Because few people know what’s realistic for LLMs

Humans learn a lot through repetition, no reason to believe that LLMs wouldn’t benefit from reinforcement of higher quality information. Especially because seeing the same information in different contexts helps mapping the links between the different contexts and helps dispel incorrect assumptions. But like I said, the only viable method they have for this kind of emphasis at scale is incidental replication of more popular works in its samples. And when something is duplicated too much it overfits instead.

They need to fundamentally change big parts of how learning happens and how the algorithm learns to fix this conflict. In particular it will need a lot more “introspective” training stages to refine what it has learned, and pretty much nobody does anything even slightly similar on large models because they don’t know how, and it would be insanely expensive anyway.

Yes, but should big companies with business models designed to be exploitative be allowed to act hypocritically?

My problem isn’t with ML as such, or with learning over such large sets of works, etc, but these companies are designing their services specifically to push the people who’s works they rely on out of work.

The irony of overfitting is that both having numerous copies of common works is a problem AND removing the duplicates would be a problem. They need an understanding of what’s representative for language, etc, but the training algorithms can’t learn that on their own and it’s not feasible go have humans teach it that and also the training algorithm can’t effectively detect duplicates and “tune down” their influence to stop replicating them exactly. Also, trying to do that latter thing algorithmically will ALSO break things as it would break its understanding of stuff like standard legalese and boilerplate language, etc.

The current generation of generative ML doesn’t do what it says on the box, AND the companies running them deserve to get screwed over.

And yes I understand the risk of screwing up fair use, which is why my suggestion is not to hinder learning, but to require the companies to track copyright status of samples and inform ends users of licensing status when the system detects a sample is substantially replicated in the output. This will not hurt anybody training on public domain or fairly licensed works, nor hurt anybody who tracks authorship when crawling for samples, and will also not hurt anybody who has designed their ML system to be sufficiently transformative that it never replicates copyrighted samples. It just hurts exploitative companies.

Remember when media companies tried to sue switch manufacturers because their routers held copies of packets in RAM and argued they needed licensing for that?

https://www.eff.org/deeplinks/2006/06/yes-slashdotters-sira-really-bad

Training an AI can end up leaving copies of copyrightable segments of the originals, look up sample recover attacks. If it had worked as advertised then it would be transformative derivative works with fair use protection, but in reality it often doesn’t work that way

See also

https://curia.europa.eu/juris/liste.jsf?nat=or&mat=or&pcs=Oor&jur=C%2CT%2CF&for=&jge=&dates=&language=en&pro=&cit=none%252CC%252CCJ%252CR%252C2008E%252C%252C%252C%252C%252C%252C%252C%252C%252C%252Ctrue%252Cfalse%252Cfalse&oqp=&td=%3BALL&avg=&lgrec=en&parties=Football%2BAssociation%2BPremier%2BLeague&lg=&page=1&cid=10711513

Math and formal logic are effectively equivalent and philosophy without conditional logic is useless. Scientifically useful philosophy is just “explorative logic” or something like it

Wine/Proton on Linux occasionally beats Windows on the same hardware in gaming, because there’s inefficiencies in the original environment which isn’t getting replicated unnecessarily.

It’s not quite the same with CPU instruction translation, but the main efficiency gain from ARM is being designed to idle everything it can idle while this hasn’t been a design goal of x86 for ages. A substantial factor to efficiency is figuring out what you don’t have to do, and ARM is better suited for that.

It’s not that uncommon in specialty hardware with CPU instructions extensions for a different architecture made available specifically for translation. Some stuff can be quite efficiently translated on a normal CPU of a different architecture, some stuff needs hardware acceleration. I think Microsoft has done this on some Surface devices.

It’s the same as IPv4 (tunnel) except as mentioned above its still hard to get an IP with the right label

Quantum mechanics still have endless ratios which aren’t discrete. Especially ratios between stuff like wavelengths, particle states, and more

Complex numbers, and a bunch more things too

Neither of these mention networks, only protocols/schemes, which are concepts. Cryptography exists outside networks, and outside computer science (even if that is where it finds the most use).

This is ridiculous rules lawyering and isn’t even done well. Such schemes inherently assume multiple communicating parties. Sure you might not need to have a network but you still have to have distinct devices and a communication link of some sort (because if you have a direct trusted channel you don’t need cryptography)

You’re also wrong about your interpretation.

Here’s how to read it:

At point A both parties create their long term identity keys.

At point B they initiate a connection, and create session encryption keys with a key exchange algorithm (first half of PFS)

At point C they exchange information over the encrypted channel.

