HN Explorer

5 comments

[ 4.2 ms ] story [ 26.2 ms ] thread
jabowery
Saurabh Kumar has submitted "Fast Cmix" to The Hutter Prize for Lossless Compression of Human Knowledge.

The Judging Committee appreciates Saurabh Kumar's compliance not only with the rules of the contest but with the provision of an executable archive which makes our job easier.

As per the Hutter Prize Award's section of The Rules:

The contribution is subject to public comments for a period of at least 30 days before the prize is awarded.

gwern
The writeup really emphasizes how far the Hutter Prize has fallen short of its original goals of contributing to AI progress. The innovations in making a smarter compressor here and beating decades of work to create an even more intelligent AI predictor here are... 'eliminating all the virtual function calls' and 'improving cache locality' and 'alternate hash-map implementations'.
jabowery
The failure is that of the machine learning world to underwrite the prize.

There is no greater potential risk adjusted return to the field.

$100/month paid out over the lifetime of the prize must be compared against the investments in machine learning over that time.

gwern
I think it got plenty of investment in the sense of zero relevant work despite many man-years. Even low-value Arxiv fodder has better payoff than that (because it's not 0). And it's obvious why it doesn't work, and never will work: the resource requirements are so stringent that they are off by many orders of magnitude on every dimension, from data to flops to RAM. Hence, from the start, the Hutter Prize has incentivized pretty much only bit-twiddling demo-scene-style work.
jabowery
If you think that $100/month incentivises "plenty of investment" compared to the tens of billions put into less principled "SOTA" measures, I suppose that says all that needs to be said about your credibility.