My impression is that for every language other than C/C++, there are three things more important than almost any feature of the language or runtime:
(1) the package system - how you search+download+reference them, how dependencies+conflicts are managed, how updates are managed (2) the build system (3) editor integration, be it Emacs or Sublime or VSCode
Are you telling me that struggles with climate change, scarcity of natural resources and so on will just vanish into thin air by embracing Socialism or perhaps a Fascist dictatorship, or Communist rule?
No one is telling you that (save for the click-baity slashdot headline). The article itself boils down to this:
1. Economies are changing to less energy-efficient energy sources for the first time in history (I assume that means wind+solar+hydro), which will disrupt things. The economic models used by governments and industry have traditionally assumed energy abundance, which is becoming a poor assumption. 2. Sink costs are rising - economic growth used to be straightforward with profligate use of energy, use of materials, creation of waste, will become less possible. 3. The models are fine at handling slow incremental changes, but we'll likely see more dramatic changes than they'll be good at handling. 4. Carbon pricing hasn't worked very well and doesn't seem likely to work in future. 5. Post-Keynesian models about the relationship between the state and the free market will give different predictions and metrics-for-success compared to neoclassical economic models. All of them still boil down to free markets with government tweaks+nudges of course. 6. The article sort of tails off into more or less wishing the whole world were like Finland - not surprising given the authors and the paper are from Finland.
[usenet daily volume of 27 terrabytes/day] Did I read that correctly? I thought Usenet was pretty much dead. Many very popular newsgroups, like rec.sports.baseball, that received 1000+ posts/day back in the day were completely empty as of a couple years ago.
Much of this traffic increase reflects not an increase in discrete users or newsgroup discussions, but instead the combination of massive automated spamming and an increase in the use of.binaries newsgroups
[I don't think it would ever scale to the quantity of data (basically, videos)]
It already has and does. I have yet to see full 50GB+ BlueRay rips posted to FB.
The upload rate to Youtube I saw for last year was 270 terrabytes/day with retention period of forever, compared to usenet daily volume of 27 terrabytes/day last year with retention period of 4 years. You reckon usenet could handle an order of magnitude more than it currently does?
IRC for "instant" communications among individuals and groups, and Usenet for public forums is why there is no need for Facebook, Twitter and other centrally-controlled systems so prone to censorship and similar abuses both by the commercial interests controlling them, and the governments able to twist the former's arms.
I loved Usenet. But I don't think it would ever scale to the quantity of data (basically, videos) that's posted to youtube or facebook or twitter.
it sounds like certain people don't like it when social media serves up a dish of politically inconvenient news and the plebs rightly get angry.
To be clear, this news story isn't about "plebs getting angry". It's about plebs illegally breaking into a refugee's house to set it on fire or otherwise damage it, and illegally assaulting refugees on the street or elsewhere.
The study shows anti-refugees folks committing illegal acts at a statistically higher rate above the norm.
No, I agree. I just feel like this list is leaving out important information. People think "big money" when they hear Apple, Google, Amazon, etc, and so a lot of people might see this and think they can get an $80-100k job after 3-4 weeks in a coding bootcamp but that's not going to happen. So that's why I asked "what kind of jobs are they really getting?"
Agreed. I'd like to make an apples-to-apples comparison:
* If someone was hired without a college degree, what are they making after 4 years in the workforce compared to a newly hired college graduate?
* If someone was hired without a college degree, what are they making after 14 years in the workforce compared to a college graduate with 10 years in the workforce? (to see if the college degree impacted long-term success)
* What are the comparative satisfaction levels of each?
I don't have mod points to upvote you, so instead I'll just write that your post was really clear and informative. Thanks. I like it when someone brings clear definitions to a debate. Or even better, like you did, when they bring clear explanations of the underlying concepts.
[I think what we've discovered is that the "capacity to actually think" is by and large unimportant for most of the needs we have]... Yeah? Who the hell is this 'we' you're referring to? Not anyone I've ever talked to. I think you're making that up, and the 'we' is actually just 'you'.
As you wrote, "that's all they've been obsessively focusing on". The "they" in your sentence are the ones who presumably think that deep learning is good enough for their purposes, else they wouldn't pouring their billions into it. I'm just describing to you what the market has perceived, which you yourself observed too.
