U.S. government limits exports of artificial intelligence software
reuters.com
reuters.com
Obviously, it failed! You can stop a single vendor with unique technology to provide hardware components with bans, but you can't stop a whole field spreading knowledge!!! People come and go, meet, talk, write, exchange knowledges... so sooner or later (and more soon now as long as the internet is not limited to US) the software will implement these ideas.
The only one that will be challenged will be "big corps" relying on IP protection. But if i remember correctly, Google has a research center in China... so knowledge will aleady be in China and won't even need to be "exported"
Thomas Sowell wrote a quite a few good books about this.
I have no doubt it applies equally to export laws.
The wiki link you provided does not have the words "export" or "espionage". Are you trying to imply that R,S, and A were suspected of espionage? You'll need a real source for that..
Not true since java 8u161, and similarly for 7 and 6. And not true for every java > 8.
https://www.oracle.com/technetwork/java/javase/downloads/jce...
http://www.oracle.com/technetwork/java/javase/8u161-relnotes...
Since 8u161, 7u171, and 6u16 you don't need to do that.
> Here at the Bouncy Castle, we believe in encryption. That's something that's near and dear to our hearts. We believe so strongly in encryption, that we've gone to the effort to provide some for everybody, and we've now been doing it for almost 20 years!
So, he printed the code in a book to be OCR’d and further compiled into the software.
No export controls on a book.
Feedback and control systems can be described in a couple pages of text and a handful of simple equations. Suppressing knowledge is never an effective strategy. We shouldn't hoard the Krabby Patty Secret Formula.
There was also the RSA t-shirt. The originals are no longer being sold, unfortunately, but these days it's easy enough to do a custom design for your own made to order. My take: https://www.customink.com/ndx/?cid=jxu0-00bx-9p0k
> The Ninth Circuit ordered that this case be reheard by the en banc court, and withdrew the three-judge panel opinion ... almost nine years after Bernstein first brought the case, the judge dismissed it and asked Bernstein to come back when the government made a "concrete threat"
Is it foolish that the US government continues to try to restrict the spread of nuclear engineering?
A parallel reasoning would be with face recognition: the algorithms are quite widespread (the knowledge) and new algorithms build on shared knowledge of the field, but the tuning of the parameters, of the input datas, the access to top hardware to speed up learning process, evaluate new strategies, and access to big databases are not.
The problem for a state to get the face recognition technology is either to obtain the parameters or to invest time and money to build it. So it's not really "bannable" because theses resources are available everywhere (and lots of companies are ready to provide it).
A more "bannable" tech would be "quantum computer" because it relies on new engineering tech, not only on knwoledge.
-- cut --
Geospatial imagery “software” “specially designed” for training a Deep Convolutional Neural Network to automate the analysis of geospatial imagery and point clouds, and having all of the following:
1. Provides a graphical user interface that enables the user to identify objects (e.g., vehicles, houses, etc.) from within geospatial imagery and point clouds in order to extract positive and negative samples of an object of interest;
2. Reduces pixel variation by performing scale, color, and rotational normalization on the positive samples;
3. Trains a Deep Convolutional Neural Network to detect the object of interest from the positive and negative samples; and
4. Identifies objects in geospatial imagery using the trained Deep Convolutional Neural Network by matching the rotational pattern from the positive samples with the rotational pattern of objects in the geospatial imagery.
Technical Note: A point cloud is a collection of data points defined by a given coordinate system. A point cloud is also known as a digital surface model.
-- cut --
[1] https://www.federalregister.gov/documents/2020/01/06/2019-27...
[2] https://s3.amazonaws.com/public-inspection.federalregister.g...
"(geospatial imagery) and (point clouds)" or "geospatial (imagery and point clouds)"?
Point 4 requires the use of geospatial imagery, so any point-cloud-only product would be exempt, it seems?
The document doesn't define "geospatial imagery", but that could surely include hobby and commercial drone footage of the ground. Perhaps even ordinary photography from security cameras that have user-define object identification features? That would make it really quite broad.
But all we need to do is not normalize color, and then everything's exempt!
But it's still confusing here. If they mean it applies to both imagery software and point cloud software, then point cloud software would be excluded in point 2 because it doesn't have pixels or color (if lidar/etc). So it must be software that uses both. That makes more sense if it's aimed at a specific existing product.
