22,905 karma · joined September 1, 2011
The main phrase of the title "model collapse is happening" is untrue and not substantiated in the article - all the true statements in the article are about the hypothetical problem, warning of the bad consequences that would likely happen if makers of major models did something they aren't doing, but they aren't doing that because that is a known issue that they're avoiding. It's like writing an article "Foot shooting epidemic is happening" with a long, solid (and true!) proof that if you'll shoot yourself in the foot, it will indeed cause serious injury...
Businesses in Scandinavia and many other countries would not treat someone knowing your personal information as any evidence of identity (because it's not); having all that information is not sufficient to impersonate you there - identity theft does happen but it would require stealing or forging physical documents or actual credentials to things like bank accounts; knowing all of what your mother or spouse would know is not enough to e.g. get credit or get valuable goods in your name.
If it's not (fully) your code, you aren't free to set the licence conditions; Linus can't do that without getting approval from 100% (not 99% or so) of authors who contributed code.
What one can do is add an informative disclaimer saying "To the best of our knowledge, installing or running this thing in California is prohibited - we permit to do whatever you want with it, but how you'll comply with that law is your business".
Like, in general, a software change to add an "age class" attribute to user accounts and a syscall "what's this attribute for the current user account" would satisfy the California bill and that's a relatively minor change (the bad part is the NY bill that allegedly requires technical verification of whatever the user claimed).
The weird issue is how should that attribute be filled for the 'root' or 'www-data' user of a linux machine I have on the cloud. Or, to put aside open source for that matter, the Administrator account on a Windows Active Directory system.
Because "user accounts" don't necessarily have any mapping (much less a 1-to-1 mapping) to a person; many user accounts are personal but many are not.
For a major example, treating a network resource like a file is neat and elegant and simple while the network works well, however, once you have unreliable or slow or intermittent connectivity, the abstraction breaks and you have to handle the fact that it's not really like a local file, and your elegant abstraction has to be mangled with all kinds of things so that your apps are able to do that.
That being said, it's not 100% used everywhere yet (Wikipedia mentions 92.6% of websites), and various means of tricking devices into downgrading to an older protocol would result in traffic that might be decrypted later.
Thank you for your testing! That's a bunch of effort which I didn't do - but checking the other claim is much more difficult; a refusal is clearly visible, but saying whether out of two different codebases one is systematically slightly less secure is quite tricky - so that's why people are complaining about the lack of any description of the methodology of how they measure that, without which the claims actually are not testable.
The same applies to web and blogs; the ability to monetize them by ads (and I do remember the "old web" before it was the case) increased the content but drowned out viewership for the true enthusiasts running things in their spare time, which IMHO were more valuable and I think that regime was better; again, losing 90% or 99% of the content wouldn't be bad in my mind, there still would be more than enough for anyone to ever "consume".
Most companies have no relation between their code and their products at all - a major food conglomerate will have hundreds or thousands of IT personnel and no direct link between defects in their business process automation code (which is the #1 employment of developers) and the quality of their products.
For companies where the product does have some tech component (e.g. refrigerators mentioned above) again, I'd bet that most of that companies developers don't work on any code that's intended to be in the product, in such a company there simply is far more programming work outside of that product than inside of one. The companies making a software-first product (like startups on hackernews) where a software defect implies a product defect are an exception, not the mainstream.
If a jury of your peers decides that you configured that firmware to protect your data from thieves, then it is not a problem.
But if a jury of your peers doesn't buy your assertion of that, and instead becomes convinced that you applied that (otherwise reasonable and legitimate) feature with the intent to destroy evidence of a crime so that it doesn't reach the court, then setting it up was a felony.
To me, namedtuples are a convenience to give a nicer syntax than ordinary tuples in scenarios where I don't want the overhead of having to store a copy of all the keys with every object, like a dict would. Dataclass seems to be even more stuff on top of a class which is effectively even more stuff on top of a dict, but all the use cases of namedtuples are those where you want much less stuff than an ordinary class has. And I don't want to have to define a custom class just as I often don't define a custom namedtuple in my code but use the one the database driver generates based on the query, which is a very common use case for namedtuples as efficient temporary storage of data that then gets processed to something else.
So if the first step isn't near-perfect (which ASR isn't) and if there is some information or "world knowledge" in the later step(s) which is helpful in deciding that (which is true with respect to knowledge about named entities and ASR) then you can get better accuracy by having an end-to-end system where you don't attempt to pick just one best option at the system boundary. Also, joint training can be helpful, but that IMHO is less important.