dplavery92

For many of us, the Inception-style CNN workloads--especially at FP32--are much more realistic than large language models that may be better suited to take advantage of the tensor cores. If I'm going to be memory bottlenecked either way, I probably don't want to spend an extra $1000 on 400 tensor cores I can't take full advantage of.

For what it's worth, the "Free to Play" sector that dominates a lot of market share isn't just the gambling-lite, pay-to-win mobile sector anymore. We're talking about major titles like Fortnite, League of Legends, DOTA2, Rocket League, Hearthstone, and Valorant, which have millions of concurrent players, dominate streaming services, and often have high-production e-sports events. Even the latest title in the Call of Duty franchise offers a Free to Play Battleground mode.

I don't know if there are any white papers with hard details yet (if anyone knows of one, please share!), but nVidia's marketing material[0] for the Ampere architecture claims the following:

"Sparsity is possible in deep learning because the importance of individual weights evolves during the learning process, and by the end of network training, only a subset of weights have acquired a meaningful purpose in determining the learned output. The remaining weights are no longer needed.

Fine grained structured sparsity imposes a constraint on the allowed sparsity pattern, making it more efficient for hardware to do the necessary alignment of input operands. Because deep learning networks are able to adapt weights during the training process based on training feedback, NVIDIA engineers have found in general that the structure constraint does not impact the accuracy of the trained network for inferencing. This enables inferencing acceleration with sparsity."

So the idea seems to be that at the end of training, there's fine tuning that can be done to figure out which weights can be zeroed out without significantly impacting prediction accuracy, and then you can accelerate inferences with sparse matrix multiplication. They consider training acceleration with sparse matrices an "active research area."

I could see it being nice for the sake of running large language models on consumer, or really cool for the few edge computing applications that can actually demand and power conventional GPUs (e.g. self-driving cars.) It's probably not a great boon to the researcher who wants to reduce their iteration timeline though.

[0] https://developer.nvidia.com/blog/nvidia-ampere-architecture...

>Objects more massive than [2 to 3 times the mass of the Sun] can only be black holes

One (pedantic) correction: objects more massive cannot be held up by neutron degeneracy pressure, but you can have more massive stars held up by radiation pressure from fusion. Certainly, there are many giant stars in the 2-8 solar mass range, and rarer stars in tens-to-hundreds of solar masses. Of course, they will eventually burn away their fusion fuel, and the core remnant may indeed become a black hole.

It's not impossible to distribute training across multiple GPUs, but it's certainly not straightforward. And if you want activations for an entire 4K image in the same model, having a lot of memory on one card is your friend.

I think it holds the market position previously known as xx80Ti

Maybe Vera Rubin belongs on this list as well.

https://en.wikipedia.org/wiki/Vera_Rubin

Usually "breeder" indicates a reactor with a fully closed fuel cycle that generates more fissile fuel than it consumes. A Pu-239 breeder would need to begin with critical mass of Pu-239 to begin this fuel cycle, in addition to the U-238 feedstock you are wanting to bombard. Kind of puts the cart before the horse. Additionally, the engineering for a liquid sodium or molten salt cooling system (water is too strong a fast-neutron moderator for this application) goes beyond the means of the hobbyist or extremist group.

I don't think this is what you meant, though. Why can't [an independent actor] recover Pu-239 from the core of a uranium reactor? I'll admit I don't know a lot about chemical reprocessing, but the proposition here is using a series of highly corrosive and/or explosive chemicals ([fuming] nitric acid, tributyl phosphate, hydrazine) to extract Pu-239 from very radioactive irradiated fuel rods with unknown actinide concentrations. Chemical reagents for this process (e.g. for PUREX, tributyl phosphate) are controlled and monitored by non-proliferation agencies. Chemical re-processing will not separate the Pu-240 from the Pu-239. Pu-239 is unsuitable for a gun-type weapon, as it is so fissile that it will blow itself apart in a partial detonation ("fizzle") before the slug and target are brought together--this means you need an implosion-type ("Fat man") weapon to effectively use Pu-239 as a fuel. A high concentration of Pu-240 exacerbates this problem. A fizzle in a populated area is still very bad, but not bad in a way that couldn't feasibly be produced by conventional explosives and/or chemical weapons much more easily. In order to produce enough Pu-239 to form a weapons core within years instead of decades or centuries (provided extraction is feasible), the independent actor probably is looking at a >1 MW reactor with the cooling towers and infrastructure that surrounds a project of that scale.

It is really only a viable threat for a terrorist organization if that organization has the clandestine support of a nation state.

EDIT: Or otherwise enough financial support, technical expertise, and freedom of activity to engage in serious construction projects and industrial chemistry. The point is that it's not something you can put together in a basement lab.

[0] https://en.wikipedia.org/wiki/PUREX

There is a huge gulf between the abundance necessary for criticality and that necessary for prompt criticality. Even rudimentary, low-yield weapons design requires HEU, likely over 50% 235 (and if you have the capability to enrich beyond 20%, the barrier to enriching up towards "weapons grade" is comparatively low.) I won't contradict the FAS's claim that this is within the reach of any _rogue state_, but it's not something you can achieve by bombarding natural ore with your garage-project fusor.

In Shayer's article (not the top-level BBC article) he makes the comment that they developed the iPod OS on Windows machines since the compiler and tools for building on ARM were not yet available on Mac. You make a good point about developing software for other platforms, though, and I'm sure there are also quite a few Mac and Linux boxes in software divisions at Microsoft for the same reason.

If Shayer's (in turn, Apple's) participation were classified, he would be looking at a serious criminal offense for writing the article. The hardware and functionality of the devices that the Bechtel engineers were integrating, about which Shayer only speculates, could well have been so classified, though.

You may be able to find uranium ore with relative ease, but enrichment is a nation-state level project that is unlikely to escape the notice of the world's intelligence and energy agencies.

I've definitely felt and experienced the situation you're describing, but I think you might be painting with too broad a brush.

Sources like ArXiv, Google Scholar, and Semantic Scholar have made it really easy for me to access tons of deep, academic/professional level knowledge in my field instantaneously and for free. I know that Physics, Math, and Computer Science are relatively uniquely privileged when it comes to ArXiv, but even paywalled journals are more accessible in the internet age than they were before. I'm even able to find public-released technical info from the DoD through DTIC that I don't think I would ever know what to request without having it come up through search engines.

Likewise, tons of undergrad classes through MIT and the Ivies are available for free on Youtube; more still are available with exercises and some form of certification through services like Coursera, EdX, and Udacity. Some graduate and professional-level courses are starting to come online through these institutions as well; the Institute for Advanced Study has started releasing its lectures on Youtube over the past few years.

I can find manuals for my car, my appliances, and all my electronics online. I have access to all sorts of do-it-yourself how-tos on Youtube when I want to do work around the house. I have StackOverflow for when I have challenges with programming languages and tools that aren't clear from the docs, and the rest of the StackOverflow network when I have questions that are more specific to my problem domain.

There's more throughout all sorts of fields: instant translations with Google translate and langauges lessons through DuoLingo, Rosetta Stone, et al; music lessons on Youtube, etc.

Again, I've been through the experience where some older materials just aren't as available as I'd like, and where the normal search tools only bring you part of the way. But I think you're taking a lot for granted here.

Ironically, in the DoD contracting world there are roles like this; SETA (systems engineering and technical assurance) and SME (subject matter expert) contracting roles from disinterested third parties are usually a required function in DoD acquisition programs. These folks are basically paid skeptics of the work being done by the prime contractors, which is a good failsafe to build in for >$100MM tax-funded programs and for weapons and vehicles that can kill people. In this particular case, though, the hokum is coming internally from a government lab, and the work is early research, so it doesn't have to meet the same validation standards.

Right, and among them are more[0] patents held by the US Navy for high-temperature superconductors and "high frequency gravity wave generators." It's clear he has at least one patron within the DoD; it's a shame that the technical oversight over this program is so credulous.

[0] https://patents.google.com/?inventor=Salvatore+Cezar+Pais

The "background" section of this particular patent [0] starts pretty sane (if obviously a little grandiose--enough to get one's guard up) but rapidly and suddenly dips off the rails. The trouble is that the mostly unintelligible details about "hyper-frequency" rotational and vibrational coupled modes are interspersed in a bunch of technically correct--if vapid--physical statements about quantum fields pervading all of spacetime, and other phenomena being emergent of excitations in these fields. (Notwithstanding the fact that uniting QFT with gravity is the biggest open problem in Physics today...)

