a local and durable app or device. It's already working, and I've started a dev blog with what I've been learning so far.

https://www.abramjackson.com/artificial-intelligence/the-archive-pt-1-purpose-and-beginnings/