The judge appears to disagree with you on this. They found that training and selling an LLM are fair use, based on the fact that it is exceedingly transformative, and that the copyright holders are not entitled to any profits thereof due to copyright. (They also did get paid — Anthropic acquired millions of books legally, including all of the authors in this complaint. This would not retroactively absolve them of legal fault for past infringements, of course.)
The trial is scheduled for December 2025. That's when a jury will decide how much Anthropic owes for copying and storing over seven million pirated books
Yes, that would by an interesting trial. But it is only about six books, and all claims regarding Claude have been dismissed already. So only the internal copies remain, and there the theory for them being infringing is somewhat convoluted: you have to argue that they are not just for purposes of training (which was ruled fair use), and award damages even though these other purposes never materialised (since by now, they have legal copies of those books). I can see it, but I would not count on there being a trial.
The fallacy in the 'fair use' logic is that a person acquires a book and learns from it, but a machine incorporates the text. Copyright does not allow one to create a derivative work without permission. Only when the result of the transformation resembles the original work could it be said that it is subject to copyright. Do not regard either of those legal issues are set in concrete yet.
Both a human and a machine learn from it. You can design an LLM that doesn’t spit back the entire text after annealing. It just learns the essence like a human.
Morally maybe, but AFAIK machines "learning" and creating creative works on their own is not recognized legally, at least certainly not the same way as for people.
Just downloading them is of course cheaper, but it is worth pointing out that, as the article states, they did also buy legitimate copies of millions of books. (This includes all the books involved in the lawsuit.) Based on the judgement itself, Anthropic appears to train only on the books legitimately acquired. Used books are quite cheap, after all, and can be bought in bulk.
Buying a book is not license to re-sell that content for your own profit. I can't buy a copy of your book, make a million Xeroxes of it and sell those. The license you get when you buy a book is for a single use, not a license to do what ever you want with the contents of that book.
Yes, of course! In this case, the judge identified three separate instances of copying: (1) downloading books without authorisation to add to their internal library, (2) scanning legitimately purchased books to add to their internal library, and (3) taking data from their internal library for the purposes of training LLMs. The purchasing part is only relevant for (2) — there the judge ruled that this is fair use. This makes a lot of sense to me, since no additional copies were created (they destroyed the physical books after scanning), so this is just a single use, as you say. The judge also ruled that (3) is fair use, but for a different reason. (They declined to decide whether (1) is fair use at this point, deferring to a later trial.)
Location: Munich, Germany
Remote: Flexible
Willing to relocate: Yes
Technologies: Python, C++
Résumé/CV: https://nicze.de/philipp/cv.pdf
Email: philipp.czerner@nicze.de
I am currently finishing a PhD in theoretical computer science and looking for a position in industry. I do a lot of programming, usually in Python and C++, and with a mathematical or algorithmic focus. For example, as part of my research I developed a BDD library that uses complexity theory to guarantee correctness with small overhead. Or recently, in my spare time, I created an AI for MTG (drafting, to be precise), using a transformer architecture with PyTorch. I am very enthusiastic about using or coming up with theoretical ideas that can improve outcomes in practice. I rely on having strong fundamentals, from mathematics over algorithms to knowledge of the actual hardware, to enable me to easily pick up new areas and tackle novel problems. Most of my research is the result of collaborations — it is important to me to both support my colleagues and be able to learn from them.
Please feel free to refer to my CV for more detailed information, or my website (https://nicze.de/philipp/). Currently, I am very interested in AI, but I am also happy to discuss any other interesting opportunities!
If you are caught driving above the legal limit of 0.05% you are fined roughly $570, are prohibited from driving for 1 month, and receive 2 “points”. Points accumulate and once you reach 8 you lose your drivers license. In this case you would keep the points for five years. Many different driving offences give you points.
For comparison, to get a similar penalty by speeding you would have to exceed the speed limit by 51 km/h (32 mph).
