Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

TurboQuant: Redefining AI efficiency with extreme compression

research.google

Post reply on HN
Print view View original

161–170 of 202 posts

Re: TurboQuant: Redefining AI efficiency with extreme compression

#161

Post by jmalicki »

amitport wrote: post #98
jmalicki wrote: post #75

Earlier quoted context omitted.

If they didn't cite your paper that's bullshit. But if they read your paper enough that they invited you to a talk, that probably means they were far enough along to independently inventing it they were going to do so anyway, and wanted to chat with someone who was also doing the thing they were already doing. Good ideas tend to reveal themselves to anyone who is aware of the problem.

To be clear, I am not claiming they stole an idea. They have made significant independent research. However, a specific part regarding the treatment of rotation with bias correction relates to prior work, and it would be appropriate to have that recognized.

If they didn't at least cite it, it is complete bullshit.

If they cited it, but you feel you deserved more credit than that... I feel you, but it's less clear cut.

Re: TurboQuant: Redefining AI efficiency with extreme compression

#162

Post by rtrgrd »

photon_lines wrote: post #100
gavinray wrote: post #80

Can someone ELI5 these two concepts please, which make no sense to me: > "TurboQuant starts by randomly rotating the data vectors. This clever step simplifies the data's geometry" I don't understand how taking a series of data and applying a random rotation could mathemetically lead every time to "simpler" geometry. If I throw a bunch of shapes on the ground, tightly packed and touching each other, then rotate all of…

The whole goal of quantisation is to put the data into 'bins' so that it can easily be 'packed' so that you can represent it using less bits (less information). You can think of it like rounding essentially (3.14159 -> 3). Now, sometimes within data, the distribution will be non-ideal for separating it out into bins (let's say that our rounding rules are simple -- we simply use a floor function so 2.45 maps to 2 and…

Added to my non-llm username list :)

Thanks so much for the explanation

Re: TurboQuant: Redefining AI efficiency with extreme compression

#163

Post by Geee »

mesuvash wrote: post #158
Geee wrote: post #139

Earlier quoted context omitted.

Is there an error in the visualization? It shows that every vector is rotated the same amount. My understanding was that they are randomized with different values, which results in a predictable distribution, which is easier to quantize.

That's actually correct and intentional. TurboQuant applies the same rotation matrix to every vector. The key insight is that any unit vector, when multiplied by a random orthogonal matrix, produces coordinates with a known distribution (Beta/arcsine in 2D, near-Gaussian in high-d). The randomness is in the matrix itself (generated once from a seed), not per-vector. Since the distribution is the same regardless of th…

Thanks. However, from this visualization it's not clear how the random rotation is beneficial. I guess it makes more sense on higher dimensional vectors.

Re: TurboQuant: Redefining AI efficiency with extreme compression

#164

Post by fc417fc802 »

mesuvash wrote: post #144

TurboQuant explained with an easy to understand (no-math) animation https://mesuvash.github.io/blog/2026/turboquant-interactive/

Someone else linked that elsewhere in the comments and while it's certainly a nice visual it seems like it's not accurately portraying the paper. Isn't the grid supposed to have a weird alignment that depends on the bit depth? And there's supposed to be a second quantization step involving the residual.

Re: TurboQuant: Redefining AI efficiency with extreme compression

#165

Post by elif »

gavinray wrote: post #80

Can someone ELI5 these two concepts please, which make no sense to me: > "TurboQuant starts by randomly rotating the data vectors. This clever step simplifies the data's geometry" I don't understand how taking a series of data and applying a random rotation could mathemetically lead every time to "simpler" geometry. If I throw a bunch of shapes on the ground, tightly packed and touching each other, then rotate all of…

i could be mistaken but from my read, the 'rotation' aspect is nothing new and not dissimilar from normal spin quant, where the importance matrix is rotated during calibration such that the local minima/maxima are more evenly smoothed and excessive/redundant quantization of parameters is avoided.

as for the J-L transformation is way above my head so i'm almost certainly mistaken but it seems to be some clever way to use a bit as a sort of pointer in order to reuse existing chunks of parameter weight data like in a jpeg or zip compression algorithm.

Re: TurboQuant: Redefining AI efficiency with extreme compression

#166

Post by mesuvash »

Geee wrote: post #163
mesuvash wrote: post #158

Earlier quoted context omitted.

That's actually correct and intentional. TurboQuant applies the same rotation matrix to every vector. The key insight is that any unit vector, when multiplied by a random orthogonal matrix, produces coordinates with a known distribution (Beta/arcsine in 2D, near-Gaussian in high-d). The randomness is in the matrix itself (generated once from a seed), not per-vector. Since the distribution is the same regardless of th…

Thanks. However, from this visualization it's not clear how the random rotation is beneficial. I guess it makes more sense on higher dimensional vectors.

Yes, this is important in high dimension. But sadly, very hard to visualize. In 2d it looks like unnecessary.

Re: TurboQuant: Redefining AI efficiency with extreme compression

#168

Post by mesuvash »

fc417fc802 wrote: post #164
mesuvash wrote: post #144

TurboQuant explained with an easy to understand (no-math) animation https://mesuvash.github.io/blog/2026/turboquant-interactive/

Someone else linked that elsewhere in the comments and while it's certainly a nice visual it seems like it's not accurately portraying the paper. Isn't the grid supposed to have a weird alignment that depends on the bit depth? And there's supposed to be a second quantization step involving the residual.

Fair point. I've updated the animation to address this. The grid now uses the correct non-uniform centroids (optimal for the arcsine distribution in 2D), so you'll see grid lines cluster near the edges where unit-circle coordinates actually concentrate, rather than being evenly spaced. The spacing does change with bit depth.

On the second quantization step: the paper's inner-product variant uses (b-1) bits for the MSE quantizer shown here, then applies a 1-bit QJL (Quantized Johnson-Lindenstrauss) encoding of the residual to make dot-product estimates unbiased. I chose to omit QJL from the animation to keep it digestible as a visual, but I've added a note calling this out explicitly.

Re: TurboQuant: Redefining AI efficiency with extreme compression

#170

Post by moktonar »

amitport wrote: post #21
mungoman2 wrote: post #19

Earlier quoted context omitted.

What they're saying is that the error for a vector increases with r, which is true. Trivially, with r=0, the error is 0, regardless of how heavily the direction is quantized. Larger r means larger absolute error in the reconstructed vector.

Yes, the important part is that the normalized error does not increase with the dimension of the vector (which does happen when using biased quantizers) It is expected that bigger vectors have proportionally bigger error, nothing can be done by the quantizer about that.

Except maybe storing another smaller vector for the difference with the original data an also quantize that maybe recursively
Post reply on HN