1) wants to acquire a geospatial imagery recognition program;
2) does not have the tech themselves;
3) can import the underlying tech without interface;
I highly doubt a restriction on GUI export will stop them.
How will anyone operate a computer without GUI? That’s impossible! /s
You can't get a contract without marketing/sales interaction with the customers.
Customers at a high enough level to sign off on a payment with 7-9 zeroes following the number are not generally programmers (or, if they were, they haven't been working at the coal face for decades): they're senior managers or civil servants.
A GUI front end is a really amazing marketing tool for any piece of software insofar as GUIs are designed to expose all the internal configuration variables and controls in a visually appealing, or even intuitive, manner that is accessible to non-programmers. Like the folks signing the big checks.
(Here's a second possibility: we know it's possible to use CAPTCHAs to crowdsource recognition of objects. Maybe they're trying to prevent export of a NN training system that uses unwitting mechanical turks for quality control?)
BTW, off topic, but I love your books!
If you think about tagging it is something you need a GUI for unless you are lucky enough to have pre-tagged samples.
Eg: the famous Silicon Valley hotdog AI. You either need to have a GUI to allow users to select hot dogs and not hot dogs in a bunch of images, or you need a bunch of images already tagged.
Edit: I overlooked “have all of the following”.
Although saying "well technically" as you get dragged off for waterboarding may not make you feel better.
In about 5 minutes I could find these:
https://github.com/search?q=geospatial+deep+learning
https://github.com/search?q=satellite+images+deep+learning
If it is considered exporting, is github the exporter, or is the developer. Just like a company might produce a metal widget but another company procures and exports it, the original company that made the widget isn't the exporter.
Evidently to the dinosaurs in Congress, a physical book is something COMPLETELY DIFFERENT than a file online.
Probably there is some precedent by considering e.g. if Lockheed or FedEx is the exporting company if there have been cases where weapons got exported to unwanted actors. Probably github is like FedEx in these cases.
Yeah almost nothing will fall under that. Build your own GUI (if you ever need one).
That will place the US in a large economical disadvantage for those.
https://law.lis.virginia.gov/vacode/title54.1/chapter42/sect...
[1] https://www2.census.gov/library/publications/decennial/1990/...
These restrictions seem to flow from a mental model that still views software as a product purchased in a shrink-wrapped box. Rather than the services based model currently extant.
Whether frameworks are at risk is limited to the wording of the ban and then the final determination comes from a judge hearing the case. The chilling effects are real and its possible framework development may very well be hampered due to the unknowns you have pointed out in your post
No.
They look like they are targeting a specific piece of commercial software.
I think the same concept applies here. Software is the machine that embodies the algorithm. Its tied to a company and to dollars.
What this will also mean is that startups that work in this space will need to watch out who they get funding from. If the VC is not US-based, CFIUS oversight may kick in.
Also, cryptography requires both sides to use the same algorithm, while companies don't need to use the same recognition algorithms.
It also helped, in the crypto case, that you could print some version of it on a T-shirt or mail it on a postcard. It looked like speech, while neural net parameters don't.
So the free speech case seems much weaker.
But Bernstein's case didn't hinge on the likely consequences of strong cryptography being widely available; rather, he argued that he had a First Amendment right to publish his research.
Sorry, I don't understand your argument.
This should not be true though, that's the point of having the law being a separate power.
AI code is just another layer in this odorous process.
These companies have "strict" code review policies but often the reviewers are just a recent previous year's new college grad, now overconfident by a small amount of work experience.
Assuming that among 1.4 billion people there are a few good coder/statistician people and using supercomputers [0] as a rough proxy for available computing power, it isn't obvious the US is even going to inconvenience the Chinese military. Presumably they are going to have a parallel AI effort anyway given that they have been investing in the area.
Chinese convnet research is absolutely state of the art already, indeed.
There was a time when a software engineer didn't get basic stuff. A time when languages like C were developed where there wasn't an associative data structure baked in for example.
It wasn't because associative data structures are a secret tech that requires great insight to uncover. It is because the field was new and people hadn't cottoned on to how basic and important having access to hash-maps is. Times moved on. Now basically all modern languages have hash-maps as a basic data type.
