work

This week I mostly just got to do normal writing. The S&P rebuttals largely haven't come in, so I've been free to just work on the papers with upcoming deadlines. It's always crazy how quickly a deadline can sneak up, and right now I'm pushing for two papers to go to Eurocrypt and maybe even getting another together for a conference before that. Interestingly, these papers are very different but all very interesting problems to me. The topics range pretty broadly between theoretical and practical and also across the subfields. This diversity has kind of re-ignited my own hesitation of working on such a big range of papers. But, I'm still optimistic that if I just keep working on what seem to be useful problems that are interesting to me, then I should succeed overall.

non-work

I started the final book in the MaddAddam Triology! I feel much more invested in the world after the second book than I did after the first, so I'm excited to see how it ends. It's a really great series to listen to casually, because it has a lot of important topics that are worth thinking about, without being too "preachy" about particular causes. Also, the second audiobook actually has the Gardener's songs sung and performed, which are fun little unique songs to listen to.

I also somewhat randomly decided to play Pokémon XD: Gale of Darkness because it's one of my friend's favorite Pokémon games, and I never played it as a kid. And, honestly, I've been impressed with the game so far. A lot of the mechanics are a nice breath of fresh air from the typical Pokémon setting. It's nice to travel quickly on the map and not have to navigate random encounters on different routes, and catching shadow pokémon is the best incorporation of the friendship mechanic I've seen in a Pokémon game. I'm still pretty close to the beginning of the game though, so I hope it holds up!

things i saw this week

• A Terence Tao post about prime gaps with or without rough numbers

• A Reuters article about Perplexity AI making a bit for Google Chrome

• A Politico article about why Wi-Fi isn't very good on trains and especially fast trains

• A WSJ article about the stagnation of home buying and jobs in America

questions

• Is there good research into how to make LLMs interact between themselves in a useful way?

  • I wonder if this could significantly amplify the abilities of the models, because collaboration seems so important for human

• How could tokenizer's lack of prefix-freeness mess up encoding and decoding of messages for LLMs? I feel like this hasn't been in issue in the papers I have seen, but could it be? Are the better/worse tokenization systems?

Thanks for reading to the end! Here's a sleepy picture of our cat from this week: