kvezach writes: To put it differently: suppose that the text was 10 bytes long.

A better way of thinking about data scaling is to ask "How many digits of Pi would it take before, say, John Tromp's 401-bit binary lambda algorithm that generates Pi would become a better model than the literal string of those apparently-random digits?" (And by "better" I mean not only that it would be shorter than those digits, but that it would extrapolate to (ie: "predict") the next digit of Pi.)

In terms of how much data humans require, this is, as I said, something about which everyone has an opinion (including obviously you, to which you are of course entitled) but on which there is no settled science: Hence the legitimacy of the 1GB limit on a wide range of human knowledge for research purposes.

Concerns about the bias implied by "The Hardware Lottery" are not particularly relevant for engineering/business decisions, but path dependencies implicit in the economics of the world are always suspect as biasing research directions away from more viable models and, in the present instance, meta-models.

There is only one Hutter Prize contest and it's for 1GB. 100MB was the original size for the Hutter Prize starting in 2006, but it was increased to 1GB in 2020, along with a factor of 10 increase in the payout per incremental improvement. See the "Hutter Prize History".

Insofar as the size is concerned: The purpose of the Hutter Prize is research into radically better means of automated data-driven model creation, not biased by what Sara Hooker has called "The Hardware Lottery". One of the primary limitations on current machine learning techniques is their data efficiency is low compared to that which natural intelligence is speculated to attain by some theories. Everyone has their opinion, of course, but it is far from "settled science". In particular, use of ReLU activation seems to indicate machine learning currently relies heavily on piece-wise linear interpolation in construction of its world model from language. Any attempt to model causality has to identify system dynamics (including cognitive dynamics) to extrapolate to future observations (ie: predictions) from past observations (ie: "the data in evidence"). Although there is reason to believe Transformers can do something like dynamics within their context windows despite using ReLU (and that this is what gives them their true potential for "emergence at scale") it wasn't until people started going to State Space Models that they started returning to dynamical systems identification (under another name, as academics are wont to gratuitously impose on their fields).

Probably the best introduction would be Hutter's 2010 paper titled "A Complete Theory of Everything (Will Be Subjective)". From there you can follow its citations in google scholar.

You didn't submit an executable archive of enwik9 purported to expand into a file that matched bit for bit. While you also failed in some other ways to qualify, that is now the first bar you must clear, before any further investment by the judging committee.

Or are you Yann LeCun out to pull my leg again?

https://twitter.com/ylecun/sta...

One big mistake they made early on with the Hutter Prize was not insisting that all contestants make their entries Open Source.

IIRC, only one entry was closed source. You may be thinking of Matt Mahoney's Large Text Compression Benchmark where the top contender is frequently closed source.

That the machine learning world has yet to recognize lossless compression as the most principled loss function is a tragedy, but it is due to a lot more than that entry. This failure stretches back to when Solomonoff's proof was overshadowed by Poppers falsification dogma in his popularization of the philosophy of science:

When a model's prediction is wrong, under Popper's falsification dogma, the model is "falsified", whereas under Solomonoff, the model is penalized by not only a measurement of the error (such as LSE), but by the literal encoding the error within the context of the model. The significance of this subtle difference is hard for people to understand, and this lack of understanding derailed the principled application of Moore's Law to science. Instead we got an explosion of statistical "information criteria for model selection", all of which are less principled than the Algorithmic Information Criterion, and now we have ChatGPT hallucinating us into genuinely treacherous territory.

Lossy compression requires a utility function. That resides in the decision theoretic framework (engineering) not the prediction framework (science). The Hutter Prize is about the latter (Algorithmic Information Theory).

Disparaging the terms "master" and "slave" is hurtful to members of the D&S community. Is this the new idea of "inclusion"?

retchdog writes:

If BERT can magically sneak in some spooky black magic, then you're admitting that enwik9 is not an adequate data set for testing, end of story.

No. All BERT must do is beat the Large Text Compression Benchmark -- a benchmark that uses enwik9 as the corpus.

Megol asks:

But let's assume the information is important, why should the raw text be repeated instead of the myriad variants that express the same thing in a different way?

For the same reason scientists don't throw out the raw measurements in their experiments just because it departs from theory. One can assert that the form of the expression is irrelevant to the underlying knowledge, but this is a sub-theory based the agent's comprehension of the entire corpus. This becomes particularly relevant when attempting to assign latent identities to sources of information so as to model their bias.

Separating the 2 questions, 1) "Why not just use BERT?" and 2) "Why exclude parallel hardware?"

1) BERT's algorithmic opacity leaves open an enormous "argument surface" regarding "algorithmic bias". There are >100,000 hits for "algorithmic bias" due to inadequately principled model selection criteria used by the ML community, thence google's algorithms. Algorithmic Information Theory cuts down the argument surface to its mathematical definition of natural science:
  Solomonoff Induction. Basically, it is provable that if classical (Cartesian) causality holds, distilling a dataset of observations to its Kolmogorov Complexity means you've factored out all the bias you possibly can if you claim to be "data-driven" in your policies. The BERT folks have yet to present a compressed enwik8 (100M) or enwik9 (1G) to demonstrate it is a superior unbiased model. They should do so. They don't have to submit an entry to the Hutter Prize for this, so they are free to use Google's entire TPU farm, or whatever. Let this be a challenge to them if they are serious about dealing with accusations of bias in their algorithms.

2) If one reads the relevant Hutter Prize FAQ answer, it becomes apparent that even a 2080Ti (or other GPU/TPU) may not be affordable to some of the best minds on the planet.

Lossy Compression Is Confirmation Bias.

One Man's Noise Is Another Man's Ciphertext.