The top-end cards became halo products. Just like few cars on the road are Mustangs, few video cards are 1080 Tis.
Checking the Steam Hardware Survey, about 60% of all cards are Nvidia GeForces in the xx50 to xx70 range - the normal, reasonably-priced cards. (AMD is only 10% of the market right now, with much of that being integrated. Their fundamental architecture was just way better-suited for mining, so their prices spiked even harder than Nvidia's did.)
Also, be aware the GF1xxx series MSRPs were already inflated by mining demand. The "normal" price for a top-end xx80 Ti is $500-$600, and the only-slightly-slower xx80 is usually $350-$400. But the 9xx series was already selling for nearly double MSRP when the 10xx series came out, they'd have been idiots to not bump up the stock prices.
Why should he? The peace negotiations are between North Korea and South Korea, and last I checked, Trump wasn't leader of either of them.
I'm sure there's a lot that's gone on behind closed doors to prepare for this, stuff we don't yet know, but I'd sooner believe it was masterminded by Xi Jinping than by Trump.
I have a suspicion, based on not much beyond "what would *I* do in his place?" contemplations.
Had Kim attempted to negotiate peace at the start of his rule, it likely would have resulted in a coup attempt. Peace is almost certainly going to ruin a lot of NK's "aristocracy" - the generals and the others who aren't at the top, but are still higher than everyone else. They'd be willing to kill to keep the power and wealth they have.
So Kim had to make his rule unquestioned. He's purged plenty of people, presumably those who'd be positioned to fight over it. He had some family murdered to keep them from even potentially being puppet replacements. And he's done plenty of internal propaganda about their nukes and missiles so the peasants and soldiers will see this as a negotiation from a place of strength, not a surrender.
Not only that, but dictators don't often retire peacefully. You die in office, of old age (if you're lucky) or to assassins, rebels or a foreign army (if you're not). And living in fear of peasant revolt or American drone strikes doesn't seem like a good way to live, to me. But if he negotiates peace, he gets to keep basically all his money, and then goes down in the history books as a generous, benevolent peacemaker.
Is this what happened? Maybe. We don't know, maybe never really will. But it's not impossible.
The webcast, being aimed primarily at a US audience, used exclusively Imperial units (feet altitude, miles per hour speed), save for one mention of the Karman line, which was defined (accurately) as 100km.
I am sure the entire engineering and operations team worked in SI, but the stuff the public was shown was not.
I was in the suburbs once, in an area I know well (I was using Maps to get home), deliberately missed a turn so I could stop for gas, and Google Maps told me to take a U-turn at the next light, despite there actually being a fast backup path by going further up the road.
Clearly not having learned my lesson with Latin, I *have* been attempting to teach myself Esperanto. I might try that when I feel like I could actually carry on a conversation, in the one-in-a-million chance that the person calling actually speaks Esperanto. I will report back when I finally give it a try.
That is reasonable and probably quite effective, but I have a more fun solution. I instead answer the phone in Classical Latin - "Salue, hic [nomen] est, quomodo audiuem?". If they respond as though nothing weird just happened, they're either a robot or a person who's indistinguishable from one, and I hang up, often after threatening some manner of debauchery with them, their parents, their progeny, their pets, or all of the above. (Latin is *great* for that sort of thing.)
If they get confused, or try greeting me again in Spanish, they're a human, and I switch back to English like nothing ever happened.
Sure, that's probably less effective, and it *has* resulted in a few odd looks in the office. But dammit, I spent three years learning that language, I am *going* to find a way to use it.
Some of the faked DK runs were *also* on camera - but they were indisputably faked. It seems that Billy was recording runs in MAME (editing either recorded inputs as in a TAS, or just splicing the video) and then playing them back on a screen while pretending to be playing. There's no proof the Pacman runs were faked but there's also no proof they were legit, since Billy is clearly not above cheating. Given the nature of the contest, it seems prudent to err on the side of caution and revoke all his records.
Twin Galaxies is part of the high scores community, not the speedrunning community. There's actually less overlap than you'd think - partially because TG was widely seen as not up to the standards of the speedrunners. (Other reasons are just the different game eras - TG/highscores tend to be pre-NES while speedruns are primarily NES and later)
The new crop of management at TG definitely seems to be headed the right direction, but I don't think most people have forgiven them yet. It's been a common criticism that TG took way, way too long with their "investigation", and was far too willing to entertain the theories of Billy's defender club. The general mood appears to be "guarded optimism".
There's been talk of trying to replace TG but so far nothing's come of it, and I think it even less likely now that they're actually purging the fraudulent records.