James Damore got in trouble, because his memo said that women don't think the same way that men do.
That's only part of what he did, and not why he got fired. Damore wrote a memo with lots of logical gaps. The reader naturally had to fill in those gaps. Most readers didn't realize they were gaps and filled them in according to their own fears and preconceptions. That's why a lot of people filled in with their presumption that he was a nasty person making sexist arguments, and other people filled in with their presumption that he was a reasonable person making valid points. The different audiences were responding to literally different memos.
I'd hoped the reason for his firing was that his logical gaps showed a woolly mind that wasn't a natural fit for software development. But alas no. He was fired because the gappiness of his memo, in a contentious area, inevitably and needlessly created workplace hostility that also interrupted a Google high-up on his vacation.
I never understood why people think image recognition is new "AI". It isn't. License plate readers (and image recognition) have been around for decades.
Image recognition was indeed around for decades. It was based on convolutions for edge-detection, haar cascades for face recognition. You'll have used these if you had a camera with face detection up to a few years ago. If you've coded with OpenCV you'll have used these APIs, e.g. https://docs.opencv.org/3.4.1/...
It was an old technology that had run its course and really was a dead end. It wasn't making progress. It required too much custom human coding for things you wanted to recognize, and it was hit-or-miss.
When machine-learning neural networks came of age in the mid 2010s it was a game-changer. It recognized far more things in far more difficult images. Neural networks of course had been around for decades too. What made them come of age was (1) vast tagged training datasets, (2) fast enough hardware to run the training at scale.
That's the point it DOESN'T 'do the job plenty well enough' it always falls short of the mark because it has ZERO capacity to actually THINK, your dog or cat has better cognitive ability, and people will trust this half-assed excuse for AI too much and disasters will happen.
I think what we've discovered is that the "capacity to actually think" is by and large unimportant for most of the needs we have -- good-enough large scale image recognition, good-enough medical imaging assessments, and others that I listed.
Disasters? You'll have to spell out why you say the "ability to think" or general-purpose intelligence will lead to fewer disasters rather than more. Personally, I reckon it would lead to more disasters just through sheer complexity and unpredictability and un-debuggability. If a "thinking" industrial robot kills a human, how the heck will we debug that or fix it? but if a half-assed-excuse robot kills a human, we can assess whether its image-recognition failed, or some other part of it failed, and address it. Similarly for a machine that controls an industrial process.
I've said it before, I'll keep saying it: until we actually understand how a biological brain produces the phenomena we call 'thinking', we will not be able to create 'machine intelligences' that match or exceed human beings. Period. Now, what they should be investing 'billions and billions of dollars' in, is research and development of newer, better instrumentation for observing a living brain in action (and I do NOT mean 'a better fMRI, I mean invent something that's a new and different approach).
There's a HUGE market today for the kinds of things that the current machine-learning approach does plenty well enough at at looks like it has a lot more growth area - license-plate recognition, facial recognition, image recognition, medical imagining recognition in the case of Deep Mind. It looks like the billions that are being invested today are a good investment.
And you propose switching that investment to a speculative thing that might bear fruit in 50-100 years time, and if it did then the result would be a general-purpose intelligence that replaces a lowly-paid human being? Why should someone invest billions in that?
No, C. P. snow didn't say that. He said that science types were lamentably under-exposed to literature and particularly the arts, and that arts types were even more ignorant of science. Neither side came off positively.
When the median house price in San Francisco is over $1.6m, using the 1/3rds rule, you need to be making $533k in salary to realistically own a home there.
Your numbers don't add up. That's because a first-time buyer (without assets) buys a starter home, not a median home. Let's flesh it out with solid scenarios:
(1) First time buyer, has rented for 5 years to build up downpayment, is now looking at a starter home. Studio rent is $2k/month in San Jose, so if you earned $100k/year after tax then you'd be paying 24% of your salary on rent, and being frugal you're saving up 25% of it. That will make $125k downpayment at the time you buy your house. Median starter home price is $450k in Oakland, $650k in San Jose, $800k in SF. Let's stay in San Jose, so you need a mortgage of $525k. At 4% interest for a 30-year fixed, you'll be paying $2.5k/month, adding up to 30% of your salary.
