building a habit
Note: This post was imported and originally posted on leaflet on 2025-06-01
Alright, I managed to write another post, which should hopefully be the beginning of a habit. I'm not sure if I'll always have a lot to write every time I post, but just writing something at least should at least keep me doing things regularly.
work
I've been trying to write as much of my contracted project as possible. It's been nice being able to focus on just one thing, and I think that because I didn't submit anything to Asiacrypt nor TCC, I'll have a good bit of time to prep my papers for Eurocrypt. But, I'm nervous the competition will be stiff. Also, because I'll be on the Eurocrypt PC, I think I'll be limited in the number of things I can submit.
Coming up, I still have a week or two more to work on this project and preparing a presentation. But, after that, I think I'll be working on those other projects in full force, which I'm very excited about.
I'm a little nervous that my projects are too disparate, though. I know that a lot of my reappointment and tenure evaluation will be based on the whole body of work that I do. And, I worry spreading myself between watermarking, PIR, structured encryption, and other interesting projects is going to undermine the narrative around the entire body of work.
So far, this hasn't dissuaded me from just working on what I find exciting, but it's something that I should probably be cognisant of. This nagging has been present for a while (especially while applying to jobs), but I wonder how much I should let it dictate what I spend my time on.
non-work
The get-together we had last night was really fun! I had the chance to meet a lot of my partner's friends from her time in NYC before we moved to the area. It was a really fun chance to grill, relax, paint, and chat.
This week, I finished reading The Frugal Wizard’s Handbook for Surviving Medieval England this week. It was the first time I read something by Brandon Sanderson outside of his Cosmere writing. But, I thought the story was entertaining and a fun sci-fi setting for a one-off book. It definitely has the typical Sanderson writing style and plot structure, though, so it won't be for everyone.
Now, I've started listening On the Edge, which I remember was popular like a year ago. I read Nate Silver's earlier book The Signal and the Noise, which I thought was good, but I'm less interested in this one so far. So, I'm not sure if I'll stick with it yet. I'd like to be reading more fiction, but I'm waiting on Oryx and Crake from the library.
Other than that stuff, I've been playing Blue Prince whenever I can. The game as opened up so much. It's really fun to play and slowly discover more about the world. I play it with a few of my friends on Discord, and it's still such an engaging game as a group, even though it's single-player.
things i saw this week
A New Yorker article about parking in NYC and the ways that people work around parking restrictions
A guide to internet safety for kids by Proton seems like good advice, although I wonder how much will be relevant even in a few years
LA is trying to get air taxis approved for the upcoming Olympics
Lots of crazy stuff in politics. Maybe I'll get the energy to write my own thoughts in the future
questions
What places in cryptography/data structures/CS in general is there a stark difference between the perfect (zero-error) version of something and the statistical version? (asked by Gabe in the context of perfect vs statistical ORAM)
Some easy examples are symmetric encryption and single-server PIR, which don't have efficient perfect versions.
Maybe also things like zero-error hashing vs noisy hashing?
There are also examples like weakly super-invertible matrices have a similar efficiency bonus compared to the error-free super-invertible matrices
What is the best way to generate LLM responses privately? (related to some questions Meenakshi asked)
There's a company Zama that is developing a lot of private machine learning, but their methods are highly interactive (I assume because FHE is too expensive)
I wonder what the most likely/common approach taken will be and what problems will occur for augmenting the LLM with like additional instructions or watermarking
My guess is that most "private LLMs" will just be run inside trusted environments in practice and won't rely on strong cryptographic primitives
Which areas of crypto need to be revisiting and given more concrete security treatment?
A lot of cryptography papers (including mine!) get away hiding things in negligible functions or have a large dependence on the security parameter
I'm no veteran of the field, but this seems like it'd be impractical for applied cryptographers to make use of these schemes. (Of course, sometimes the theoretical results are already interesting enough)
Thanks for reading to the end! Here's a picture of our cat as a reward: