Sean Moubry

We dropped a new music discovery feature in Earbuds called Dreamer.

You know those times you understand something COMPLETELY, but no one else does? And you can’t understand why other people can’t understand it? And then it turns out you’re wrong? That’s so funny.

Error rate is as close to a North Star metric as you’re going to get in software engineering because most other issues (latency, throughput, utilization, exhaustion, data integrity, etc) will eventually manifest as an error. I mean, once you build something people want.

You think the Queen ever got asked for the Wi-Fi password?

Of all the technology choices we made building earbuds.app, using tailscale.com is the only one that has just worked from the start, has never given us a headache, and just keeps getting better.

In case you were wondering.

It’s so easy to overfit building recommendation systems. Not just in the obvious technical sense, but in a broader sense too — the experience of an app often degrades the more you give people what you think they want. More of what you like isn’t always better.

If an LLM gives you invalid JSON, you can send it right back, and ask it to fix the invalid JSON at zero temperature. This works pretty well, at least for simple documents. Imagine if you fine-tuned an LLM specifically for correcting JSON syntax errors.

I spent way too much time re-learning Photoshop and Final Cut Pro to make this little demo video.

How many tokens is your codebase?

This is pretty crazy when you think about it: /bin/bash -c "$(curl -fsSL raw.githubusercontent.com/Ho…)"

The Browser Company is building Arc on Windows using Swift. speakinginswift.substack.com…