(2) Ten-year home owner, has accrued $400k in equity, is now looking at a mid-range home. Median house price is $1.1M in San Jose, so needs a mortgage of $600k or $2.8k/month. I trust that you're earning more money now, say $130k/year after tax. That works out at 25% of your salary.
That's a very questionable assumption, because if you are encoding a 48MP JPEG rather than a 24MP JPEG, you are still forcing the format to store information for twice as many pixels, whether or not there is meaningful detail in those pixels.
I don't know about jpeg compression in particular. But if compression is done with gabor wavelets then it doesn't need to store information for twice as many pixels. Nor fractal compression. And if you encode audio in the frequency domain then you don't need to store information for twice as many samples.
Yes, it's compressed, but just go ahead and do this experiment for yourself - an aggressively compressed, high-resolution image will probably display noticeable compression artifacts, while a barely compressed image at half the resolution will probably look just as good as the original, for realistic viewing conditions.
I did do the experiment myself, which is why I'm posting! I digitized all my old photo negatives. My target filesize was 3MB/photo. I spent a day just experimenting with compression, and determined that 9MB/photo with 90% compression to achieve 3MB, gave a better output than 4MB/photo with 50% compression to achieve the same 3MB filesize. Of course if you zoom in a lot in the 9MB/photo images then you see compression artifacts, but then if you zoom in the same amount to the 4MB/photo images then you just see blocky pixels. When you zoom out enough that compression artifacts are no longer obvious on the 9MB images, well, you still see the blocky pixels on the 4MB images.
As to whether my results would scale up to 48MB vs 24MB? that's hard to predict. I can see it going either way.
We don't need 48 MP taking up space on our phones and hard drives.
I don't think you've got the right metric there either. The right metric should be: "If I jpeg-encode a 48MP image to X number of bytes, vs jpeg-encode a 24MP image to the same number of bytes, which one produces the best image?" (for various values of X, I guess from about 0.5mb to 8mb)
I'd assume the 48MP one will produce the best image -- because we haven't pre-emptively thrown away information at the sensor stage, and so we've left it to the encoder to decide which information to throw away to achieve that filesize, and the software can make a better decision about what to throw away.
As a Seattle person and Amazon prime member, I'd say that 95% of my recycling is just Amazon cardboard box packaging. I see that the "transfer station" (where you take rubbish yourself) has a dedicated machine just for squishing cardboard boxes, so I assume they get recycled from there. As for whether cardboard gets recycled when I leave it in the unsorted recycling bin? or when big cardboard boxes are stacked next to the recycling bin for them to take away? I'd love to know what happens to them.
The issue is having the same source tree always build an identical cdrom.
Reproducible builds are hugely important in industry, but I think that producing an identical cdrom is the least of the benefits. The biggest benefit is build and test time. Imagine if a complete build of the entirety of Windows (all different platforms, all different versions) took 20+ hours on the build servers. But if each build step is a pure function from inputs to outputs, and you're 100% sure that ephemera mentioned in the article like timestamps, timezones, build order, paths on disk are absent, then suddenly it becomes correct to CACHE intermediate results. I'd expect to see at least 10x improvement in build+test times across a large organization, maybe 100x. This performance increase opens the door to being able to test within hours whether a change is safe for the product as a whole, rather than having to wait a day.
(I guess this wouldn't be an advantage in a mythical ideal world where every single component is precisely specced and fully unit-tested, and if something passes its unit-tests then you can blindly trust that it will have no unintended consequences upon the product as a whole.)
I don't think hyper loops are real-world feasible. Even if the technology works, any aggrieved destructive fool - and these exist everywhere in the world, China included - can put the entire system at risk in a way that aircraft are not threatened by. It's easier to guard an airport in such a way that man-portable missiles are out of range of aircraft taking off/landing, than it is to guard the entire length of some long-distance piped network that basically needs to maintain vacuum sealing in its entirety.
Did you just disprove the existence of oil pipelines?
That should never happen, because you communicate this to the waitperson every time, right?
If you can suggest a form of words and a manner of explaining this that wouldn't irritate the heck out of the waitperson, and wouldn't make my fellow dinners squirm (who share my preferences but don't want to violate the social contract), then please suggest one! I don't believe one exists.
Slashdot Mirror
My impression is that for every language other than C/C++, there are three things more important than almost any feature of the language or runtime:
(1) the package system - how you search+download+reference them, how dependencies+conflicts are managed, how updates are managed
(2) the build system
(3) editor integration, be it Emacs or Sublime or VSCode
Are you telling me that struggles with climate change, scarcity of natural resources and so on will just vanish into thin air by embracing Socialism or perhaps a Fascist dictatorship, or Communist rule?
No one is telling you that (save for the click-baity slashdot headline). The article itself boils down to this:
1. Economies are changing to less energy-efficient energy sources for the first time in history (I assume that means wind+solar+hydro), which will disrupt things. The economic models used by governments and industry have traditionally assumed energy abundance, which is becoming a poor assumption.
2. Sink costs are rising - economic growth used to be straightforward with profligate use of energy, use of materials, creation of waste, will become less possible.
3. The models are fine at handling slow incremental changes, but we'll likely see more dramatic changes than they'll be good at handling.
4. Carbon pricing hasn't worked very well and doesn't seem likely to work in future.
5. Post-Keynesian models about the relationship between the state and the free market will give different predictions and metrics-for-success compared to neoclassical economic models. All of them still boil down to free markets with government tweaks+nudges of course.
6. The article sort of tails off into more or less wishing the whole world were like Finland - not surprising given the authors and the paper are from Finland.
[usenet daily volume of 27 terrabytes/day] Did I read that correctly? I thought Usenet was pretty much dead. Many very popular newsgroups, like rec.sports.baseball, that received 1000+ posts/day back in the day were completely empty as of a couple years ago.
I got the number from https://en.wikipedia.org/wiki/... -
Much of this traffic increase reflects not an increase in discrete users or newsgroup discussions, but instead the combination of massive automated spamming and an increase in the use of .binaries newsgroups
[I don't think it would ever scale to the quantity of data (basically, videos)]
It already has and does. I have yet to see full 50GB+ BlueRay rips posted to FB.
The upload rate to Youtube I saw for last year was 270 terrabytes/day with retention period of forever, compared to usenet daily volume of 27 terrabytes/day last year with retention period of 4 years. You reckon usenet could handle an order of magnitude more than it currently does?
IRC for "instant" communications among individuals and groups, and Usenet for public forums is why there is no need for Facebook, Twitter and other centrally-controlled systems so prone to censorship and similar abuses both by the commercial interests controlling them, and the governments able to twist the former's arms.
I loved Usenet. But I don't think it would ever scale to the quantity of data (basically, videos) that's posted to youtube or facebook or twitter.
it sounds like certain people don't like it when social media serves up a dish of politically inconvenient news and the plebs rightly get angry.
To be clear, this news story isn't about "plebs getting angry". It's about plebs illegally breaking into a refugee's house to set it on fire or otherwise damage it, and illegally assaulting refugees on the street or elsewhere.
The study shows anti-refugees folks committing illegal acts at a statistically higher rate above the norm.
No, I agree. I just feel like this list is leaving out important information. People think "big money" when they hear Apple, Google, Amazon, etc, and so a lot of people might see this and think they can get an $80-100k job after 3-4 weeks in a coding bootcamp but that's not going to happen. So that's why I asked "what kind of jobs are they really getting?"
Agreed. I'd like to make an apples-to-apples comparison:
I don't have mod points to upvote you, so instead I'll just write that your post was really clear and informative. Thanks. I like it when someone brings clear definitions to a debate. Or even better, like you did, when they bring clear explanations of the underlying concepts.
The US has a lot of subcultures. The ones you are dissing for low life expectancy are likely not the ones that you think you are dissing.
I honestly don't have a clue what you mean here!
Cuba is down to 33rd in the world in life expectancy.
That's terrible - only one notch above the USA, at 34th in the world!
[I think what we've discovered is that the "capacity to actually think" is by and large unimportant for most of the needs we have] ... Yeah? Who the hell is this 'we' you're referring to? Not anyone I've ever talked to. I think you're making that up, and the 'we' is actually just 'you'.
As you wrote, "that's all they've been obsessively focusing on". The "they" in your sentence are the ones who presumably think that deep learning is good enough for their purposes, else they wouldn't pouring their billions into it. I'm just describing to you what the market has perceived, which you yourself observed too.
James Damore got in trouble, because his memo said that women don't think the same way that men do.