Speaking as someone whose daily-use computers range from an Intel HD 4400 to a GeForce1080 Ti:
Intel's integrated graphics are perfectly fine for what they needed to be - something that can run a basic compositor and decode video. For those use cases, what they've got offers enough performance, and does so with the least power draw and the smallest die area. And, arguably, their drivers are the best from a standards-following standpoint. They don't do the "hand-write optimized shaders for specific high-profile games" bullshit Nvidia and AMD are doing, but Intel's doing way better than eg. Qualcomm or ARM drivers.
Intel's architecture isn't bad. It's only weak because they're producing such small designs. I don't know how well it will scale up, but graphics is inherently pretty scalable, so I wouldn't be surprised if it you get a full 2x the performance out of 2x the execution units. And with dedicated memory and a nice, wide GDDR5 or HBM memory bus, they should do a lot better at feeding the thing. (And they've got a GDDR5 interface from Xeon Phi to reuse)
I don't think trying to make CPU/GPU numbers comparable between vendors is a good idea - whatever standard is used WILL be abused and exploited, to the detriment of actual performance if need be.
But within each vendor, there should be general ways to tell performance based on a model number and a simple, consistent numbering scheme.
* Some number needs to indicate relative performance. A higher number here should indicate higher performance in every reasonable usage. These do not need to be on an absolute scale - eg. if a CPU700 is twice as fast as a CPU500, a CPU600 could be anything from 110% to 190% the speed of the CPU500.
* Some number should indicate generation or featureset. If you added Feature X in Generation 2, there should be a 2 somewhere in the model number, and anything with a 2 or higher in that spot should have Feature X. Socket compatibility might also be good to put here - one ought to be able to say "this motherboard will work with any [vendor] CPU starting with [digit]", which is sadly not currently the case.
* Certain features ought to be knowable given only the model number. For CPUs, core count is important - I really hate how Intel will label dual-core mobile chips as "i7" or "i5", markers that absolutely indicate core/thread count on desktop. And it seems AMD is copying them now. On the GPU side, I would go with either CU count, as AMD is doing with Vega, or memory bus width.
The signal-to-noise ratio in a lot of subjects gets so low that, even if you can immediately and perfectly identify a bot as such, you'll still be unable to have useful communication because you'll have to scroll through dozens of bots to get to an actual person - who is probably just arguing with bots.
For instance, click on *any* trending hashtag. Top one right now, for me, is "#MondayMotivation", a recurring hashtag for people looking for some vague platitudes to try to motivate themselves on a Monday morning. First two (again, for me, at the time I clicked the link) were relatively on-topic, by "verified" twits. One's @DisneyAnimation posting a gif from Mulan's training montage, the next is some platitudes from @VexKing, whoever that is. Both of these are hours old - two and six hours, respectively.
Next one is some idiot yammering about how the Mueller probe is fake news, only dems colluded, #maga. It's either a bot or a human who fails the turing test. I could see a tenuous connection to the topic, but there's no actual connection given. The hashtag was just thrown on there to get views and responses.
Then there's a RT/follow "competition", obviously a bot, no relevance whatsoever to the topic.
I'm not going to bother combing through the rest, because I've got better shit to do, but you can see how roughly half the "above the fold" tweets were noise. And it's pretty apparent that Twitter's doing some sorting to put "verified" twits at the top, to try to boost the signal.
You wrote this long ass post when you have no idea what you're talking about.
I will readily concede that I do not speak Chinese, in any form. But I am interested in linguistics, albeit as a hobbyist, and I know enough to know what questions need to be asked, if not the answers.
Even if only for replication, I would have expected the paper to make *some* mention of the language side of things. Instead, they seem to have left it at "we threw this giant corpus of text at it, and hired a bunch of people to evaluate samples manually and they said we did pretty good" with no further detail. I'd expect even an English text corpus to mention whether it was US, UK, Commonwealth or some other variety of English, so seeing that never stated explicitly when dealing with a "language" with far more internal variety is worrying.
Mandarin and Cantonese and so forth are very different spoken, but really are almost the EXACT SAME FUCKING THING when written down. Especially if you avoid idioms and slangs and so forth.
So, under conditions where machine translation has trouble, you have to be cognizant of differences between topolects. Why, then, would a paper on improving the state of machine translation not even address the issue? Even a passing "regional variations did not present a problem" or "so far the net is only trained on MSM" would have been enough.
You seem to have heard this somewhere, which is why you say "The nature of the logographic writing system elides a lot of differences" (and seriously, who talks that way) but then you babble on with your ignorant bullshit.
