Tool author here. Sure, out of all bait in this thread I will bite this one. Curiously, it seems like Konstantinos has opened this ticket, i.e. https://github.com/iczelia/bzip3/issues/177, and open-sourced his algorithm here -- https://codeberg.org/kagiannis/gdcc-2021. As the allegation is very serious, I will also copy the edited version of this response to my website.
I have not seen his tool before, or even was aware of it. However, once you take a closer look, you notice that these two tools could not be more dissimilar. Agiannis' compressor uses a compact context representation to group bytes, followed by RLE and prefix coding. Bzip3 combines repetition removal (here via a run-length code -- prior to the BWT -- and LZP of Charles Bloom), a full Burrows–Wheeler transform, and a richer (thus slower) statistical arithmetic coder. The use of RLE for post-coding the BWT output dates as far back to Julian Seward, perhaps even further. The RLE and LZP are applied before the Burrows-Wheeler transform (as opposed to the implementation in `text', which makes a big difference). Bzip3 uses a proper SAIS library for the forward and backward transforms. Konstantinos' entropy coder seems to use FPC (bytewise prefix codes over adaptively selected subblocks), bzip3 uses an idea similar to this of bcm, which itself descends from Mahoney and ancient work of JS Vitter on arihtmetic coding, where a bitwise arihtmetic coder is input mixed probability estimates from exponential-moving averages with probability refinement.
You are welcome to conduct your own analysis, but this is the gist of it -- perhaps Konstantinos has convinced himself that he had invented run-length coding?
sorry palaiologos. I just looked into the code and it is different enough from my compressor. I just looked the description and it looked similar sorry.
Just tested it with enwik9
agiannis_text enc 0m16,319s dec 0m15,594s size 270365918
The difference is in scaling. The top US labs have oom more compute available than chinese labs. The difference in general tasks is obvious once you use them. It used to be said that open models are ~6mo behind SotA a year go, but with the new RL paradigm, I'd say the gap is growing. With less compute they have to focus on narrow tasks, resort to poor man's distillation and that leads to models that show benchmaxxing behavior.
That being said, this model is MIT licensed, so it's a net benefit regardless of being benchmaxxed or not.
1. electricity costs are at most 25% of inference costs so even if electricity is 3x cheaper in china that would only be a 16% cost reduction.
2. cost is only a singular input into price determination and we really have absolutely zero idea what the margins on inference even are so assuming the current pricing is actually connected to costs is suspect.
I think the US created Starlink for military use. It provides world wide coverage and very small latency that helps a lot with UAVs. UAVs that aren't in line of sight need satellite communication. They just allow the public to use part of it, so as to reduce the cost of the system.
The real big-ticket customers will be hedge funds getting trans- and intercontinental financial data a few ms ahead of competitors.
It would not be surprising if Starlink charged them 100x as much for each ms of latency boost. They would be paying not so much for the ms ahead of fiber as for their lower-paying competitors to get less ahead of fiber.
In fintech, they say "a microsecond is an aeon, a millisecond an eternity". You can do millions of computations in a ms.
The US didn't create Starlink, SpaceX did and they did it without a government grant. They did what they thought would solve the issue of internet connectivity and make a lot of money with it.
Sure they will have many government including military, emergency services, coast guard, NASA and so on.
Starlink has many use-cases for the military they certainty didn't create it with that application as the primary goal. The primary goal is making money to fund SpaceX.
reply