'AI' is in that early phase where the engineering world is still getting excited over stuff that will basic practice eventually. BERT and GTP2 are signs of how much computer power Google/etc's researchers have access to, not signs that the architectures are fundamentally complicated or somehow hard to work out if you live in China. AlphaGo for instance was breathtaking as a standalone project, but not hard to implement.
I have a belief that someone in the USSR worked out a way of doing fluid dynamics that has enabled the Russians to develop hypersonics and super cavitation. This is probably rather straightforward - in the style of NS - of you know the principles. No one in the West ( or China) knows those principles, so Western torpedos and reentry vehicles are rather poor Vs Russian ones. Once you grasp how something works the fact that it's rather easy to apply in comparison to the process of getting the insight shouldn't detract from the value of the insight .
The surprise was that in the mid-2000s suddenly GPU became so powerful that LSTM could be used to achieve interesting results. The story here isn't the models, it is the computers running the models.
That is a really weird historical fantasy. If you pull out your copy of volume 3 of Knuth and look at chapter 6, it is obvious that associative data structures were some of the first ones to be developed in the field.
The reason why hash tables became so popular is the explosion in main memory size starting in the late 1990s. The trade-offs between the possible associative data structures became less important for a lot of applications, especially when you consider how much needed to be done on secondary storage and specifically on tapes in the 1960s through the 1980s.
Also, how effective could this regulation be with so much knowledge and open source code already disseminated in the field?
1. Siri / Alexa and similar for voice recognition and doing basic tasks.
2. Face recognition for uploaded pictures on Facebook.
3. Lots of people use FaceID on iPhones.
4. Tesla and other SDC systems: yes, it's not good enough for general use, but the fact that it mostly works in California is pretty cool..
"Mindbending" is subjective, but these all use DNNs, and are used by millions to billions of people every day. So it's incorrect to suggest that all DNN use-cases could be replaced with "doing basic math with a multivariate normal distribution".
You want mindbending and scary? Mindbending: deepfakes. Scary: automated ai-based law enforcement.
This is Fear based Decision Making 101. All it leads to is more absurd outcomes such as Endless Wars, Huge Monopolies, More consolidation of power and resources in the hand of few therefore more inequality.
These people and this thinking style would have more credibility if they had stopped Wars, reduced inequality, disrupted monopoly and oligopolies. They have not done that.
The can't imagine a Chinese AI team and American AI team working together to solve problems in humane way. They can't imagine constructing orgs that push that through. They can't imagine punishing their own who cross lines out of fear that the other side wont.
When we allow Fear based thinking to dominate decision making Imagination dies. Outcomes are consistently shit. And way below the potential of what people collaborating and communicating across artificial bullshit boundaries are capable off.
Pick a side and don't back Fear based Decsion makers in your org. These guys hold back progress, are the reason climate change research is hidden, endless Wars keep getting funded and monopolies cling to power way past their expiry date.
How can it be the age of information and knowledge when fear wins?
This is going to persist regardless of this decision
Wonder what the real world impact of this will be. Not much, I expect.
Here's a thought experiment I use to imagine the impact of AI:
Imagine you've got a million people at your disposal. At zero cost and with no downtime, these people can remotely operate robots, understand text, interact using natural language, or classify objects in images, all with human-level intelligence and accuracy. Now what?
Obviously there are areas where AI can outperform humans, like mechanical accuracy and mathematical computation. But in general, I find this experiment works pretty well.
Perhaps a post-Singularity AI will have wild capabilities beyond our comprehension, but that is outside the scope of this experiment.
All other countries probably have smart researchers and engineers, but no one has the data machine that Google has...
I'm based in Australia, and we started playing today with analysis of geospatial imagery for bush fire imagery (because the country is on fire).
ECCN0D521No.1 from https://s3.amazonaws.com/public-inspection.federalregister.g... covers pretty much everything I'm doing.
I guess if I want to look at the positive side, it means I don't have to compete with any US vendors if I want to sell my work.
Until the Australian government gets the same idea
"The rule will likely be welcomed by industry, Lewis said, because it had feared a much broader crackdown on exports of most artificial intelligence hardware and software"
Or is it that training data provides experience, which improves the algorithms? Or the application of the algorithms?
Not to mention nowadays Deep learning is pretty much a big data game.
Assuming it is - China is also competitive on developing algorithms. A few months ago there was a post on explosion of AI papers submitted by authors at Chinese research institutions, with no signs of slowing down.
