Tomer Solberg

@hikettei What about rescaling first and then applying the Hadamard rotations?

(6/n) Then I went back and tried to push (1) harder after the rotation. Diagonal Hessian weighting (diag(W^T W)), Fisher-weighted quantization, and low-rank SVD error correction. Basically nothing moved. Once the rotation has done its job, the Hessian is close to uniform across channels, so there isn't much structure left to exploit. SmoothQuant-style per-channel scaling also seems to work against the rotation: after rotating, the scaling gets smeared back across channels and the smoothing mostly disappears. So the same step that removes the outliers also removes the signal that most follow-up methods need. Does anyone have ideas on what could still help at this point? :)

(1/n) I recently joined @trymirai, where we are working on LLM inference targeting Apple Silicon. Lately I've been digging into quantization. LLM inference is mostly memory-bound. The byte/FLOP ratio is high enough that a lot of the machine's time goes to moving data around instead of doing compute. Quantization helps with that in general, but on Apple Silicon there's an extra payoff: the GPU has a fast W8A8 path. If both weights and activations are INT8, you can use that path for prefill and speculative-decoding verification. Weights are easy since they're static and can be quantized offline. Activations are where the real pain starts.

CICLE V3.5: Sumcheck with Lambda Functions by @Ingo_zk @ingonyama/icicle-v3-5-sumcheck-with-lambda-functions-5f17d88910cb" target="_blank" rel="nofollow noopener">medium.com/@ingonyama/ici…

@RenaudDUBOIS10 @OmerShlomovits The proof is only required if the ownership is disputed. Not every image needs a proof

@TomerSolberg @OmerShlomovits but the validity/zk proof, itself ? Where is it written ? The aim here seems to be able to verify publicly the watermark no ?

We unlocked a new ZKP use case: Meta uses the Stable Signature algorithm to watermark their genAI images (Imagine). This watermark is robust to image manipulation but the problem is that the only one who can detect it is Meta. Anyone else who knows the watermark extractor can reverse the process and remove Meta watermarks from all of their images. The solution is ZK: Meta can attach to every image a watermark + proof that a watermark can be extracted from this image. Clients can now easily verify if an image is watermarked or not. This particular ZKP requires proving a CNN, and doing it fast as Meta AI generates many images daily. Our solution is based on ideas from previous works zkDL and zkCNN, runs on GPU end to end, and with some optimizations takes only a few seconds to run.

@sergey_nog @OmerShlomovits That's a good question. I think the number of samples you'll need increases exponentially with the bit length of the WM but It's definitely something to look into

@OmerShlomovits what are the guarantees that the watermark extractor cannot be reversed engineered after sufficiently many samples?

@RenaudDUBOIS10 @OmerShlomovits Meta's stable signature model discusses the robustness of the watermark. You'll have to distort the image by quite a lot to remove it as long as the decoder/extractor networks remain private

@OmerShlomovits What is the exact model then ? If one can erase the watermark bc the process only lie on the extractor knowledge, it can just publish the image without proof. If it is authenticity of image+proof, then use classic signature.

This is T2 tower field T2~ F2[x,y]/(x^2+x+1,y^2+xy+1) Represent each number as a multilinear polynomial in x,y where the coefficients are the binary representation of the number. So 1~1, 2~x, 4~y, 8~xy This gives 6~x+y, and 2*6~x*(x+y)=x^2+xy=1+x+xy~1+2+8=11

@prasincs @kobigurk Actually no. Since this is a field it can't have 0 divisors so 2*6 is definitely not 0. Usually in GF(2^4) 2*6=12 (just look up galois field calculator), but I guess tower field representation is different than the standard one so it could be 11

@kobigurk Wouldn’t 2*6 also be 0 then instead of 11? (2+2+2+2+2+2 =0 because 2 is additive inverse of itself)

Guys this is just a field with characteristic 2

We have successfully closed $21M in seed funding 🥳 The round was led by @IOSGVC, @walden_catalyst and @__geometry__ This is an important step on our mission to democratize #ZKPs by enhancing their speed, scalability, and cost-effectiveness, making them accessible to all 1/3

@Antalpha_Labs Nice to hear that you found my write up useful! I considered adding Lasso, but I'm currently working on other projects and couldn't find the time

@TomerSolberg consider adding lasso? 🫡

Salute 🫡 to @Ingo_zk Brief history of lookup arguments, A not so brief history of lookup arguments by @1dot2 @SECBIT_IO Will be sharing some notes later while we still taking time to digest 🤯 @ingonyama/a-brief-history-of-lookup-arguments-a4eeeeca2749" target="_blank" rel="nofollow noopener">medium.com/@ingonyama/a-b…

more great educational content, by @TomerSolberg and @godstar28 👑 youtube.com/watch?v=FfeXX6…

@Ingo_zk’s @TomerSolberg breaks down Fast Danksharding.

A Brief History of Lookup Arguments - A new review paper! 1) uses of lookup protocols 2) plookup 3) cq, LIKE YOU HAVE NEVER SEEN BEFORE 4) evolution from plookup to cq github.com/ingonyama-zk/p…