That's only part of what he did, and not why he got fired. Damore wrote a memo with lots of logical gaps. The reader naturally had to fill in those gaps. Most readers didn't realize they were gaps and filled them in according to their own fears and preconceptions. That's why a lot of people filled in with their presumption that he was a nasty person making sexist arguments, and other people filled in with their presumption that he was a reasonable person making valid points. The different audiences were responding to literally different memos.
I'd hoped the reason for his firing was that his logical gaps showed a woolly mind that wasn't a natural fit for software development. But alas no. He was fired because the gappiness of his memo, in a contentious area, inevitably and needlessly created workplace hostility that also interrupted a Google high-up on his vacation.
I never understood why people think image recognition is new "AI". It isn't. License plate readers (and image recognition) have been around for decades.
Image recognition was indeed around for decades. It was based on convolutions for edge-detection, haar cascades for face recognition. You'll have used these if you had a camera with face detection up to a few years ago. If you've coded with OpenCV you'll have used these APIs, e.g. https://docs.opencv.org/3.4.1/...
It was an old technology that had run its course and really was a dead end. It wasn't making progress. It required too much custom human coding for things you wanted to recognize, and it was hit-or-miss.
When machine-learning neural networks came of age in the mid 2010s it was a game-changer. It recognized far more things in far more difficult images. Neural networks of course had been around for decades too. What made them come of age was (1) vast tagged training datasets, (2) fast enough hardware to run the training at scale.
That's the point it DOESN'T 'do the job plenty well enough' it always falls short of the mark because it has ZERO capacity to actually THINK, your dog or cat has better cognitive ability, and people will trust this half-assed excuse for AI too much and disasters will happen.
I think what we've discovered is that the "capacity to actually think" is by and large unimportant for most of the needs we have -- good-enough large scale image recognition, good-enough medical imaging assessments, and others that I listed.
Disasters? You'll have to spell out why you say the "ability to think" or general-purpose intelligence will lead to fewer disasters rather than more. Personally, I reckon it would lead to more disasters just through sheer complexity and unpredictability and un-debuggability. If a "thinking" industrial robot kills a human, how the heck will we debug that or fix it? but if a half-assed-excuse robot kills a human, we can assess whether its image-recognition failed, or some other part of it failed, and address it. Similarly for a machine that controls an industrial process.
I've said it before, I'll keep saying it: until we actually understand how a biological brain produces the phenomena we call 'thinking', we will not be able to create 'machine intelligences' that match or exceed human beings. Period. Now, what they should be investing 'billions and billions of dollars' in, is research and development of newer, better instrumentation for observing a living brain in action (and I do NOT mean 'a better fMRI, I mean invent something that's a new and different approach).
There's a HUGE market today for the kinds of things that the current machine-learning approach does plenty well enough at at looks like it has a lot more growth area - license-plate recognition, facial recognition, image recognition, medical imagining recognition in the case of Deep Mind. It looks like the billions that are being invested today are a good investment.
And you propose switching that investment to a speculative thing that might bear fruit in 50-100 years time, and if it did then the result would be a general-purpose intelligence that replaces a lowly-paid human being? Why should someone invest billions in that?
Shooting the guy who entered while she was dressing would have helped educate him and others for the future...
Problem: company has a policy that I don't like. Solution: injure/kill the low-paid employee who's carrying out that company policy. ???
Shrug. I work at Facebook, am pretty conservative in many respects, and have felt entirely welcome in the two years I've been here.
No, C. P. snow didn't say that. He said that science types were lamentably under-exposed to literature and particularly the arts, and that arts types were even more ignorant of science. Neither side came off positively.
When the median house price in San Francisco is over $1.6m, using the 1/3rds rule, you need to be making $533k in salary to realistically own a home there.
Your numbers don't add up. That's because a first-time buyer (without assets) buys a starter home, not a median home. Let's flesh it out with solid scenarios:
(1) First time buyer, has rented for 5 years to build up downpayment, is now looking at a starter home. Studio rent is $2k/month in San Jose, so if you earned $100k/year after tax then you'd be paying 24% of your salary on rent, and being frugal you're saving up 25% of it. That will make $125k downpayment at the time you buy your house. Median starter home price is $450k in Oakland, $650k in San Jose, $800k in SF. Let's stay in San Jose, so you need a mortgage of $525k. At 4% interest for a 30-year fixed, you'll be paying $2.5k/month, adding up to 30% of your salary.