Given two research labs, with one having a bit better equipment, while the other having a more proven history of publishing innovative research, it would be disingenuous to say that the better equipped lab is ahead until they have actually produced some research that puts them ahead. It might help them gain lead, but it also might result in nothing. Better equipment is just one of many components that affect the chances of success. Until that lead is acquired, I don't really think it would be appropriate to say that they had done so.
Note: the lab with a history of published innovative research in my analogy isn't supposed to represent the US or any country in specific. This was just an example to better illustrate the point I was making. The only thing that should matter for whether someone is ahead or not in this situation is the actual proof of being ahead, not "the opportunities that could lead to them being ahead". Otherwise, we should also start immediately trusting all those articles that pop up once every few months about how some random city is "about to become the next Silicon Valley, here are the reasons why".
Most of the time whatever interesting thing (posted to HN) originated in some Chinese startup/university.
But that's just my gut feel.
I highly recommend the Frontline documentary on AI in China.
https://www.pbs.org/wgbh/frontline/film/in-the-age-of-ai/
They see AI as the next industrial revolution and have decided to make sure they are at its forefront. And likely they will, which means that we’ll be the ones trying to import their tech.
The ResNet Project [1] of 2015. It was used as the core algorithm behind Google’s AlphaGo in 2017.
The 4 computer scientists behind the paper were Chinese nationals. They were all educated by the Chinese educational system, and got their PhD there (one guy was from Hong Kong). They worked at Microsoft at the time, so Microsoft paid them a salary for their work, but I think Microsoft benefited more from their research, as did the other Silicon Valley and American companies.
Three of them went to start or lead other Chinese unicorn companies, and one guy went to Facebook in Silicon Valley, so Facebook benefited here.
Page 4 and 9 have the technical definitions.
My off the cuff interpretation is that the rule would only cover convolutional neural nets that are trained to identify and determine the orientation of specific objects in geospatial imagery. If the neural net's input/output aren't wrapped in a GUI it sounds like they still might be OK to export without a license
2.Reduces pixel variation by performing scale, color, and rotational normalization on the positive samples;
3. Trains a Deep Convolutional Neural Network to detect the object of interest from the positive and negative samples; and
4. Identifies objects in geospatial imagery using the trained Deep Convolutional Neural Network by matching the rotational pattern from the positive samples with the rotational pattern of objects in the geospatial imagery.
What counts as "geospatial imagery"? Could this apply to any training UI for self-driving cars, maps, street view, etc.?
So you can be an American company with tons of "friends" in the EU, Asia, Latin American and now Africa, doing stuff (research, product) and you would just happen to buy/sell from/through these independent actors. Fiction-Google: “Oh but that's not us! It's Oogleg, a Swiss company! It's true 95% of our private shareholders also have shares in Oogleg, but that's only circumstancial, these are large funds you know... they actually have shares in 95% of businesses altogether through ETFs and mutual funds dilution. + some legalese blabla.”
There goes your protectionism, State governments! You'll get your import taxes for physical goods and on-prems services but overall, it certainly won't impede or even touch the thriving heads, the global leaders of the business world. Not anymore. That was in another time, before global networks.
And actually, we might think Fortune 100, perhaps 1,000; but in truth it's probably much more (cue 80% of GDP in the form of SMBs) because how do you enforce a restriction on remoting to contribute to some repo somewhere?
Note that this is true as of 2020, factually from a technical standpoint, but given a few decades and some generalized country-based firewalls (it's coming, in all likelihood) + convenient surveillance and you get all the means necessary to enforce such policies anew.
> Did you think that companies can't interact if they're not subsidiaries of the same organization?
Of course not :) I however wonder if defending anti-trust from a subsidiary strategy would work — at least in France, I'm pretty sure taking half your execs and hiring them in a subsidiary which you control will NOT get you past anti-trust regulation.
You might say "but it's legal!" and the judge will kindly ask you not to mock the court by disingenuously failing to address the case at hand — are you or are you not effectively in a monopoly, or cartel situation? Legal or not in terms of legal structure doesn't matter because antitrust is 'above' in the hierarchy of norms (so to speak, my law studies are really far away now, and I was more into public than private law).
Case in point though, shareholding is even legally restricted in some sectors (e.g. media, and that was a strong motivation for e.g. Facebook trying not to be filed as a media group, at least in the EU).
