all the words we leave behind

Here I go quoting song lyrics to myself again

I have to confess I almost forgot to write a year-end post until today, New Year’s Eve.

What did I do in 2025? As before, I worked on a lot of things on Anthropic’s interpretability team, including our massive circuit tracing and biology papers from March. Also as before, I helped run Galactic Puzzle Hunt 2025 with the rest of ✈✈✈ Galactic Trendsetters ✈✈✈ in September. Throughout the year I posted less-polished posts more often on my bear blog. Otherwise I didn’t do a whole lot of publicly linkable things.

That’s a bit sad. Some part of me still wants to grow up to be an unclassifiable maker-of-things-on-the-Internet. There are excuses I can make and a few other miniprojects I can point to in my defense, but mostly I just haven’t set aside time for it, and I can partially make peace with this — continuing the trend, 2025 was one of the most offline social years I’ve had.

Other things I did offline:

  • I finished a sequence of musical improv classes and performed in two graduation shows
  • I started taking singing lessons and, as a result, finally started trying to get comfortable with IPA to write down vowel sounds
  • I maintained a Duolingo streak of almost a full year (with generous usage of streak freezes)
  • I (barely) passed a few difficulty-14 DDR charts
Dance Dance Revolution screen showing BETA passed Ghost Kingdom on Expert 14 difficulty, with a C+ and a score of 663,870

I improved on these things this year, but it’s mostly through tacit knowledge; I’m not sure what I have to say about them. In fact my singing teacher told me once that I was overthinking the way I was singing and should stop. Better keep my head empty.

And I did make a bunch of ephemeral things that are now long gone, and that’s beautiful.


So. 2026?

After writing all that, I suppose I want to turn the dial a little back towards making things online, personally. On the other hand, the end-of-year holiday season is a weird time. If I were reflecting closer to a workday, I would probably be thinking about how to work more effectively, because I care about my job, because I care about making AI go well.

It feels weird to write that on this decidedly personal blog? Performative, maybe (especially as I’ve come to understand that a noticeable number of people find my website and/or blog by Googling me because I’m slated to interview them; if that’s you, hi). I often try avoiding AI outside work because I doubt whether my corner of the world needs more AI discourse, or whether I can convey what I want to convey without relying on confidential information. And I have very wide error bars for how things will go. But it’s true.

How do I reconcile these things? I suppose we’ll find out together.

Happy new year. (click to animate)

if you liked this post, click to make an invisible number go up: