September 2024 - nexa - server-nexa.polito.it

Database di accordi commerciali tra aziende AI e piattaforme ed editori per l'uso di contenuti

by Antonio Vetro'

Un database che riporta gli accordi che le aziende di AI hanno stipulato con piattaforme ed editori per utilizzare i loro contenuti per l'addestramento di Large Language Models (LLM) https://petebrown.quarto.pub/pnp-ai-partnerships/ Sviluppato da ricercatori della Columbia Journalism School, non ci sono però le fonti usate. A.

Sept. 6, 2024

Re: [nexa] draft 0.0.9 di Open Source AI Definition (da: large language model e open washing)

by Antonio

D'accordo su tutto, ma io sono ancora più "critico" > Even if the weights are handy to modify an AI system, they are in no > way enough to study it. No, i pesi non sono sufficienti per modificare (e ovviamente creare qualcosa di sensato). A questo sono arrivato solo costruendomi, l'anno scorso, un microLM. https://github.com/opensignature/stories/tree/main I pesi, tutti nel file story.h, sono poco più di 250000, niente a confronto dei miliardi degli LLM. Eppure dubito che si riesca a modificarli per ottenere qualcosa di altrettanto funzionante. Il "peso" è il risultato di una serie di operazioni matematiche irreversibili, se li modifichi ottieni frasi senza senso, parole (composte da token) senza senso, ecc. A.

Sept. 6, 2024

Durov: la sua "versione" di cosa accaduto

by Damiano Verzulli

Fonte: lui, direttamente su Telegram, qui: https://t.me/durov/342 Copio/Incollo, qui sotto. Bye, DV ------------------------- ❤️ Thanks everyone for your support and love! Last month I got interviewed by police for 4 days after arriving in Paris. I was told I may be personally responsible for other people’s illegal use of Telegram, because the French authorities didn’t receive responses from Telegram. This was surprising for several reasons: 1. Telegram has an official representative in the EU that accepts and replies to EU requests. Its email address has been publicly available for anyone in the EU who googles “Telegram EU address for law enforcement”. 2. The French authorities had numerous ways to reach me to request assistance. As a French citizen, I was a frequent guest at the French consulate in Dubai. A while ago, when asked, I personally helped them establish a hotline with Telegram to deal with the threat of terrorism in France. 3. If a country is unhappy with an internet service, the established practice is to start a legal action against the service itself. Using laws from the pre-smartphone era to charge a CEO with crimes committed by third parties on the platform he manages is a misguided approach. Building technology is hard enough as it is. No innovator will ever build new tools if they know they can be personally held responsible for potential abuse of those tools. Establishing the right balance between privacy and security is not easy. You have to reconcile privacy laws with law enforcement requirements, and local laws with EU laws. You have to take into account technological limitations. As a platform, you want your processes to be consistent globally, while also ensuring they are not abused in countries with weak rule of law. We’ve been committed to engaging with regulators to find the right balance. Yes, we stand by our principles: our experience is shaped by our mission to protect our users in authoritarian regimes. But we’ve always been open to dialogue. Sometimes we can’t agree with a country’s regulator on the right balance between privacy and security. In those cases, we are ready to leave that country. We've done it many times. When Russia demanded we hand over “encryption keys” to enable surveillance, we refused — and Telegram got banned in Russia. When Iran demanded we block channels of peaceful protesters, we refused — and Telegram got banned in Iran. We are prepared to leave markets that aren’t compatible with our principles, because we are not doing this for money. We are driven by the intention to bring good and defend the basic rights of people, particularly in places where these rights are violated. All of that does not mean Telegram is perfect. Even the fact that authorities could be confused by where to send requests is something that we should improve. But the claims in some media that Telegram is some sort of anarchic paradise are absolutely untrue. We take down millions of harmful posts and channels every day. We publish daily transparency reports (like this or this ). We have direct hotlines with NGOs to process urgent moderation requests faster. However, we hear voices saying that it’s not enough. Telegram’s abrupt increase in user count to 950M caused growing pains that made it easier for criminals to abuse our platform. That’s why I made it my personal goal to ensure we significantly improve things in this regard. We’ve already started that process internally, and I will share more details on our progress with you very soon. I hope that the events of August will result in making Telegram — and the social networking industry as a whole — safer and stronger. Thanks again for your love and memes 🙏 ================ -- Inviato dal mio dispositivo Android con K-9 Mail. Perdonate la brevità.