I, apparently, talk that way. I usually try not to overdo it, since it can come off as pretentious, but I typed that first post on a tablet and didn't feel like using less concise phrasing, since typing it was a pain already. (Seriously, most of my posts are *way* longer than that one)
And the difference between simplified and traditional is nothing to a computer. A simple lookup table is really the only difference. So not relevant if the machine translator uses simplified.
That does not track with what I've read. First, if you're in Unicode, a lookup table is unnecessary - thanks to Han Unification, Simple and Traditional glyphs are mapped to the same Unicode code points, so a lookup table shouldn't even be necessary unless you're accepting inputs, or doing internal processing, in a non-Unicode encoding.
Second, I regularly see complaints about Han Unification causing problems. Mostly it's in the context of Kanji/Hanja but unless I'm misremembering, there are problems trying to treat Traditional and Simplified Chinese characters identically.
Anyway, if you actually understood Chinese a little bit you could just plug in a text and look at the translation (which to my mind was very impressive, if not perfect). But instead you don't know anything of what you're talking about...so you posted a long ass post about it?
I made a post raising questions. I probably ought to have written it in a more "this seems off to me, can someone who knows better comment?" way, but I guess I assumed my tone would carry through. And I realize now that the way I wrote it seemed like I was speaking with more authority than I actually claim. Guess that'll teach me not to post on/. at 2am on my tablet...
Please lurk more and post less. You have nothing to contribute to Slashdot.
Even if you disagree with the points I raised, this kind of incivility is uncalled for.
I read the MS blog and skimmed the actual paper. It gives a decent overview of the system design but has basically no details on the linguistics side of things. They just hired a bunch of people to do manual translation, both for training and for testing, but the only details of the results are a single table summarizing what categories of errors occurred.
A lot of relevant information was missing. To start with, saying "Chinese language" is like saying "European language" - there isn't one unified "Chinese", but rather a variety of languages, topolects and dialects, with some level of mutual intelligibility, but it varies considerably. Not all variants use the same writing system - most use Hanzi, but there's the whole Traditional vs. Simplified issue, and some obscure varieties use entirely different systems (eg. Dungan is written using Cyrillic, despite being closer to Mandarin than many Hanzi-using topolects). And secondary writing systems abound - for teaching and for computer usage, both the Latin alphabet and Bopomofo syllabary are used, in the mainland and Taiwan, respectively.
From context, they seem to be aiming for Mandarin Chinese, the most common variety, and they only accept input in Simplified Hanzi, but they don't make that at all clear from the paper. Was the training corpus exclusively Mandarin, or did it include Cantonese or Hakka or Minnan? Was it entirely Mainstream Mandarin, or were regional dialects like Sichuanese included? The nature of the logographic writing system elides a lot of differences, but I can't see how you could completely ignore the issue. At the very least, I would expect it would be a problem for false negatives in the validation - these are issues for human translators as well. Did they dig deeper into the reported translation issues, and find any were a case of "oh, the news article was written in MSM but quoted someone using Dalian dialect" and then have to figure out whether the human or the machine was more accurate? I didn't read the paper thoroughly but I didn't see any mention at all of any of this crap.
Anyways, they may or may not have made progress on the AI front. I am even less qualified to judge that than I am the linguistics side of it. But there's so many things *not* discussed in the paper that I can't help but feel like they're overstating their results. Guess I'll have to wait for the language blogs to pick up on it.
Also, while the People's Republic of China's international influence is growing, their language is not growing similarly. Few people are learning Mandarin in order to better speak with Chinese people - rather, Chinese people are learning English in order to better speak with the rest of the world. The only real growth area for Mandarin comes from displacing other Sino-Tibetan languages within PRC borders - trying to displace Cantonese from Hong Kong, or exterminate the regional languages of the provinces that aren't very happy being in the PRC. In fact, Mandarin is probably in net negative growth right now - you can wander the halls of a Chinese university and overhear people conversing in English, because all the academic publications are in English and they learned all their terminology as English.
Vulkan is easier to use properly than OpenGL. OpenGL is designed around an architecture that no longer exists - no graphics chip made in the past two decades was strictly fixed-function. And because *every* modern graphics feature is an optional extension to OpenGL, every OpenGL program that wants to take advantage of, say, programmable shaders or compressed textures, has to spend a few hundred lines of code telling the drivers and runtime that yes, it knows what a fucking shader is and can you please let me use them now? Compare modern OpenGL to Direct3D 11 - features nobody actually uses (like fixed-function lighting) are removed, features everyone has (like shaders) are built-in, and while a basic demo *is* a few hundred lines of C, it's much more grokkable. And D3D11 has options for compatibility down to D3D9FL1 (GeForce 5000, GMA 900, Radeon 9000 series) hardware - you can code to D3D11 and still run on older hardware if you limit yourself to certain features or make the right fallback options.