I think this is an illustrating case in how much easier it is to create high-falutin' nonsense than it is to debunk it. Most of the claims are somewhat vague but they touch on areas where someone needs a lot of knowledge and precision to coherently dispute them.

I highly recommend reading a patent lawyer's research into the history of this particular patent from the last time it was posted here[1]. Apparently the examiner tried to reject the patent claim several times on technical bases, against the protest of some Navy brass, before finally admitting it without comment months later.

[0] https://patents.google.com/patent/US10144532B2/en [1] https://news.ycombinator.com/item?id=19763445

Any works you produce using the GPL'd asset must also be GPL'd. So I can put a price-tag on my movie, comic book, AI model, or tutorial, but anyone else can also copy and redistribute my work with no legal recourse from me. It's hard to see the feasibility of commercial use.

I think that's a fair argument to make for GPL when you're talking about software, e.g. GIMP, which you could use to produce commercial art assets that you then sell. But for data, whether these images or audio clips or weather measurements or whatever, the requirement that derivative products also be licensed under GPL makes their use effectively non-commercial.

Our "toy" up there is marvel of global technological cooperation, the subject of an international treaty only a few years after the fall of the Soviet Union, maintained by engineers, scientists, and spacefarers from fifteen different countries. It's a beacon of non-violent collaboration and humanist hope, and a continued platform for science that helps us understand not just how to continue to explore, but also the nature of our own planet and our own biology. And its total $150B cost, between all those fifteen nations since the start of its component developments thirty five years ago, is about 40% of the sum total of SBA loans our government doled out in the last four weeks. Once you pack the tremendous cost of our bailouts to large corporations and financial institutions plus the personal checks being disbursed now, it's less than eight percent the size of the $2T stimulus package the US just passed. I think your perspective needs some adjustment.

Maybe you meant centuries, and not millennia? English, even it's oldest forms, hasn't been around for millennia. Anyway, as your own etymology link demonstrates, the word's first use in its present meaning, as an executive leader of a republic--especially of something the size and importance of a nation state--is first in the US Constitution. Prior to that it generally meant leaders of more modest organizations: "heads of religious houses, hospitals, colleges and universities."

It doesn't appear to me that the article linked in the rebuttal thread appears to address any of the points made in this article. Instead, the author offers that he's written two books and thousands of blog posts in Markdown, and claims (without substantiation) that 98% of all code documentation could be written in Markdown.

Where does that leave us on any of this article's concerns about extensibility and cross-reference? I'm persuaded by the need for a richer set of tags and non-local references for writing code documentation on most of my projects.

Generate a random int from [0,199]. If you generated a 0, you die.

Rationally, you pretty confident that you won't die? Sure. But what behavior changes are you willing to tolerate not to have to take the chance? I sure wouldn't take the bet on a lark, even though the expectation is that I live.

Now think about extending that same game to your family and friends, to the school down the street, to the shopping mall, and to the elder's home in town. Some people are going to roll 0, and there's a real risk that some of those people are people you know and care about. And even if they're not, your community will still be dramatically affected. It could be your car mechanic, your office's custodial staff, or the greybeard in your office who knows how to decipher the old FORTRAN code.

I'm not willing to be flippant about that. Even 1 in 200 people can be devastating emotional, logistical, and financial toll on a community. I'm not saying panic, but I don't think it's smart or responsible to downplay the risks of infection in a disease that is currently spreading exponentially (or, at least, maintains a positive growth ratio.) Canceling gatherings, temporarily closing schools, working from home, etc.--these are all inconvenient, they make our lives and business harder, they're having a negative financial impact. But they're also totally the reasonable course of action in the face of a pretty serious threat.

No one claimed the opposite; someone merely asked for evidence for an assertion.

Since that assertion is important to a larger argument, it's reasonable to ask why we should believe it.

This, I believe, is the exactly the behavior the parent commenting was describing (and bemoaning.) If you you have to check against every term in a linked list (an operation proportional to the length of the list), then your hash table ceases to be constant-time once you get collisions. In that case, if you have a hash table of fixed size, but arbitrarily increase the number of entries in the table, then the look-up time tends as O(n).

