Ten Links from This Week

Cassia Scheffer

08 Mar 2025 — 4 min read

LLMs

As a side note, I've called this category "AI" in the past. But most of these links are actually about LLMs, or predictive text thingamajigs that appear somewhat smart sometimes. I still see a future where my job changes dramatically because of LLMs, but I'm tired of the hype. I'm trying to focus my learning about LLMs on to practical side: How do they work? How can I make an AI assistant?

With that said, on with the links.

LLMs Might Help Us Make New User Experiences

This is an intriguing post about how LLMs might fit into dynamic, generative user experiences. More interestingly, it is a post about how interfaces need new paradigms as technologies allow for new devices.

Make Your Own Coding Assistant

Do coding assistants seem magical? They're not. They're a collection of prompts and code wrapped around an LLM.

From Design doc to code: the Groundhog AI coding assistant (and new Cursor meta)

Okay, but How do LLMs Work?

One thing I love the most about this is that the author breaks Claude when they ask Claude to explain how it works.

Here's another take on the same subject that breaks down how predictions are made.

Weird Computers

TL;DR LLMs are backward computing. They are very good at being generally correct but almost always specifically wrong.

LLMs Are Weird Computers

A perspective on AI models as an inverted computing paradigm

Code

Remember vibe coding? Big eew. Here's the opposite side of this. You don't need to "just talk to superwhisper." People have been coding with their voice for a lot longer than LLMs.

API Performance Testing

A good breakdown of more than just p99 metrics. Having worked on an API team for a few years now, I'm a bit tired of performance testing that just looks at one or two metrics and says, "Yup, it's bad."

Code Reuse

I am tired of DRY code, and this quote explains it all:

What was already complex and convoluted is now even more complex and convoluted, and half of the code is just function calls. This doesn’t make the code any easier to understand, but it does make it almost impossible to work with.