OpenGL is easier to use for basic demos but is harder to use in practice. In particular, optimizing OpenGL is sheer madness, because everything is abstracted. You're basically just rearranging stuff to figure out how to make the drivers generate the right code. Whereas on Vulkan you actually interact with the API in the same way the hardware works - you ask a special malloc() for memory on the GPU, put some data in it, then tell your vertex shader how to interpret that data as triangles, then tell your pixel shader how to turn those triangles into pixels. It only looks like more code if you were relying on OpenGL defaults instead of explicitly coding what you're doing. Since very few apps do Gouraud-shaded, fixed-light, untextured rendering anymore except for graphics API tutorials, any practical program in Vulkan will be easier to write than the equivalent OpenGL program.
What Vulkan does is make trivial stuff harder, but hard stuff easier. You can write a multi-threaded OpenGL renderer, I believe, though I've never been masochistic enough to try. It's easier on Vulkan - still not trivial, but what multithreaded code is?
"New language on an existing branch within a fairly well-studied family" seems fairly niche for a/. article. I'm sure the Austro-asiatic linguistics blogs are all over this, but new languages get discovered all the time. I'd only expect to see it on non-linguistics news sites if there was something special about it - if it was an isolate, or contained an unusual feature, for instance.
(Also: the article summary misspelled "Austro-asiatic", omitting the "R".)
I have no doubt that someone in this clusterfuck of an administration thinks "privatizing" the ISS would be a good idea, but I can't even see how it's a full-fledged idea, let alone a good one.
What's to be privatized? The entire station? Operations? Transit and resupply? All of it?
Private industry doesn't want to own the ISS. It was tailor-made for science, not for tourism, so it would make an awful space hotel. There isn't much demand for science on it from private industry - there's some, but the bulk of it is for NASA et al. trying to figure out how to do deep space exploration better. And let's not forget, half the ISS belongs to other countries. You could probably convince Russia to part with it for enough cash, but Japan? Canada? Seems unlikely. Not to mention, the ISS is nearing end-of-life - it's planned to be de-orbited sometime in the 2020s, because it's just not worth the cost of keeping it running past its designed lifespan.
Operations (replacing NASA's Johnson Space Center with private contractors) is vaguely doable but it doesn't play to private industry strengths. It's a one-off thing, no economies of scale, and it's so safety-critical that you can't shave much cost without risking lives. It's a zero-income project so the only way to squeeze a better profit out is to reduce expenses, and I just don't see how you could do that by any meaningful amount without inviting disaster. If this happens, it's a money-grab - some contractor with lots of donations to the GOP and/or direct connections to Trump will get a contract for several times what we currently pay, and they still will probably fuck it up.
As for replacing transit and resupply... we're already doing that. The Commercial Cargo Development program started under the Bush presidency, and Commercial Crew Development started under Obama. First crewed flights are expected this year. So this is just more of the Trump regime taking credit for stuff Obama (and Bush) did, while doing their best to burn everything to the ground.
I have noticed a lot of tech/computer nerds have a significant interest in language nerdery. I've seen/. threads devolve into arguments over correct Latin grammar. This certainly piques the interest of people who have a bit of language nerd in them, because it's as much about knowledge of old writing systems and abbreviations as it is ability to look at squiggly lines and pattern-match.
Slight correction: the core ran out of ignition fluid (a mix of triethylborane and triethylaluminum, ignites on contact with LOX (or most anything, really)), not fuel. A similar setup was used for both the Saturn V's F-1 engines, and the SR-71's J58 engines.
And a status update: the second stage re-lit just fine, and in fact exceeded expectations - the aphelion of the orbit is well past Mars, just shy of Ceres in fact.
I upgraded to 10 pretty quickly after it was out - first upgrading my Win8 machine (I actually bought a disc copy of Windows 8), then eventually my Win7 machines as it got better. My entire array of Windows boxes are now on 10, and I haven't felt any desire to go back.
Google Translate can also produce seemingly-sensible results when given senseless inputs. Getting some meaningful output is only a weak suggestion that they have meaningful inputs. They should not have published without finding at least one Hebrew scholar who would take a look at their work - and the fact that they couldn't convince anyone to do so is itself suggestive.
Slashdot Mirror
The top-end cards became halo products. Just like few cars on the road are Mustangs, few video cards are 1080 Tis.