There are many additional related offences you could commit, with different consequences. Repeat offences to the above, for example, are punished more severely: you get 3 months instead of 1 and the fine is doubled and tripled for the second and third offence, respectively. Already with a blood alcohol level of 0.03% you risk legal consequences, e.g. if you make an error while driving. If you endanger someone else (or property) with that level you are committing a crime, will lose your license, and can go to prison. If you are in your probationary period (two years after acquiring your license), any nonzero level is an offence.
Losing your license is generally temporary. You are blocked from re-acquiring it for some time, depending on the offence (at least 6 months, but can be multiple years). You have to complete an MPU, which certifies your ability to safely drive. For alcohol based offences, this would include demonstrating that you have reduced your consumption significantly. This can be quite harsh; you may, for example, be required to show complete abstinence for a period of one year. Of course, you are also looking at costs close to $1000 for the MPU alone. It is possible to get permanently blocked from driving, but it's quite difficult, I believe.
I have a Framework 13 (AMD 7640U), running Arch Linux, and overall it is nice. It is convenient to have an HDMI port again and I did not encounter issues with hardware. (Sleep drains the battery somewhat quickly over longer time periods, but I put it in hibernate anyway for that.)
However, the screen broke down after a few months and support initially refused to replace it, citing “customer induced damage”. As far as I can tell, this is both untrue and illegal (under German law, within 12 months after purchase all defects are presumed to be due to the original condition of the product, for which the seller is liable; that presumption can be overcome, but you would need some reason). They relented eventually, but it certainly soured my opinion on both the product and the company.
Let us say that an index i is bad, if every finite subsequence of s starting at i is red (i.e. for every j ≥ i we have χ(s_i ... s_j) = red). Two cases:
Case 1: there are infinitely many bad indices. Here we go to the first bad index then the second, and so on. The colour of w₀ does not matter, and since subsequent words start at a bad index, they will all be red.
Case 2: there are finitely many bad indices. Then there is some k which is larger than all bad indices. We start by going to k (again, the colour of w₀ does not matter). Since k is not bad, there must be some blue word starting at k. We take that one and move to a larger index. Again, that index is not bad. We repeat this process to find our sequence.
Nice proof, similar to the one I posted just now but simpler -- you realised that we only need one special category ("bad" rather than "hard-red" + "hard-blue"). Gonna leave mine up though :)
The full decision can be found here [1]. The consumer protection agency did also seek that LinkedIn be forced to respect DNT, but the court did not grant this relief, reasoning that it was overly broad in two ways. First, it did not specify precisely enough what is meant by DNT — in particular, the suit did not limit itself to the DNT header, but referred to any kind of configured signals sent by the browser. Second, it described the behaviour that LinkedIn is supposed to cease when encountering such a signal in an overly broad manner.
If upheld, the judgement certainly seems to open the door for future litigation, and one might even hope for potential targets to adjust their behaviour in anticipation of it, but I would not hold my breath there.
The name of the city is “Garching bei München” which translates to “Garching near Munich”. This disambiguates it from „Garching an der Alz“. (Although Jülich is just called Jülich.)
I use these as well (3M Aura). They are much more comfortable than the more common types of FFP2 masks and I can wear them for prolonged periods of time without issue, but I would not classify them as “no discomfort whatsoever”. Also, they have a very good fit – other FFP2 masks I had to fiddle around with quite a bit to get a good seal around my nose.
A very important thing is that at least for some types of 3M Aura (the 9320+, at least), the nose clip is significantly stronger than that of cheap masks. So if you're used to pushing it onto your nose very strongly with cheap masks, you need to unlearn that with the Aura masks, or it will obstruct your breathing and become very uncomfortable.
With the Aura masks (at least those with a strong clip), it's not necessary to push it down very strongly; you just need to lightly shape it, and then the strength of the clip will ensure that it stays in shape, and the foam around the nose will ensure a good seal.
It seems to take some practice to find the right clip shape, but once you've found it, they're quite comfortable IMO - and most of the time I've seen people complain about discomfort with these masks, it seems to be because of the nose clip being too tight.
Someone I know wanted to have a movable white rectangle on the screen to cover up things (for a presentation). They had a creative solution: open a “10h white background” video and used Firefox's picture-in-picture feature. Unfortunately, the recommendation algorithm picked up on this, and started recommending a bunch of similar videos...
