What you are seeing here is probably the effect of window size. BZip has to perform the BWT strictly block-wise and is quite memory-hungry, so `bzip2 -9` uses a window size of 900KB, if I recall correctly. Dictionary-based algorithms are more flexible in this regard, and can gain a substantial advantage on very large and repetitive files. The article kind of forgets to mention this. Not that BZip isn't remarkably efficient for its simplicity, but it's not without limitations.
> Every time you look up something related to Radon, it's always cited as "the second leading cause of lung cancer after smoking"
> I wonder if that's really true.
That claim is in fact based on extremely poor research methodology. It is made by combining the linear no-threshold model of radiation damage (which contradicts everything we know about cellular repair and hormesis) with evidence from "case-control studies", a kind of retroactive hand-waving that has nothing whatsoever to do with a "controlled trial", despite the name.
Maybe a bit late to comment, but for what it's worth: There is no primer that could be actually useful, because the "tax optimization" landscape is fragmented and constantly shifting. Everything depends on where you live, where you do business, how much much money is involved, etc.
But there is a central driving force behind it all: governments constantly fight for "tax justice" with one hand and create various "incentives" and exceptions with the other, in an effort to briefly gain the upper hand over other countries in the zero-sum game of attracting international capital. The former tends to plug all possible loopholes for the "ordinary wealthy", while the latter always leaves options for the truly big fish, they just don't stay the same decade-over-decade.
HN seems to be full of Anthropic fanboys for some reason. Probably because Dario is the only big boss in AI right now that successfully pulls off the I'm not a sociopath act.
Probably more like Claude was slightly better than GPT-xx when the IDE integrations first got widely adopted (and this was also the time where there was another scandal about Altman/OpenAI on the front page of HN every other week) so most programmers preferred Claude, then it got into a virtuous cycle where Claude got the most coding-related user queries and became the better coding model among SOTA models, which resulted in the current situation today.
At the risk of stating the blatantly obvious, this will be a rebranded Chinese phone, if it happens at all. The photo on their website [1] is a quick and sloppy Photoshop job (note identical lenses and lack of flash), and the specs and pricing are totally implausible for a US-made phone. Compare that to the 2000$ Purism charges for their comically under-powered Liberty phone [2] that is mostly US-made.
It's legalized graft, not a subsidy. The tens of billions flowing into SLS do not bolster productive capability in the civilian or military aviation sector, they tie up engineers in a nonsensical, dead-end project, and totally mess up incentives on top of that.
I think you are misusing Berkson's Paradox here. It applies when you sample two extremes, i.e. when you look at the richest 0.1% and the most moral 0.1% and notice that the two appear mutually exclusive, even though they might actually be uncorrelated in the general population. When you look only at the richest 0.1% and notice their lack of morals compared to the general population, that is a legit correlation.
Elsewhere in the world (under IFRS accounting rules) capitalization of R&D costs has been a firm requirement for a while. The US has been somewhat unique in allowing them to be expensed instead, until recently.
Yeah, seems I was wrong about that. Apparently most IFRS countries allow expensing R&D for tax purposes, regardless of accounting. Many even have an R&D superdeduction nowadays.
If the business has some revenue, but is not yet profitable after deducting development costs, it can become profitable on paper (and owe tax) if R&D is capitalized instead.
reply