Checking the Steam Hardware Survey, about 60% of all cards are Nvidia GeForces in the xx50 to xx70 range - the normal, reasonably-priced cards. (AMD is only 10% of the market right now, with much of that being integrated. Their fundamental architecture was just way better-suited for mining, so their prices spiked even harder than Nvidia's did.)
Also, be aware the GF1xxx series MSRPs were already inflated by mining demand. The "normal" price for a top-end xx80 Ti is $500-$600, and the only-slightly-slower xx80 is usually $350-$400. But the 9xx series was already selling for nearly double MSRP when the 10xx series came out, they'd have been idiots to not bump up the stock prices.
Why should he? The peace negotiations are between North Korea and South Korea, and last I checked, Trump wasn't leader of either of them.
I'm sure there's a lot that's gone on behind closed doors to prepare for this, stuff we don't yet know, but I'd sooner believe it was masterminded by Xi Jinping than by Trump.
I have a suspicion, based on not much beyond "what would *I* do in his place?" contemplations.
Had Kim attempted to negotiate peace at the start of his rule, it likely would have resulted in a coup attempt. Peace is almost certainly going to ruin a lot of NK's "aristocracy" - the generals and the others who aren't at the top, but are still higher than everyone else. They'd be willing to kill to keep the power and wealth they have.
So Kim had to make his rule unquestioned. He's purged plenty of people, presumably those who'd be positioned to fight over it. He had some family murdered to keep them from even potentially being puppet replacements. And he's done plenty of internal propaganda about their nukes and missiles so the peasants and soldiers will see this as a negotiation from a place of strength, not a surrender.
Not only that, but dictators don't often retire peacefully. You die in office, of old age (if you're lucky) or to assassins, rebels or a foreign army (if you're not). And living in fear of peasant revolt or American drone strikes doesn't seem like a good way to live, to me. But if he negotiates peace, he gets to keep basically all his money, and then goes down in the history books as a generous, benevolent peacemaker.
Is this what happened? Maybe. We don't know, maybe never really will. But it's not impossible.
The webcast, being aimed primarily at a US audience, used exclusively Imperial units (feet altitude, miles per hour speed), save for one mention of the Karman line, which was defined (accurately) as 100km.
I am sure the entire engineering and operations team worked in SI, but the stuff the public was shown was not.
I was in the suburbs once, in an area I know well (I was using Maps to get home), deliberately missed a turn so I could stop for gas, and Google Maps told me to take a U-turn at the next light, despite there actually being a fast backup path by going further up the road.
Clearly not having learned my lesson with Latin, I *have* been attempting to teach myself Esperanto. I might try that when I feel like I could actually carry on a conversation, in the one-in-a-million chance that the person calling actually speaks Esperanto. I will report back when I finally give it a try.
That is reasonable and probably quite effective, but I have a more fun solution. I instead answer the phone in Classical Latin - "Salue, hic [nomen] est, quomodo audiuem?". If they respond as though nothing weird just happened, they're either a robot or a person who's indistinguishable from one, and I hang up, often after threatening some manner of debauchery with them, their parents, their progeny, their pets, or all of the above. (Latin is *great* for that sort of thing.)
If they get confused, or try greeting me again in Spanish, they're a human, and I switch back to English like nothing ever happened.
Sure, that's probably less effective, and it *has* resulted in a few odd looks in the office. But dammit, I spent three years learning that language, I am *going* to find a way to use it.
"Bravely" is an adverb.
Some of the faked DK runs were *also* on camera - but they were indisputably faked. It seems that Billy was recording runs in MAME (editing either recorded inputs as in a TAS, or just splicing the video) and then playing them back on a screen while pretending to be playing. There's no proof the Pacman runs were faked but there's also no proof they were legit, since Billy is clearly not above cheating. Given the nature of the contest, it seems prudent to err on the side of caution and revoke all his records.
Twin Galaxies is part of the high scores community, not the speedrunning community. There's actually less overlap than you'd think - partially because TG was widely seen as not up to the standards of the speedrunners. (Other reasons are just the different game eras - TG/highscores tend to be pre-NES while speedruns are primarily NES and later)
The new crop of management at TG definitely seems to be headed the right direction, but I don't think most people have forgiven them yet. It's been a common criticism that TG took way, way too long with their "investigation", and was far too willing to entertain the theories of Billy's defender club. The general mood appears to be "guarded optimism".
There's been talk of trying to replace TG but so far nothing's come of it, and I think it even less likely now that they're actually purging the fraudulent records.