I have absolutely no idea how this would fly in the US. I bow to your expertise, here.
A good example, I think, will be the shareholding structure of Libra (if it ever comes to fruition), where many actors essentially hide their participation behind layers of companies, like some onion (there was a good infographic which you might google on the topic). It's legal, technically, but would it stand in front of a supreme court antitrust case?
As far as I know from history, even legal lines tend to become blurry in major antitrust cases because these are, by essence, out-of-bounds of 'normal' operation, they're fringe cases that sometimes requires a new ad hoc law to take where we want to (I seem to remember elements of Teddy Roosevelt's opposition with Rockefeller, details of the Bell system breakup too, but I'm really not sure. Here in the EU, it's really common —all things considered— to just make new law whenever the current letter fails to live up to the desired spirit).
Thank you for the remarks, I'll probably refrain from speaking about antitrust in the US until I have a better understanding of those.
And it seems common for A, B, C, ... to jointly govern Z, without antitrust problems, like (first thing I could find) the W3C.
https://en.wikipedia.org/wiki/North_American_Aerospace_Defen...
What prevents a Chinese company from opening a business in the US to get access to US technology? I mean...nothing...that's exactly what they do.
More seriously, computer vision is going to be important and it appears to be far less known than machine learning and has higher barriers to entry. I'd exchange a few introductory machine learning books for more good computer vision introductions.
Any suggestions on how to get started with computer vision?
While we patiently await for a HN user (or, let's be honest, one of the ancient cryptologist-lawyers who come out of the woodwork every time something like this happens and sue the government) to fix this by suing the government on free speech grounds, don't forget that git, mercurial, fossil, bazaar and more are all decentralized, can't actually be censored at scale, and can be effectively hosted and mirrored trivially.
I actually think it's a well-intentioned law, and it's not like it'll harm most people, but it's still something that should be stood against on principle.
> Under a new rule which goes into effect on Monday, companies that export certain types of geospatial imagery software from the United States must apply for a license to send it overseas except when it is being shipped to Canada.
"sensors, drones, and satellites" used to target anything means that you can't even send a Ring camera to Europe.
This seems to target one or a few products so they dislike that someone uses the software for that purpose
https://en.wikipedia.org/wiki/TERCOM
There was a big drone attack on a Saudi Arabian oil processing facility last year: https://en.wikipedia.org/wiki/2019_Abqaiq%E2%80%93Khurais_at...
Has that now changed?
Is publishing personal code on GitHub "exporting"?
https://www.bis.doc.gov/index.php/all-articles/107-about-bis...
I work on ITAR-regulated software and, even though, the software itself is exported all over the world, I would not be able to write it if I had been a national of a restricted country, working in the US on a temporary visa.
AI tech is too powerful to be monopolised, if not democratised it might become another 'semiconductor' industry.
I think economic is the keyword here. From what I gather this is not the first time the US is doing something like this. I am pretty sure other countries have done the same in their particular areas of concern. They're just not as mighty and famous as the US so nobody pays attention. So much for free market.
Anyways I think it is a little too late and all it will accomplish is - opening a window of opportunity for other players.
Also because it formulated way too broad and has an escape clause (apply for a license) then it might offer an unfair advantage right inside the US. Big companies will get it and for smaller it nay be more difficult. Same as patent system. Company like Apple can patent my cat with little troubles. Me: not so much and I speak from experience.
Anyone can slap their code onto a private Gitlab installation in an hour. Hosting a tarball on an HTTPS server is even more trivial.
Google suggests Nvidia GPUs are probably made in Guangdong by PCPartner.
The actual chips, I don't know, but TSMC does have fabs in China.
What financially or technologically significant exports are going to stop? How military or nonmilitary are they?
Companies don't sell stuff from GitHub, they sell proprietary stuff. It may well be based on open source code, but they own it to sell (license) it.
There's lots of examples of products in the geospatial domain that are not "AI" yet are restricted or even classified.
For example, ship detection from space-based radar. There are numerous public papers on the topic yet any software that purports to do this is subject to ITAR rules in the US and CGP rules in Canada.
Just because you may know how to do something doesn't prevent a government from restricting you from selling it, or talking about it. Even if it is "public". Machine guns are an old tech and yet are restricted. As they should be.