Jeremy Kun

Author of Math ∩ Programming currently doing homomorphic encryption @ Google. Based in Portland, OR.

Linkgraph (Books, newsletter, etc): https://jeremykun.com/linkgraph
PIM Book: https://pimbook.org
PMFP Book: https://pmfpbook.org

Last Notes

2026-04-02 13:19:14 UTC

I think the real question about GH availability metrics is: does having zero nines actually impact their business? Or are they slacking on reliability because they can and their users won't leave.

2025-11-22 21:57:40 UTC

- reply

many companies have internal-only LLMs trained on proprietary source code for their employees to use. The obvious answer on why not to make these public is security and IP protection. This seems obvious....

2025-07-09 20:00:20 UTC

I found one of those 4-digit number clickers and showed it to my toddler. He has been playing with it and using it to count people who walk past our house.
We use it to talk about place value, the idea of offloading our memory to a physical device (I keep pretending I forget the current number and he is delighted to remind me by showing me the clicker), and he learned about "overflow", which he describes as "when the number gets so big that the computer gets tired." I told him that it's like if you filled up a cup too much and it overflowed, but then instead of staying full the cup suddenly became empty.
My son also asked: how does this computer work without any energy? (electricity) And I realized that yes, this is a computer.
https://media.mathstodon.xyz/media_attachments/files/114/825/036/893/578/126/original/b75c481a6099e75c.png

2025-06-10 23:28:29 UTC

- reply

+1 why bother repeating a known bullshitter?

2025-01-08 02:05:12 UTC

I'll be at the Joint Mathematics Meeting in Seattle (starting tomorrow).
Archived at: https://www.jeremykun.com/shortform/2025-01-07-1747/

2025-01-04 19:25:34 UTC

- reply

ah, definitely NOT common 😔 but not completely unused

2025-01-04 03:50:36 UTC

- reply

Correct, they can't get away with pure FHE because it would be too slow to do a nearest neighbor search across all images. Sharding means that Apple sends each device all the cluster centroids, and the client requests a search within a particular cluster by including the centroid. The centroid would be public, and leak some info about what the client wants to identify, but the anonymization network + DP is supposed to alleviate this.
In my side of the world, we're not betting on hybrid protocols, and instead hardware acceleration, but basically all production FHE applications today are hybrid protocols like this. I wonder if the same author would feel the same way if they saw the other places this same technology was used.

2025-01-04 03:47:00 UTC

- reply

So at worst they can't see your particular image of the Eiffel Tower (that would require cracking the crypto), but they might know you're asking to identify a building landmark because the centroid that leaked was the centroid corresponding to "building landmarks."
Finally, I do love the line, "Tell me honestly that you understand that! How many people in the world understand it?" because I can say that I do :) I read the papers, though not in so much detail as to give a serious security review, but enough to understand how it is intended to work.

2025-01-04 03:46:46 UTC

- reply

I think the author brings up some decent points about the rollout, and I am generally in favor of opt-in features whenever possible. More so when the feature is based on new/novel research. It's also reasonable for someone who had never heard of homomorphic encryption to live by the credo, "anything that phones home is not private," because this HE stuff is really fantastical. Finally, his point about bugs is good--any additional surface area is a risk no matter how awesome the math, and while that's an overly broad point, it seems particularly relevant to this feature. It's good to be suspicious, that's the lifeblood of the security industry.
On the other hand, it's a rich juxtaposition to say "the only way to guarantee privacy is to not send data off the device" alongside "I don't understand what homomorphic encryption is." Unless he's trying to make a point about the possibility of cracking the crypto, which is valid but not something I often hear, then just knowing that only _encrypted_ data leaves the device should suffice for his purposes to guarantee privacy. And that is true of Apple's implementation (provided the code matches their papers), with the exception of the centroid of the cluster being requested, and they use this anonymization network technique (the part that smells riskiest to me) to hide which centroid the device is querying (Apple provides the centroids to the device).

2024-11-22 16:19:57 UTC

There's a bulletin board at my local bagel shop, on it is an advertisement for ABACUS LESSONS and every single phone number slip has been taken.

2024-11-12 18:58:24 UTC

My god, someone implemented tic tac toe in a single (looped) call to printf: https://github.com/carlini/printf-tac-toe

2024-09-23 03:03:51 UTC

- reply

seems like you could get big wins for cache locality by doing this well.

2024-09-05 23:06:22 UTC

- reply

I'm not staying super informed about what is going on in the blockchain world here, but there is a lot of funding to do FHE and ZKP on ethereum.
I wrote jeremykun.com/fhe-in-production in case you want to see what is actually being used.

2024-09-05 21:41:28 UTC

As my FHE career has progressed I find myself getting a LOT of attention from blockchain folks who want to do FHE on the ethereum blockchain (as well as zero-knowledge practitioners funded by various blockchain startups).
I'm not a huge fan of this development. Doing my best to keep my distance from it, and see what I can do about getting non-blockchain FHE work in production ASAP :)

2024-09-02 15:12:53 UTC

- reply

there are native options, but they require a lot more work for the author to write the equations. https://developer.mozilla.org/en-US/docs/Web/MathML

2024-09-02 14:24:14 UTC

- reply

because today supporting latex rendering requires the server to send additional data to the client, which adds to bandwidth costs for the instance maintainer. Browsers/phones support Unicode (hence all languages) natively, which makes foreign language support free.

2024-09-02 05:57:34 UTC

- reply

seems unlikely to force other instances to support it, more likely to happen via a browser extension