Because global warming means not just a slightly less comfortable summer season, but also rising ocean levels that displace low-lying coastal communities (to include not just island nations like Micronesia and The Maldives, but also large sections of major harbor cities like New York City, Los Angeles, Mumbai, and Sydney); more violent and more frequent tropical storms and monsoons; drastic changes to crop viability in semi-arid agricultural regions such as those in California, North Africa, Yemen, and Afghanistan; increased wildfires in California and Australia; worsened air quality in and around major cities (especially megacities); acidification of our oceans; and heightened turmoil in and around arid regions with existing water and refugee crises.

I think there's a latent suggestion in your language that reducing consumption discards all of the technological social progress of the past three centuries, while climate change presents a negotiable inconvenience. The balance is quite the opposite. You can keep cellphones, the internet, food security, modern medicine, and clean, running water. If anything, many of these advances are secured, rather than threatened, by curbing human-driven climate change.

>If you want to go back to stone age, please feel free to go: I'm not retaining you.

Beyond misjudging the gravity of the problem at hand, I think this comment is also a little short-sighted in appreciating the nature of the problem and our collective responsibility for it. The choices we make as consumers, designers, and voters have impacts that affect the whole world, including populations with clear stake but no voice in those decisions. It's particularly callous to tell someone whose island will be washed away by rising sea levels that they're free to eschew any technology they want to curb ecological impact while our own emissions (by absolute value or per capita) far exceed their own contribution to atmospheric CO2 levels.

Harmonics occur at integer multiples of the fundamental. If a body resonates at F, you will see harmonics at F, 2F, 3F...

You will not see harmonics at fractional (1/2 F, etc...) frequencies of the fundamental, though you will likely still see some backscatter.

Retroreflectors[0], to include the Luneburg lenses[1] mentioned in the article, function by directing radar energy parallel to its incidence direction. These reflectors will not significantly increase the plane's RCS for passive/bistatic receivers. Even if these reflectors did increase RCS omni-directionally, it is unlikely that a reflector designed to operate at 3 Ghz (10 cm) would also be designed to operate at 187.5 MHz (1.6 meters).

[0]https://en.wikipedia.org/wiki/Retroreflector [1]https://en.wikipedia.org/wiki/Luneburg_lens#Radar_reflector

Median personal income has not, in fact, been keeping pace with inflation. The median personal income (men and women) in Sep 2016 was $31,099.[0] Which means the $47,600 tuition ($62200 with mandatory room and board)[1] was 1.53x (2.00x) the median personal income in 2016, versus 0.94x in 1940. I'd consider that more than slightly faster growth, especially since most of that proportional growth has happened in the years from 1980 to present, rather than the full 79 years since 1940.[2]

[0] https://fred.stlouisfed.org/series/MEPAINUSA672N [1] https://finaid.yale.edu/sites/default/files/files/EA_FYYE_Cl... [2] https://yalealumnimagazine.com/articles/4074-the-cost-of-yal...

I don't know who.

I think we have seen this sort of approach succeed in a major US TV production in Netflix's Narcos. There are English-speaking characters, yes, but most of the drama happens entirely in Spanish.

I think this series would also have been better fully in Russian with English subtitles.

While 5 megatons definitely does seem like an overestimation, I think the fear was that it was a much larger mass of molten corium, including not just the fuel itself, but also molten silicon and lead, from the material dumped on the exposed reactor to quench the fires and prevent fallout. There was an estimated 5000 tons of this suppressant material, although I don't know how much could have reasonably been expected to melt (and I think later analysis shows that very little of it actually landed on the exposed core itself.)

Then when the molten corium hits the water, the water could act as a neutron moderator for a runaway fission reaction. You have to think this would be more a nuclear "fizzle" than the prompt criticality that is necessary for a >10kt (let alone megaton) bomb, but it would still be bad.

A 5 megaton explosion that levels Kiev? Probably not, but then again I'm not the nuclear physicist who came up with that number.

The idea that blowing the core material into the atmosphere and causing further meltdowns in the other three reactors would be a major radiological disaster for Northern Europe, though, I can appreciate.