Lol, yeah, I do not think you understand the Kangaroo algo.
It's all laid out for you in various readings/papers.
I never said 9 billion kangaroos. Do you understand the algo? When I say "find" x amount of tames and wilds, it is referring to the points/distances found by each type of kangaroo. You store tame and wild points (Based on DP used) and distances, that are generated from the tame and wild kangaroos, hopping all around.
2^66.05 - 2^32 (DP size that I stated) = 2^34.05 stored DPs. 2^33.05 tames and 2^33.05 wilds. 2^33.05 = 8,892,857,981; so roughly 9 billion points and distances stored (tames and wilds) to solve, on average. Could be a little higher, could be a little lower. So no, I was not "kidding". And yes, my times are based on math, and the space complexity is what I said, roughly 9 billion points & distances per tame and wild, to solve. I can't give you exact amount of GBs required because each Kangaroo program stores points differently, different amount of bytes and different formats, binary vs plain text. One would need to calculate it based on their DP and how the points/distances are stored.
But yes, you could set out 2 kangaroos, 1 tame, and 1 wild, and eventually solve, in many many years, or you could get lucky and solve within minutes, hours, days.
I doubt whoever solved 120/125, if they used the kangaroo algo, set a DP of less than 28. They would have an enormous amount of DP overhead, that JLP explains well in his github:
DP overhead according to the range size (N), DP mask size (dpBit) and number of kangaroos running in paralell (nbKangaroo).
110 and 115 were both solved with DP 25. I know that during the 115 run, the grid sizes for the GPUs were choked down and another part of the code was reduced, to prevent a massive DP overhead. And when finally solved, I do believe total DPs stored (points w distances) was a smidge over the expected total of 2^33.55
So your estimated times include the DP overhead?
Because I did a lot of simulations on lower puzzles to get min/avg/max jumps and stored footprints, and, as you can guess, when the DP criteria kicks in, storage goes down, number of jumps until a collision goes up. And there's a lot of ways to improve either space or time, but not both.
So it is not fair to still call the time complexity anywhere near O(sqrt(n)) when you're using 2*sqrt(sqrt(n)) stored items instead of the average 2*sqrt(n).Steps until a collision can range between expected average / 10, and 3 * the expected average.
And JLP kangaroo is not some godly reference, I did not use it at all, for one I don't even agree with the way jumps and hashmap keys are used, but that's another story.
No one says you need to have 50/50 tames and kangs. And no one says you even need to store the entire traveled distance. There are various techniques.
Well, just with HOPE you could try to execute the normal transaction and hope the block will be mined quickly afterward so it'll reach your address before anyone else replaces the TX 
