iiinova — where your digital self keeps evolving

Holden Lee

@oldheneel.bsky.social

94 followers 24 following 3 posts

Mathematician & writer

Posts Replies Media

Holden Lee @oldheneel.bsky.social · 14/07/2026

Why do machine learners call them "diffusion models" when probabilists came up with the same process first? Let’s rebrand diffusion models as "stochastic localizers".

020

Reposted by Holden Lee

Juspreet Singh Sandhu @freegaussians.bsky.social · 10/07/2026

Building on the spin-glass & free-probability estimates we developed in PHA 3 @oldheneel.bsky.social, @jnshi.bsky.social and I observed that the "regularity" of the ASL process + conductance can be used to show WPI for a target Gibbs. The framework is applicable under mild regularity conditions 1/3

111

Reposted by Holden Lee

Christopher Honey @chrishoney.bsky.social · 15/06/2026

Some findings from our national survey of U.S. population attitudes toward AI. First, feelings about AI are actually very similar across political parties (and many other demographics). They vary most by people's experience with AI tools. Explore the data: futurerealities.org/poll/2026/ 🧵

Bar chart titled "Positive feeling toward AI varies by experience, not by party," showing the percentage who feel positive about AI overall. By experience using AI: never used AI 12%, tried it a few times 24%, use it weekly 50%, skilled daily user 80%. By political party: Republicans 39%, Democrats 36% (nearly identical).

1164

Reposted by Holden Lee

Juspreet Singh Sandhu @freegaussians.bsky.social · 20/05/2026

Update: We fixed a small bug about the MLSI for our final Glauber sampler; we now use entropy-contraction of the polarized walk instead (since Glauber might not work). More exposition for theorem-5.1 also added. Many thanks to reviewers for a careful reading! 1/2

131

Reposted by Holden Lee

Juspreet Singh Sandhu @freegaussians.bsky.social · 06/05/2026

With @oldheneel.bsky.social and @jnshi.bsky.social (+Ewan), we extend the PHA framework developed to optimize the SK model to become a negligible TVD-error sampler for beta < 1/2. 2 key conceptual observations: PHA implements ASL at high-temperature + Jarzynski's gives a way to rejection sample 1/2

363

Holden Lee @oldheneel.bsky.social · 21/08/2025

Thank you! That was very much on my mind when I was writing it!

010

Holden Lee @oldheneel.bsky.social · 10/07/2025

My first professionally published short story "Beyond the Fold" is in the latest issue #436 of Beneath Ceaseless Skies! Featuring origami magic, lesbian love, and life choices.

2161