Jeremy Kun

Jeremy Kun

@jeremykun.com

Portland-based mathematician and software engineer. Building a homomorphic encryption compiler at Google. https://jeremykun.com/linkgraph https://heir.dev

1.5K followers 425 following1.5K posts

Jeremy Kun reposted

Sophie Huiberts

@sophie.huiberts.me

Too many people are calling on the community to "change how we value papers". Not enough people are posting "this is what we look for in our next hiring round".

Jeremy Kun reposted

Sophie Huiberts

@sophie.huiberts.me

New video is up! Why is the simplex method so fast? At STOC'26 we presented the best theory yet

youtube.comAnalyzing the Simplex Method By-the-BookYouTube video by Sophie Huiberts

I went to a comedy show, and the comedian was doing crowd work, and the crowd member said they worked "in AI" and the theater was overwhelmed by boos. After the booing subsided, the comedian's next line was, "well we've lost the room"

One small benefit of the vulnpocalypse: my wife now thinks it's a great idea for me to install a home server rack for local backups.

Pinterest board for closet design levels of excitement.

In our family we do squares up to 16 and powers of 2 up to 16

Liℤℤy Kodpuak née Acid Lich (can’t spell sheaves without she)@acidlich.bsky.social

What you mean to tell me that you don’t have the first 19 squares memorized

CKKS — Key Switching and Relinearization https://www.jeremykun.com/2026/10/21/ckks-key-switching-and-relinearization/

Truly inspiring (though I could only quickly sketch this one up before a meeting).

Kate@katef.bsky.social

vaguely distressing state machines, a small experiment in emotion with automata and alt text.

If I had nothing else to do today, I'd probably clean up one or two of these state machines and put them on a t-shirt.

Ok this convinced me to try it, just to see where the cameras are near me. The app is very cute tbh. Gotta have more whimsy in my life

MapQuest@mapquest.com

The Roads Have Eyes! 👁️ Turn on the Surveillance Layer in Map Options to see known license plate readers from Flock Safety, Motorola Solutions (Vigilant), Genetec, Axis Communications, Axon, Leonardo (ELSAG), Neology, Rekor, PlateSmart, Ubicquia, Verkada, and other outdoor surveillance cameras.

Indeed, I'm only two appliances away from converting my home/car to fully electric!!

Jeremy Kun@jeremykun.com

I've been wanting to get one but nervous about whether it will fit in the counter space where my current gas stove lives, and how to deal with cutting stone or filling the gap if it doesn't.

My hot take on the OpenAI proof dump is: the papers need to be thoroughly digested before we can understand the scope of the "damage" to math as a profession.

Some papers have been retracted. Most of the papers are still not understood. Some have been found to be a bit less miraculous than advertised.

See for example www.argmin.net/p/faster-fou... As Recht describes it, the FFT "breakthrough" requires a computational model that uses super-polynomially precise complex arithmetic. As such, it can't be reasonably implemented and incurs intense numerical instability.

argmin.netFaster Fourier TransformsThe OpenAI algorithm for the FFT is less a breakthrough and more a malicious exploit.

Jeremy Kun reposted

Sophie Huiberts

@sophie.huiberts.me

Guest blog by me

terrytao.wordpress.comThis is normal bad work[This is a guest post by Sophie Huiberts, being an expanded version of a post on her own blog. This blog post was initially written in a different file format and converted using AI. — T.] Wh…

The existence of Rich Hickey implies the existence of Poor Coverup.

CKKS — Addition and Multiplication https://www.jeremykun.com/2026/10/11/ckks-addition-and-multiplication/

Found an old card that made me laugh

Jeremy Kun reposted

@littmath.bsky.social

extremely concerning that OpenAI has made 3 sign errors—an odd number—upsetting this delicate balance

Daniel Litt@littmath.bsky.social

that the mathematical literature as a whole is so reliable comes down to the fact that, over time, we have made an even number of sign errors

Jeremy Kun reposted

@josefoshea.bsky.social

What a weird world we live in - Ukraine hits a data centre deep in Russia - Irish twitter users see racist posts about IPAS centres in Offaly fall off a cliff

Alina Pereverzieva 🇺🇦@perevaa.bsky.social

And now the best part: Bot traffic from the #Yandex network has plunged to near zero following the strike on its Vladimir data center. Bot traffic includes automated requests from search crawlers, scrapers and other online services. Twitter must be a ghost town by now.

Posting my preliminary proof of P vs NP scribbled on a soiled napkin, attached via cocktail spear to a receipt with the current date and time. Just in case I get scooped by OpenAI.

One possible solution: the community just decides how to ascribe credit appropriately.

Jeremy Kun@jeremykun.com

New fraud dropped: arxiv.org/abs/2610.10072 > The result is also contained in a paper released by OpenAI on October 6, 2026, in which the proof strategy and specific choices of notation are identical to a preliminary version of the present paper that was uploaded to ChatGPT on September 8, 2026.

arxiv.orgInvariant Primes in Lubin-Tate Space and Hovey-Strickland at Every HeightLet $H_n$ be the one-dimensional Honda formal group of height $n$ over $\mathbf F_{p^n}$ and let \[ R_n=W(\mathbf F_{p^n})[[u_1,\ldots,u_{n-1}]],\qquad A_n=R_n/(p). \] We prove, for every height $n$ a...

New fraud dropped: arxiv.org/abs/2610.10072 > The result is also contained in a paper released by OpenAI on October 6, 2026, in which the proof strategy and specific choices of notation are identical to a preliminary version of the present paper that was uploaded to ChatGPT on September 8, 2026.

arxiv.orgInvariant Primes in Lubin-Tate Space and Hovey-Strickland at Every HeightLet $H_n$ be the one-dimensional Honda formal group of height $n$ over $\mathbf F_{p^n}$ and let \[ R_n=W(\mathbf F_{p^n})[[u_1,\ldots,u_{n-1}]],\qquad A_n=R_n/(p). \] We prove, for every height $n$ a...

