[dead]
Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
41–50 of 55 posts
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#42[flagged]
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#43Earlier quoted context omitted.
So then ... "lossy"
theres a big difference between 99% quality and 30%. near lossless is a good name for the first one. if you treat it in a binary way where everything short of 100 falls into one "lossy" bucket you lose all the practical differences that make one encoding much better than another.
> theres a big difference between 99% quality and 30%.
sure
> if you treat it in a binary way where everything short of 100 falls into one "lossy" bucket you lose all the practical differences that make one encoding much better than another.
no; lossless is an inherently binary term. and I don't lose all the practical differences of better lossy encoders by understanding that; I'm not just going to start using mp3 96k because I have an understanding of lossless vs lossy encoders...
Lossless is an objectively binary term.
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#44Hey breadislove; amazing article, I’ll be sending mixedbread an email in the morning that may interest you (email will be @pm.me) I have also been working in compression and performance engineering, and managed to get a 99+% compression unlock versus conventional approaches (100+KB down to 1KB) in the scenario of 30 minute massive multiplayer game replays for a “game+engine” I’m developing I think there’s a synergy b…
Any way I can read about this or the use case? I have a hobby interest
Yes soon I’ll be launching my game and engine, and will have a blog post - just keep an eye on Show HN over the following week
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#45Hey breadislove; amazing article, I’ll be sending mixedbread an email in the morning that may interest you (email will be @pm.me) I have also been working in compression and performance engineering, and managed to get a 99+% compression unlock versus conventional approaches (100+KB down to 1KB) in the scenario of 30 minute massive multiplayer game replays for a “game+engine” I’m developing I think there’s a synergy b…
to which email did you send it? can u send it to support please?
Sent to the support email with the subject line “Hackernews …”
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#46Earlier quoted context omitted.
So then ... "lossy"
theres a big difference between 99% quality and 30%. near lossless is a good name for the first one. if you treat it in a binary way where everything short of 100 falls into one "lossy" bucket you lose all the practical differences that make one encoding much better than another.
I agree with you somewhat, and I like what is described in the article. But I also feel like we are diluting the meaning of the word to make things sound better. Lossy/Lossless is inherently binary, and it carries a specific meaning. It would not detract from the work at all if it was described differently.
You can't be a little bit on fire :)
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#47``` We evaluated several precision pairings across our internal retrieval benchmark suite. Scores are NDCG@10 averaged across the suite, scaled to 0–100. NDCG@10 (Normalized Discounted Cumulative Gain at rank 10) measures how well the top 10 results are ordered against the ideal ranking, rewarding relevant documents more when they appear higher, with 100 being a perfect ranking. The full-precision baseline averages 9…
Yeah, what bugs me about stuff like that is like they spend all this time and then they output several or minimal real testing to prove the theory It's like you're building your model to And just because it takes a long time to compute and do the testing, you'd rather publish your article and then try to get credit on something that hasn't really been proven. Look, prove your results. Study it. Ruggedize it. Make sure it works. Then, show us.
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#48Earlier quoted context omitted.
theres a big difference between 99% quality and 30%. near lossless is a good name for the first one. if you treat it in a binary way where everything short of 100 falls into one "lossy" bucket you lose all the practical differences that make one encoding much better than another.
> theres a big difference between 99% quality and 30%. sure > if you treat it in a binary way where everything short of 100 falls into one "lossy" bucket you lose all the practical differences that make one encoding much better than another. no; lossless is an inherently binary term. and I don't lose all the practical differences of better lossy encoders by understanding that; I'm not just going to start using mp3 96…
Lossless is objective for information theory. To get from the real world to digital world you need an analog to digital converter, this process is by definition lossy. We are interested in the real world, and information is pure but never represents exactly reality.
Lossyness is baked into our problem statement here.
Using terms like near lossless means we think we are very close to reality for what we’re trying to do
Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#49Re: Asymmetric Quantization: Near-Lossless Retrieval with 97% Storage Reduction
#50``` We evaluated several precision pairings across our internal retrieval benchmark suite. Scores are NDCG@10 averaged across the suite, scaled to 0–100. NDCG@10 (Normalized Discounted Cumulative Gain at rank 10) measures how well the top 10 results are ordered against the ideal ranking, rewarding relevant documents more when they appear higher, with 100 being a perfect ranking. The full-precision baseline averages 9…
Ask a SOTA LLM when Newton was born without any access to internet : the answer is Lossless for our shared culture understanding of this question. Not Near-lossless, lossless. Ask the same LLM when YOU were born, the answer is just wrong for almost anyone in the world, not lossy. Between the two there is a whole new field of Lossyness to study.
90% depends entirely on what the measure means here, do you understand what "Normalized Discounted Cumulative Gain at rank 10" means to the set of data that we are comparing ?
Sometimes coming up with new codecs (compressors decompressors) means coming up with new ways to interpret artifacts of the real world. And this is exactly why LLM are so powerful and they are like a giant Lossy (but Near-Lossless for various use cases) ZIP file / Database of the whole knowledge of the training data.
Nobody is trying to manipulate you here, humanity just has to find new explanations for complex topics.
Lossy-ness is binary
Lossless is objective for information theory. To get from the real world to digital world you need an analog to digital converter, this process is by definition lossy. We are interested in the real world, and information is pure but never represents exactly reality. Lossyness is baked into our problem statement here.
Using terms like near lossless means we think we are very close to reality for what we’re trying to do