Speaking as someone whose daily-use computers range from an Intel HD 4400 to a GeForce1080 Ti:
Intel's integrated graphics are perfectly fine for what they needed to be - something that can run a basic compositor and decode video. For those use cases, what they've got offers enough performance, and does so with the least power draw and the smallest die area. And, arguably, their drivers are the best from a standards-following standpoint. They don't do the "hand-write optimized shaders for specific high-profile games" bullshit Nvidia and AMD are doing, but Intel's doing way better than eg. Qualcomm or ARM drivers.
Intel's architecture isn't bad. It's only weak because they're producing such small designs. I don't know how well it will scale up, but graphics is inherently pretty scalable, so I wouldn't be surprised if it you get a full 2x the performance out of 2x the execution units. And with dedicated memory and a nice, wide GDDR5 or HBM memory bus, they should do a lot better at feeding the thing. (And they've got a GDDR5 interface from Xeon Phi to reuse)
I don't think trying to make CPU/GPU numbers comparable between vendors is a good idea - whatever standard is used WILL be abused and exploited, to the detriment of actual performance if need be.
But within each vendor, there should be general ways to tell performance based on a model number and a simple, consistent numbering scheme.
* Some number needs to indicate relative performance. A higher number here should indicate higher performance in every reasonable usage. These do not need to be on an absolute scale - eg. if a CPU700 is twice as fast as a CPU500, a CPU600 could be anything from 110% to 190% the speed of the CPU500.
* Some number should indicate generation or featureset. If you added Feature X in Generation 2, there should be a 2 somewhere in the model number, and anything with a 2 or higher in that spot should have Feature X. Socket compatibility might also be good to put here - one ought to be able to say "this motherboard will work with any [vendor] CPU starting with [digit]", which is sadly not currently the case.
* Certain features ought to be knowable given only the model number. For CPUs, core count is important - I really hate how Intel will label dual-core mobile chips as "i7" or "i5", markers that absolutely indicate core/thread count on desktop. And it seems AMD is copying them now. On the GPU side, I would go with either CU count, as AMD is doing with Vega, or memory bus width.
The signal-to-noise ratio in a lot of subjects gets so low that, even if you can immediately and perfectly identify a bot as such, you'll still be unable to have useful communication because you'll have to scroll through dozens of bots to get to an actual person - who is probably just arguing with bots.
For instance, click on *any* trending hashtag. Top one right now, for me, is "#MondayMotivation", a recurring hashtag for people looking for some vague platitudes to try to motivate themselves on a Monday morning. First two (again, for me, at the time I clicked the link) were relatively on-topic, by "verified" twits. One's @DisneyAnimation posting a gif from Mulan's training montage, the next is some platitudes from @VexKing, whoever that is. Both of these are hours old - two and six hours, respectively.
Next one is some idiot yammering about how the Mueller probe is fake news, only dems colluded, #maga. It's either a bot or a human who fails the turing test. I could see a tenuous connection to the topic, but there's no actual connection given. The hashtag was just thrown on there to get views and responses.
Then there's a RT/follow "competition", obviously a bot, no relevance whatsoever to the topic.
I'm not going to bother combing through the rest, because I've got better shit to do, but you can see how roughly half the "above the fold" tweets were noise. And it's pretty apparent that Twitter's doing some sorting to put "verified" twits at the top, to try to boost the signal.
You wrote this long ass post when you have no idea what you're talking about.
I will readily concede that I do not speak Chinese, in any form. But I am interested in linguistics, albeit as a hobbyist, and I know enough to know what questions need to be asked, if not the answers.
Even if only for replication, I would have expected the paper to make *some* mention of the language side of things. Instead, they seem to have left it at "we threw this giant corpus of text at it, and hired a bunch of people to evaluate samples manually and they said we did pretty good" with no further detail. I'd expect even an English text corpus to mention whether it was US, UK, Commonwealth or some other variety of English, so seeing that never stated explicitly when dealing with a "language" with far more internal variety is worrying.
Mandarin and Cantonese and so forth are very different spoken, but really are almost the EXACT SAME FUCKING THING when written down. Especially if you avoid idioms and slangs and so forth.
So, under conditions where machine translation has trouble, you have to be cognizant of differences between topolects. Why, then, would a paper on improving the state of machine translation not even address the issue? Even a passing "regional variations did not present a problem" or "so far the net is only trained on MSM" would have been enough.
You seem to have heard this somewhere, which is why you say "The nature of the logographic writing system elides a lot of differences" (and seriously, who talks that way) but then you babble on with your ignorant bullshit.
