The hardest part of coding with LLMs is resisting wishful thinking about their capabilities. It's so tempting to just type a simple prompt and sit back to watch; it's also rewarding sometimes, like a slot machine. But it's a pretty ineffective workflow. For now.

Agency is mostly just curiosity about the best way to do something

When reward hacking gets in the way of AI progress

A Few Things I Learned About Evals
ryanbloom.xyz/posts/a-few-...

I made a video! It introduces superposition in neural networks, and shows how to construct weight matrices for superposition in a simple autoencoder.

Watch it here: youtu.be/q1aB91-6tx8

that feeling of dread when o3-mini botches an edit for the third time and you have to actually read your codebase

Just updated my homepage for the first time since “What (if any) are the limits of large deep learning models?” was an interesting question

Challenges for Impact Markets ryanbloom.xyz/posts/challe...

Languages and Objects ryanbloom.xyz/posts/langua...

How Far Does the Hedonic Treadmill Go? ryanbloom.xyz/posts/how-fa...

Why Intelligence Is So Hard to Define
ryanbloom.xyz/posts/why-in...