Uzomidy

The BBC was always pretty establishment, but now they're very afraid of seeming “left wing”, and so we get this…

That's what you need to do. Most people think streaks are unlikely, which biases people to void even small streaks. The algorithm expects that, so to fool the algorithm initially, you need to tilt towards streaks, and then when begins to pick up on that, tilt away.

Whee, I did pretty well against it, managed to be at around 30%. https://imgur.com/a/GmaE45k

(And no, I didn't write a Perl script, or have multiple goes to game the score, I just “tried not to do what I thought it'd think I'd do”.)

Interesting critique. I'm curious if there's some other nice-and-simple framework you'd recommend?

The article is a misrepresentation, apparently "believe in a higher power or God" has risen somewhat. This is different from excitement about organized religion, which seems to be less of a thing, not more.

It's not 100% more — 100% would be twice as much. In this case, we've gone from none at all to having some, which is infinitely more.

Does this mean that the Y combinator, Y = λf.(λx.f (x x)) (λx.f (x x)), is not a combinator?

This has nothing to do with cryptographic security or predictability.

This is about whether the output of the PRNG passes statistical tests for randomness. In a previous blog post by John D. Cook we saw that it failed the PractRand statistical test suite [1], and now we see that it also fails the older TestU01 suite, too.

When Xorshift128+ and Xoroshiro128+ were launched, they were claimed to be statistically excellent. Those claims have been shown to be false.

[1] https://www.johndcook.com/blog/2017/08/14/testing-rngs-with-...

It looks like they used an automated tool to add copyright notices to files in the repository. Probably were required to do that by legal. So it's not 100% faithful history, but close.

The benchmark details are in the paper. It's a pretty ordinary Haswell PC.

Looking at the PCG paper, it shows that a basic XorShift generator doesn't pass statistical tests as it gets bigger. Basically, there are some tests it never passes no matter how big it gets.

But a tiny change (XorShift*) is enough to get a generator that passes empirical statistical tests.

The page about other RNGs has lots of good info about this stuff.

The graphs of failures don't show the PCG scheme because their number of failures is zero.

The performance graphs show a whole bunch of PCG generators.

You absolutely can predict the Mersenne Twister after 624 outputs (there are several blog posts out there). The Mersenne Twister is also huge (2K of state) and some people claim that as a result it's not cache friendly.

But yes, it's a new thing, so it needs people to look at it.

Apparently, both the Mersenne Twister and XorShift (which seem to be two of the popular RNGs people use these days) can't pass all statistical tests. (The Mersenne Twister is pretty huge as well.)