(2) Ten-year home owner, has accrued $400k in equity, is now looking at a mid-range home. Median house price is $1.1M in San Jose, so needs a mortgage of $600k or $2.8k/month. I trust that you're earning more money now, say $130k/year after tax. That works out at 25% of your salary.
That's a very questionable assumption, because if you are encoding a 48MP JPEG rather than a 24MP JPEG, you are still forcing the format to store information for twice as many pixels, whether or not there is meaningful detail in those pixels.
I don't know about jpeg compression in particular. But if compression is done with gabor wavelets then it doesn't need to store information for twice as many pixels. Nor fractal compression. And if you encode audio in the frequency domain then you don't need to store information for twice as many samples.
Yes, it's compressed, but just go ahead and do this experiment for yourself - an aggressively compressed, high-resolution image will probably display noticeable compression artifacts, while a barely compressed image at half the resolution will probably look just as good as the original, for realistic viewing conditions.
I did do the experiment myself, which is why I'm posting! I digitized all my old photo negatives. My target filesize was 3MB/photo. I spent a day just experimenting with compression, and determined that 9MB/photo with 90% compression to achieve 3MB, gave a better output than 4MB/photo with 50% compression to achieve the same 3MB filesize. Of course if you zoom in a lot in the 9MB/photo images then you see compression artifacts, but then if you zoom in the same amount to the 4MB/photo images then you just see blocky pixels. When you zoom out enough that compression artifacts are no longer obvious on the 9MB images, well, you still see the blocky pixels on the 4MB images.
As to whether my results would scale up to 48MB vs 24MB? that's hard to predict. I can see it going either way.
We don't need 48 MP taking up space on our phones and hard drives.
I don't think you've got the right metric there either. The right metric should be: "If I jpeg-encode a 48MP image to X number of bytes, vs jpeg-encode a 24MP image to the same number of bytes, which one produces the best image?" (for various values of X, I guess from about 0.5mb to 8mb)
I'd assume the 48MP one will produce the best image -- because we haven't pre-emptively thrown away information at the sensor stage, and so we've left it to the encoder to decide which information to throw away to achieve that filesize, and the software can make a better decision about what to throw away.
Seattle sends its unsorted recycling to China.
As a Seattle person and Amazon prime member, I'd say that 95% of my recycling is just Amazon cardboard box packaging. I see that the "transfer station" (where you take rubbish yourself) has a dedicated machine just for squishing cardboard boxes, so I assume they get recycled from there. As for whether cardboard gets recycled when I leave it in the unsorted recycling bin? or when big cardboard boxes are stacked next to the recycling bin for them to take away? I'd love to know what happens to them.
The issue is having the same source tree always build an identical cdrom.
Reproducible builds are hugely important in industry, but I think that producing an identical cdrom is the least of the benefits. The biggest benefit is build and test time. Imagine if a complete build of the entirety of Windows (all different platforms, all different versions) took 20+ hours on the build servers. But if each build step is a pure function from inputs to outputs, and you're 100% sure that ephemera mentioned in the article like timestamps, timezones, build order, paths on disk are absent, then suddenly it becomes correct to CACHE intermediate results. I'd expect to see at least 10x improvement in build+test times across a large organization, maybe 100x. This performance increase opens the door to being able to test within hours whether a change is safe for the product as a whole, rather than having to wait a day.
(I guess this wouldn't be an advantage in a mythical ideal world where every single component is precisely specced and fully unit-tested, and if something passes its unit-tests then you can blindly trust that it will have no unintended consequences upon the product as a whole.)
I don't think hyper loops are real-world feasible. Even if the technology works, any aggrieved destructive fool - and these exist everywhere in the world, China included - can put the entire system at risk in a way that aircraft are not threatened by. It's easier to guard an airport in such a way that man-portable missiles are out of range of aircraft taking off/landing, than it is to guard the entire length of some long-distance piped network that basically needs to maintain vacuum sealing in its entirety.
Did you just disprove the existence of oil pipelines?
That should never happen, because you communicate this to the waitperson every time, right?
If you can suggest a form of words and a manner of explaining this that wouldn't irritate the heck out of the waitperson, and wouldn't make my fellow dinners squirm (who share my preferences but don't want to violate the social contract), then please suggest one! I don't believe one exists.