Sept. 6, 2024

Re: [nexa] AI Training is Copyright Infringement

by GC F

La nozione dicotomia idea/espressione sta ad indicare che il diritto d'autore non protegge dati, idee, fatti storici, formule matematiche, etc ma solo "espressioni", quindi un processo di text-and-data-mining che è teso all'estrazione di dati, in quanto tali non oggetto di protezione autoriale, non dovrebbe a rigor di logica causare una violazione del diritto d'autore. Questo anche nel caso in cui per estrarre l'elemento improteggibile dato/idea si debba preliminarmente fare una copia (diritto esclusivo) di ciò che è proteggibile invece, l'"espressione" in cui è espressa l'idea/dato. Questa conclusione è palesemente accettata dalla giurisprudenza statunitense fin dai tempi del caso Baker v Selden (1884) e a seguito dell'affermazione della dottrina del fair use, non invece in diritto UE, dove pare prevalente l'opinione che una copia intermedia anche se effettuata per estrarre e utilizzare elementi non protetti dal diritto d'autore, quindi nel pubblico dominio, comunque violi tali diritti e sia fonte di responsabilità per violazione. Questa la ragione per cui in diritto EU abbiamo introdotto eccezioni e limitazioni specifiche per il text-and-data-mining. Sostenendo che la dicotomia/idea espressione sia la grundnorm del diritto d'autore, sostengo anche che qualsiasi conclusione che porti a identificare una violazione in processi di utilizzo di espressioni proteggibili per estrarre elementi improteggibili sia incompatibile con i principi generali e strutturali del diritto d'autore. Giancarlo On Fri, Sep 6, 2024 at 1:12 AM Giacomo Tesio <giacomo(a)tesio.it> wrote: > Salve Giancarlo, > > posso approfittare della tua competenza per un chiarimento? > > Cosa c'entra la > > > "dicotomia idea/espressione", forse la grundnorm del diritto d'autore > > ? > > Dentro un software programmato statisticamente non c'è alcuna mente che > possa elaborare idee, solo un intricato sistema automatico di collage > delle espressioni. Lo chiamano "machine learning" per ingannare coloro > che non sanno come funziona quel tipo di software, ma non c'è nessuno > che apprenda alcunché, nemmeno dentro una "rete neurale artificiale". > > > Giacomo >

Sept. 6, 2024

Re: [nexa] AI Training is Copyright Infringement

by Giacomo Tesio

Salve Giancarlo, posso approfittare della tua competenza per un chiarimento? Cosa c'entra la > "dicotomia idea/espressione", forse la grundnorm del diritto d'autore ? Dentro un software programmato statisticamente non c'è alcuna mente che possa elaborare idee, solo un intricato sistema automatico di collage delle espressioni. Lo chiamano "machine learning" per ingannare coloro che non sanno come funziona quel tipo di software, ma non c'è nessuno che apprenda alcunché, nemmeno dentro una "rete neurale artificiale". Giacomo

Sept. 6, 2024

Re: [nexa] draft 0.0.9 di Open Source AI Definition (da: large language model e open washing)

by Giacomo Tesio

Ops! il mio post è stato "nascosto"... On Fri, 6 Sep 2024 00:55:55 +0200 Giacomo Tesio <giacomo(a)tesio.it> wrote: > Qui trovi la mia controproposta: > > https://discuss.opensource.org/t/draft-v-0-0-9-of-the-open-source-ai-defini… Fortunatamente, la Wayback Machine è stata più veloce: http://web.archive.org/web/20240905230145/https://discuss.opensource.org/t/… Riporto comunque il contenuto di seguito, caso mai richiedessero di cancellarlo anche di lì (mi è già capitato in passato...) ``` Totally agree with @thesteve0. Systems based on machine learning techniques are composed of two kind of software: a virtual machine (with a specific architecture) that basically maps vectors to vectors and a set of “weight” matrices that constitute the software executed by such virtual machine (the “AI model”). The source code of the virtual machine can be open source, so that given the proper compiler, we can create an exact copy of such software. In the same way, the software executed by the virtual machine (usually referred to as “the AI model”) is encoded in a binary form that the specific machine can directly execute (the weight matrices). The source code of such binary is composed of all the data required to recreate an exact copy of the binary (the weights). Such data include the full dataset used but also any random seed or input used during the process, such as, for example, the initial random value used to initialize an artificial neural network. Even if the weights are handy to modify an AI system, they are in no way enough to study it. So, any system that does not provide the whole dataset required to recreate an exact copy of the model, cannot be defined open source. Note that in a age of supply chain attacks that leverage opensource, the right to study the system also has a huge practical security value as arXiv:2204.06974 showed that you can plant undetectable backdoors in machine learning models. Thus I suggest to modify the definition so that Data information: Sufficiently detailed information about all the data used to train the system (including any random value used during the process), so that a skilled person can recreate an exact copy of the system using the same data. Data information shall be made available with licenses that comply with the Open Source Definition. Being able to build a “substantially equivalent” system means not being able to build that system, but a different one. It would be like defining Google Chrome as “open source” just because we have access to Chromium source code. When its training data cannot legally be shared, an AI system cannot be defined as “open source” even if all the other components comply with the open source definition, because you cannot study that system, but only the components available under the os license. Such a system can be valuable, but not open source, even if the weights are available under a OSD compliant license, because they encode an opaque binary for a specific architecture, not source code. Lets properly call such models and systems “freeware” and build a definition of OpenSource AI that is coherent with the OpenSource one. ``` Giacomo