I, apparently, talk that way. I usually try not to overdo it, since it can come off as pretentious, but I typed that first post on a tablet and didn't feel like using less concise phrasing, since typing it was a pain already. (Seriously, most of my posts are *way* longer than that one)
And the difference between simplified and traditional is nothing to a computer. A simple lookup table is really the only difference. So not relevant if the machine translator uses simplified.
That does not track with what I've read. First, if you're in Unicode, a lookup table is unnecessary - thanks to Han Unification, Simple and Traditional glyphs are mapped to the same Unicode code points, so a lookup table shouldn't even be necessary unless you're accepting inputs, or doing internal processing, in a non-Unicode encoding.
Second, I regularly see complaints about Han Unification causing problems. Mostly it's in the context of Kanji/Hanja but unless I'm misremembering, there are problems trying to treat Traditional and Simplified Chinese characters identically.
Anyway, if you actually understood Chinese a little bit you could just plug in a text and look at the translation (which to my mind was very impressive, if not perfect). But instead you don't know anything of what you're talking about...so you posted a long ass post about it?
I made a post raising questions. I probably ought to have written it in a more "this seems off to me, can someone who knows better comment?" way, but I guess I assumed my tone would carry through. And I realize now that the way I wrote it seemed like I was speaking with more authority than I actually claim. Guess that'll teach me not to post on /. at 2am on my tablet...
Please lurk more and post less. You have nothing to contribute to Slashdot.
Even if you disagree with the points I raised, this kind of incivility is uncalled for.
I read the MS blog and skimmed the actual paper. It gives a decent overview of the system design but has basically no details on the linguistics side of things. They just hired a bunch of people to do manual translation, both for training and for testing, but the only details of the results are a single table summarizing what categories of errors occurred.
A lot of relevant information was missing. To start with, saying "Chinese language" is like saying "European language" - there isn't one unified "Chinese", but rather a variety of languages, topolects and dialects, with some level of mutual intelligibility, but it varies considerably. Not all variants use the same writing system - most use Hanzi, but there's the whole Traditional vs. Simplified issue, and some obscure varieties use entirely different systems (eg. Dungan is written using Cyrillic, despite being closer to Mandarin than many Hanzi-using topolects). And secondary writing systems abound - for teaching and for computer usage, both the Latin alphabet and Bopomofo syllabary are used, in the mainland and Taiwan, respectively.
From context, they seem to be aiming for Mandarin Chinese, the most common variety, and they only accept input in Simplified Hanzi, but they don't make that at all clear from the paper. Was the training corpus exclusively Mandarin, or did it include Cantonese or Hakka or Minnan? Was it entirely Mainstream Mandarin, or were regional dialects like Sichuanese included? The nature of the logographic writing system elides a lot of differences, but I can't see how you could completely ignore the issue. At the very least, I would expect it would be a problem for false negatives in the validation - these are issues for human translators as well. Did they dig deeper into the reported translation issues, and find any were a case of "oh, the news article was written in MSM but quoted someone using Dalian dialect" and then have to figure out whether the human or the machine was more accurate? I didn't read the paper thoroughly but I didn't see any mention at all of any of this crap.
Anyways, they may or may not have made progress on the AI front. I am even less qualified to judge that than I am the linguistics side of it. But there's so many things *not* discussed in the paper that I can't help but feel like they're overstating their results. Guess I'll have to wait for the language blogs to pick up on it.
Mandarin is Sino-Tibetan, not Indo-European.
Also, while the People's Republic of China's international influence is growing, their language is not growing similarly. Few people are learning Mandarin in order to better speak with Chinese people - rather, Chinese people are learning English in order to better speak with the rest of the world. The only real growth area for Mandarin comes from displacing other Sino-Tibetan languages within PRC borders - trying to displace Cantonese from Hong Kong, or exterminate the regional languages of the provinces that aren't very happy being in the PRC. In fact, Mandarin is probably in net negative growth right now - you can wander the halls of a Chinese university and overhear people conversing in English, because all the academic publications are in English and they learned all their terminology as English.
Vulkan is easier to use properly than OpenGL. OpenGL is designed around an architecture that no longer exists - no graphics chip made in the past two decades was strictly fixed-function. And because *every* modern graphics feature is an optional extension to OpenGL, every OpenGL program that wants to take advantage of, say, programmable shaders or compressed textures, has to spend a few hundred lines of code telling the drivers and runtime that yes, it knows what a fucking shader is and can you please let me use them now? Compare modern OpenGL to Direct3D 11 - features nobody actually uses (like fixed-function lighting) are removed, features everyone has (like shaders) are built-in, and while a basic demo *is* a few hundred lines of C, it's much more grokkable. And D3D11 has options for compatibility down to D3D9FL1 (GeForce 5000, GMA 900, Radeon 9000 series) hardware - you can code to D3D11 and still run on older hardware if you limit yourself to certain features or make the right fallback options.
