anon_tor_12345

[flagged]

i'm gonna get downvoted for elitism but it's very clearly "aspirational upvotes" from people that don't actually know any ML and won't ever end up learning any. to wit: no one that actually studies seriously spends this much time wringing their hands over which reference to use.

you don't keep cycling through references that seem better because that's a sure way to never make any progress - you pick a reference and grind through it. maybe with occasional double checking against some other reference sure but you never end up ditching the first one because the cost of changing notations/formalisms/etc is very very high and not worth paying almost ever (since intrinsically they're all talking about the same thing anyway). this is the reason that profs today still teach from the books they learned from 30 years ago.

you don't need vivado to modify the hdl, you need vivado (or some other proprietary xilinx toolchain) to produce the bitstream which actually deploys to the fpga.

i read polya's book as an undergrad to my detriment. 10 years later and having "solved" several problems, i can authoritatively say that it is not how difficult problems are solved. the real "how to solve it" for problems that aren't exercises is closer to what feynman said in one of his books: a good physicist [or mathematician] has 5-10 problems on their mind at all times and when they learn of a new technique they apply it to their problem. such learning happens by reading monographs, papers, and going to conferences. that is to say that math (or theoretical physics) is socially produced/developed.

there is some slight difference but i don't remember what it is. i wrote a blog post about abel-ruffini based on those two videos and i remember there being some things that were clearer in one rather than the other (but i don't remember which!)

is this the first instance of an article title that ends with a question mark where the answer is in fact yes?

>with people he didn't consider mentally slow.

your whole point is that no such people existed?

>You sure about that? Society has MASSIVELY benefited from von Neumann's work. If the cost of that was a few people's hurt feelings at his inability to interact with them a way that doesn't hurt their feelings it was a small price to pay.

The flawed premise implicit in this is that he (or even someone else) wouldn't have produced all the same things while being cordial. More importantly the even greater flaw is the assumption that he wouldn't have produced even more if he'd been easier to work with

Lol this is exactly what I preemptively alluded to - how we (as a culture) enable this kind of behavior. It becomes even worse when you realize that at least the ultra wealthy pay people for the right to abuse them and von Neumann et al are abusing, frequently, poorly paid junior scientists.

>That gets old quickly no matter what level you're operating on.

If you can't handle being around people that are differently abled from you then the answer is not to take your frustrations out on them. The answer is to stop being around people. No community/culture/society owes anyone, not even people at this level, some kind of pampered sphere of existence that revolves around them.

lol people don't understand that science at gerard 't hooft's level is just as petty and vulgar as your local bar.

what i mean is very accomplished scientists brag and sneer and bluster all the time - they feel entitled to and we worship and empower them to. examples abound; read any of feynman's books; landau had a scale by which he measured other scientists; von neumann would chew people out who couldn't keep up with him; the cat fight right now around Mochizuki's proof; etc etc etc. if you've never been around these people you think they're saints when in fact they're almost universally assholes.

it's not unlike ultra-wealthy people competing with each other for the greatest monument to their wealth...

i'm sure someone will respond to me to say something like they need to be this way to accomplish (just like with the slavish worship of the ultra wealthy) what they've accomplished and to this i always present to them john bardeen, who won two nobels and still managed to be a good neighbor.

E.g. Pavan lol

Without reading the paper, I think you have it a little backwards - the IR doesn't itself allow for more general functions. More general functions are possible (in theory) because the frontend (this Triton language) is decoupled from the backend (CUDA) through the IR as an interface. In this way the Triton IR is no less domain specific than XLA (because both are IRs that represent sequences of operators that run on GPU (or TPU or whatever). I guess in theory Triton could be eschewing all of eg cuDNN but most likely it's not as NVIDIA's closed source kernels perform best on their closed source hardware.

Edit: should've read the post before commenting. Looks like they are in fact using LLVM's PTX backend (ie generating cuda kernels from scratch). Kudos to them