I guess they've worked out to always get exactly the same text back? Hopefully this won't be a repeat of the photocopiers that used a similar compression technique and occasionally substituted letters in the copy.
For this, I'd look at comparable work in the audio codec space, including Lyra (https://arxiv.org/abs/2102.09660) and SoundStream (https://arxiv.org/abs/2107.03312).
For text compression, I think it would be novel if you could get a lossy but semantically equivalent decompression that was resilient to the exact hardware used for inference. I don't think that's what happened here, though, given the requirement for "exact same GPU and model".
He has also held the ‘world record’ for data compression since 2019 with nncp http://mattmahoney.net/dc/text.html#1085
Robert Metcalfe just won the award for Ethernet. Go scan through the list of Turing winners again. It is not just awarded to pure theorists. There’s a huge slant to academics who also found (or caused) massive commercial success, especially in recent years.
Bellard could get hit by a bus tomorrow and have
>the most important video tool ever
>the most important emulation tool ever
>the best text compression tool ever
…on his resume. And with those tools he’s not just wrapping algorithms cooked up by others, but also innovating on fundamental theory. VirtualBox and the Android emulator? That’s QEMU. Every video platform in existence? Thin wrappers to FFmpeg. Seems deserving to me, and certainly seems as or more impactful that many names on that winners list.
Or is it actly ok.
A summary is a form of lossy text compression, and it's extremely useful.
That's not what this is but if it can actually produce "semantically equivalent" decompressed text it's fascinating.
https://googleprojectzero.blogspot.com/2021/12/a-deep-dive-i...
>JBIG2 doesn't have scripting capabilities, but when combined with a vulnerability, it does have the ability to emulate circuits of arbitrary logic gates operating on arbitrary memory. So why not just use that to build your own computer architecture and script that!? That's exactly what this exploit does. Using over 70,000 segment commands defining logical bit operations, they define a small computer architecture with features such as registers and a full 64-bit adder and comparator which they use to search memory and perform arithmetic operations. It's not as fast as Javascript, but it's fundamentally computationally equivalent.
>The bootstrapping operations for the sandbox escape exploit are written to run on this logic circuit and the whole thing runs in this weird, emulated environment created out of a single decompression pass through a JBIG2 stream. It's pretty incredible, and at the same time, pretty terrifying.
Wow... that's really out of the box. They built their own VM from a JBIG compression stream vuln and wrote an exploit to run on it... Not bad.