OpenGL is easier to use for basic demos but is harder to use in practice. In particular, optimizing OpenGL is sheer madness, because everything is abstracted. You're basically just rearranging stuff to figure out how to make the drivers generate the right code. Whereas on Vulkan you actually interact with the API in the same way the hardware works - you ask a special malloc() for memory on the GPU, put some data in it, then tell your vertex shader how to interpret that data as triangles, then tell your pixel shader how to turn those triangles into pixels. It only looks like more code if you were relying on OpenGL defaults instead of explicitly coding what you're doing. Since very few apps do Gouraud-shaded, fixed-light, untextured rendering anymore except for graphics API tutorials, any practical program in Vulkan will be easier to write than the equivalent OpenGL program.
What Vulkan does is make trivial stuff harder, but hard stuff easier. You can write a multi-threaded OpenGL renderer, I believe, though I've never been masochistic enough to try. It's easier on Vulkan - still not trivial, but what multithreaded code is?
Function as a useful currency?
"New language on an existing branch within a fairly well-studied family" seems fairly niche for a /. article. I'm sure the Austro-asiatic linguistics blogs are all over this, but new languages get discovered all the time. I'd only expect to see it on non-linguistics news sites if there was something special about it - if it was an isolate, or contained an unusual feature, for instance.
(Also: the article summary misspelled "Austro-asiatic", omitting the "R".)
I think 20% is a pretty low estimate. Tack on another zero and you might be in the right ballpark.
I have no doubt that someone in this clusterfuck of an administration thinks "privatizing" the ISS would be a good idea, but I can't even see how it's a full-fledged idea, let alone a good one.
What's to be privatized? The entire station? Operations? Transit and resupply? All of it?
Private industry doesn't want to own the ISS. It was tailor-made for science, not for tourism, so it would make an awful space hotel. There isn't much demand for science on it from private industry - there's some, but the bulk of it is for NASA et al. trying to figure out how to do deep space exploration better. And let's not forget, half the ISS belongs to other countries. You could probably convince Russia to part with it for enough cash, but Japan? Canada? Seems unlikely. Not to mention, the ISS is nearing end-of-life - it's planned to be de-orbited sometime in the 2020s, because it's just not worth the cost of keeping it running past its designed lifespan.
Operations (replacing NASA's Johnson Space Center with private contractors) is vaguely doable but it doesn't play to private industry strengths. It's a one-off thing, no economies of scale, and it's so safety-critical that you can't shave much cost without risking lives. It's a zero-income project so the only way to squeeze a better profit out is to reduce expenses, and I just don't see how you could do that by any meaningful amount without inviting disaster. If this happens, it's a money-grab - some contractor with lots of donations to the GOP and/or direct connections to Trump will get a contract for several times what we currently pay, and they still will probably fuck it up.
As for replacing transit and resupply... we're already doing that. The Commercial Cargo Development program started under the Bush presidency, and Commercial Crew Development started under Obama. First crewed flights are expected this year. So this is just more of the Trump regime taking credit for stuff Obama (and Bush) did, while doing their best to burn everything to the ground.
I have noticed a lot of tech/computer nerds have a significant interest in language nerdery. I've seen /. threads devolve into arguments over correct Latin grammar. This certainly piques the interest of people who have a bit of language nerd in them, because it's as much about knowledge of old writing systems and abbreviations as it is ability to look at squiggly lines and pattern-match.
Slight correction: the core ran out of ignition fluid (a mix of triethylborane and triethylaluminum, ignites on contact with LOX (or most anything, really)), not fuel. A similar setup was used for both the Saturn V's F-1 engines, and the SR-71's J58 engines.
And a status update: the second stage re-lit just fine, and in fact exceeded expectations - the aphelion of the orbit is well past Mars, just shy of Ceres in fact.
Guess I'm part of the 1% now.
I upgraded to 10 pretty quickly after it was out - first upgrading my Win8 machine (I actually bought a disc copy of Windows 8), then eventually my Win7 machines as it got better. My entire array of Windows boxes are now on 10, and I haven't felt any desire to go back.
Google Translate can also produce seemingly-sensible results when given senseless inputs. Getting some meaningful output is only a weak suggestion that they have meaningful inputs. They should not have published without finding at least one Hebrew scholar who would take a look at their work - and the fact that they couldn't convince anyone to do so is itself suggestive.