KEY PHRASE: specific choices of notation Damning.

inb4 "AWS storage bucket was default readable and the AI agent found it" Someone is vibecoding something

A quick update on our homomorphic encryption compiler HEIR (https://heir.dev): our partners at Belfort have a hotword detector in FHE that runs in 350ms. Archived at: https://www.jeremykun.com/shortform/2026-10-08-1120/

View full thread

Belfort also has an early access program for their FHE-based cloud platform which you can sign up for here (http://eml.belfortlabs.com/). You can forgive me for the advertisement, but they intend to use HEIR as a primary tool. So if you use their platform,

you'll be helping find bugs and improvemenets in the HEIR compiler!

A quick update on our homomorphic encryption compiler HEIR (https://heir.dev): our partners at Belfort have a hotword detector in FHE that runs in 350ms. Archived at: https://www.jeremykun.com/shortform/2026-10-08-1120/

View full thread

Of course, this could all be done locally, but it's an impressive showcase of how fast FHE has become. And to my delight, how useful our compiler has become for compiling FHE workloads!

There's still a lot more work to do, but this milestone makes me proud of the community of academics and practitioners we've grown around this compiler. Hooray for collaboration!

A quick update on our homomorphic encryption compiler HEIR (https://heir.dev): our partners at Belfort have a hotword detector in FHE that runs in 350ms. Archived at: https://www.jeremykun.com/shortform/2026-10-08-1120/

View full thread

More specfically, the breakdown of runtime is: 15ms for client-side encryption, 350ms for server-side evaluation, 600ms for network latency (my Portland-based laptop to Belfort's servers which are likely in Belgium) for a total round-trip of less than one second.

The one caveat is that they pre-load the one-time setup encryption keys, which currently take a minute or so to generate in the browser and upload to the server. So if you want to play the game securely, you need that extra step.

A quick update on our homomorphic encryption compiler HEIR (https://heir.dev): our partners at Belfort have a hotword detector in FHE that runs in 350ms. Archived at: https://www.jeremykun.com/shortform/2026-10-08-1120/

View full thread

Your audio is encrypted under FHE, transitted to their server, and a hotword detector runs on it to determine which word you said (still encrypted). The client receives the encrypted result, decrypts it, and updates the game state.

Not only was this demo compiled with HEIR to target Belfort's (proprietary) GPU platform, but the inference step runs in 350ms.

A quick update on our homomorphic encryption compiler HEIR (https://heir.dev): our partners at Belfort have a hotword detector in FHE that runs in 350ms. Archived at: https://www.jeremykun.com/shortform/2026-10-08-1120/

More specifically, they launched a live demo at hotword.belfortlabs.com (https://hotword.belfortlabs.com/), that has a little game inspired by the NY Times Connections puzzle. Instead of typing in your guess for the word that connects a group of clues, you say it out loud.

Belfort did a "NYTimes Connections" game, but where you speak the word and it uses a word detection model server-side, with all computation encrypted under FHE. Wow! hotword.belfortlabs.com Secret words 🟩🟩🟩: all 3 in 27.7 s, guessed out loud and judged under fully homomorphic encryption.

hotword.belfortlabs.comBelfort · Encrypted Keyword GuessingLive encrypted keyword spotting under FHE. The server never sees your audio.

My favorite part is the word detector model was compiled with our HEIR compiler for FHE (heir.dev)

heir.devHEIR: Homomorphic Encryption Intermediate Representation

I reckon I have about two more years while my 6yo still thinks my jokes and bits are funny.

Currently he is really impressed that I can do the Gollum voice, though he has never seen the films.

TIL there's a conference for negative results. CFAIL: The Conference for Failed Approaches and Insightful Losses in Cryptology. Missed it for this year, but I'll definitely keep an eye on it! www.cfail.org/call-for-pap...

cfail.orgCall for Papers | CFAIL

All operating systems are welcome in this household.

Jeremy Kun reposted

🐔 Brian Bucklew 🐔 ₑͤ＞∿＜ₑͤ ∞🌮

@unormal.bsky.social

unfortunately it was unsealed and reading deeper it explicitly differentiates between generative ai and the legal search tool fingfx.thomsonreuters.com/gfx/legaldoc...

Jeremy Kun@jeremykun.com

3rd circuit rules AI training on copyrighted material is NOT fair use. Wow! chatgptiseatingtheworld.substack.com/p/breaking-n...

chatgptiseatingtheworld.substack.comBREAKING News: Third Circuit affirms summary judgment rejection of fair use defense by ROSS Intelligence in AI trainingOpinion under seal for now.

3rd circuit rules AI training on copyrighted material is NOT fair use. Wow! chatgptiseatingtheworld.substack.com/p/breaking-n...

chatgptiseatingtheworld.substack.comBREAKING News: Third Circuit affirms summary judgment rejection of fair use defense by ROSS Intelligence in AI trainingOpinion under seal for now.

Congrats to my colleagues at CMU for releasing their Cerium library: github.com/CMU-CAOS/Cer... It's a multi-GPU FHE (CKKS) implementation that achieved really impressive performance on Llama-8B inference (to the tune of 20 seconds per token). I am excited to integrate it into HEIR!

github.comGitHub - CMU-CAOS/CeriumContribute to CMU-CAOS/Cerium development by creating an account on GitHub.

Oops, I mis-remembered; it's 8s/token for Bert-base, 130s/token for Llama-8B. Still impressive!

It's incredible that "how it works" has become my most triggering phrase.