At point D the session keys are automatically deleted (second half of PFS)

At point E the long term key of one party is leaked. The contents from B and C can not be recovered because the session key is independent of the long term key and now deleted. This is forward secrecy. The adversary can’t compromise it after the fact without breaking the whole algorithm, they have to attack the clients as the session is ongoing.

This is motivated for example by how SSL3.0 usually was used with a single fixed RSA keypair per server, letting user clients generate and submit session encryption keys - allowing a total break of all communications with the server of that key is comprised. Long term DH secrets were also often later used when they should be single use. Then we moved on to ECDH where generating new session secrets is fast and everybody adopted real PFS.

Yes compromising the key means you often get stuff like the database too, etc. Not the point! If you keep deleting sensitive data locally when you should then PFS guarantees it’s actually gone, NSA can’t store the traffic in their big data warehouse and hope to steal the key later to decrypt what you thought you deleted. It’s actually gone.

And both of the above definitions you quoted means the same as the above.

In any case, both of these scenarios create an attack vector through which an adversary can get all of your old messages, which, whether you believe violates PFS by your chosen definition or not, does defeat its purpose (perhaps you prefer this phrasing to “break” or “breach”).

Playing loose with definitions is how half of all broken cryptographic schemes ended up insecure and broken. Being precise with attack definitions allows for better analysis and better defenses.

Like how better analysis of common attacks on long running chats with PFS lead to “self healing” properties being developed to counter point-in-time leaks of session keys by repeatedly performing key exchanges, better protecting long term keys by for example making sure software like Signal make use of the OS provided hardware backed keystore for it, etc. All of this is modeled carefully and described with precise terms.

Edit: given modern sandbox techniques in phones, most malware and exploits doesn’t survive a reboot. If malware can compromise your phone at a specific time but can’t break the TPM then once you reboot and your app rekeys then the adversary no longer have access, and this can be demonstrated with mathematical proofs. That’s self healing PFS.

Anyone can start a forum.

Fair point, but my cryptography forum (reddit.com/r/crypto) has regulars that include people writing the TLS specifications and other well known experts. They’re hanging around because the forum is high quality, and I’m able to keep quality high because I can tell who’s talking bullshit and who knows their stuff.

decoupled

Assuming it’s a pun

Tasker also have the autonotification plugin which let you set any rules and actions you want on notifications

Once again reminding you that I run a cryptography forum (I’ve done so for one 10 years, I keep up to date on the field) and it’s a term defined by professional cryptographers.

https://www.sectigo.com/resource-library/perfect-forward-secrecy

https://link.springer.com/referenceworkentry/10.1007/978-1-4419-5906-5_90

https://www.sciencedirect.com/topics/computer-science/forward-secrecy

Literally all definitions speak of network traffic and leaked / extracted encryption keys. PFS is about using short term keys that you delete so that they can not leak later.

Backup and sync via a separate mechanism is not a PFS violation. In particular because they’re independent of that same encrypted session. It’s entirely a data retention security issue.

Matrix.org supports message log backup via the server, and does so by uploading encrypted message logs and syncing the keys between clients. You can delete the logs later, or delete your keys, or even push fake logs if you want. It’s still happening outside of the original encrypted session and the adversary can’t confirm what actually was said in the original session.

I don’t know why you think that PFS is broken if a local client has to be breached to recover encrypted data from a cloud backup, but PFS is not broken if a local client has to be breached to recover the same data from the client itself. Literally the only difference is where the data is stored, so either chat logs available to the client break PFS or they do not

Matrix does it this way

I run a cryptography forum, I know the exact definition of these terms. Message logs in plaintext is very distinct from forward secrecy. What forward secrecy means in particular is that captured network traffic can’t be decrypted later even if you at a later point can steal the user’s keys (because the session used session keys that were later deleted). Retrieving local logs with no means of verifying authenticity is nothing more than a classical security breach.

You can transfer messages as a part of an account transfer on Signal (at least on Android). This deactivates the app on the old device (so you can’t do it silently to somebody’s device)

Message logs doesn’t break forward secrecy in a cryptographic sense, retaining original asymmetric decryption keys (or method to recreate them) does. Making history editable would help against that too.

What Signal actually intends is to limit privacy leaks, it only allows history transfer when you transfer the entire account to another device and “deactivate” the account on the first one, so you can’t silently get access to all of somebody’s history

You can still push old message history from your main device to your other devices, you can re-encrypt

Am encrypted container doesn’t help if the directory is mounted and accessible or if the key is in plaintext. Also doesn’t help if the process isn’t isolated. You need a bunch of extra measures like using the OS keystore set to only allow the correct program to retrieve the key, keeping secrets only in process memory, etc.

Tldr it’s a lot of work to do it right. If you do it the simple way like throwing it all in SQLite with encryption active you still leak metadata.

Const