Sept. 5, 2024

Re: [nexa] AI Training is Copyright Infringement

by GC F

e daremo il benvenuto al "third enclosure movement" e l'ennesima vittoria del "copyright maximalism"... Giancarlo On Thu, Sep 5, 2024 at 11:57 PM GC F <gcfrosio(a)gmail.com> wrote: > Studio commissionato da una lobby di parte che "sorprendentemente" > fornisce prova che serve gli interessi di quella parte ("we now have proof > that" - sic!). Al di là dei contenuti specifici, e siamo in molti con > posizioni differenti o perlomeno più caute (e infatti, "This study > challenges the prevailing European legal stance" e aggiungerei > "internazionale"), ma qual'è il valore scientifico visto il contesto? Poi, > aggiungerei, il fatto che l'intero rapporto sia redatto in tedesco non > aiuta neppure l'accesso alla comunità internazionale per vagliare premesse, > sviluppo del ragionamento e conclusioni. A una preliminare lettura tramite > traduzione automatizzata di ToC e intro, non vendo riferimenti > importanti alle questioni salienti, eg quali "dicotomia idea/espressione", > forse la grundnorm del diritto d'autore, potenziale liceità della copia > digitale intermedia per usi trasformativi, distinzione input e output (e se > questo output finale violi o meno i diritti autoriali quale opera > derivata), distinzione tra responsabilità delle piattaforma che addestra la > macchina a produrre "infinite" potenzialità lecite e illecite e > responsabilità dell'utente finale che fornisce alla macchina quei "suitable > prompts" che conducono a generare materiali illeciti, circonvenendo tra > l'altro gli strumenti tecnologici a tutela dei diritti autoriali che le > piattaforme generative hanno implementato nella creazione dell'algoritmo. > Comunque, "we now have proof"...e allora ci dimenticheremo dei tanti dubbi > che la "prevailing legal stance" si pone... > > On Thu, Sep 5, 2024 at 10:02 PM Daniela Tafani <daniela.tafani(a)unipi.it> > wrote: > >> Press Release: A computer scientist and a legal scholar shed light on the >> black box of processing steps in AI training - for the first time on this >> scale. >> >> The presentation of the interdisciplinary study “Copyright & Training of >> Generative AI - Technological and Legal Foundations” took place today in >> the European Parliament. >> >> In spring, the Copyright Initiative commissioned Prof. Dr. Tim W. Dornis >> (University of Hannover) in collaboration with Prof. Dr. Sebastian Stober >> (University of Magdeburg) with a tandem expert opinion on the technological >> and legal aspects of training generative AI models. Their interdisciplinary >> research provides urgently needed new insights into the technically >> necessary intermediate steps in the training of generative artificial >> intelligence. For the first time on this scale, a computer scientist and a >> legal scholar are jointly creating evidence regarding the processing steps >> in AI training. During the event, many open questions about protected >> materials were answered in a well-founded, reliable manner and in line with >> the current state of the art. >> >> The work of Prof. Dornis and Prof. Stober focuses on the copyright >> assessment of the processing of protected material in AI training: >> >> “As a closer look at the technology of generative AI models reveals, the >> training of such models is not a case of text and data mining. It is a case >> of copyright infringement – no exception applies under German and European >> copyright law,” says Prof. Dornis. Prof. Stober explains that “parts of the >> training data can be memorized in whole or in part by current generative >> models - LLMs and (latent) diffusion models - and can therefore be >> generated again with suitable prompts by end users and thus reproduced.” >> Axel Voss, MEP and host of today's event in the European Parliament, >> expressly thanks the scientists Dornis and Stober and is pleased that >> >> “the study not only proves that the training of Generative AI models is >> not covered by text and data mining, but that it also provides further >> important indications and suggestions for a better balance between the >> protection of human creativity and the promotion of AI innovation.” >> “This study is explosive because it proves that we are dealing with >> large-scale theft of intellectual property. The ball is now in the >> politicians' court to draw the necessary conclusions and finally put an end >> to this theft at the expense of journalists and other authors,” >> commented Hanna Möllers, legal advisor to the DJV and representative of >> the European Federation of Journalists (EFJ). >> >> Katharina Uppenbrink, Managing Director of the Initiative Urheberrecht, >> emphasizes: >> >> “It is a groundbreaking result if we now have proof that the reproduction >> of works by an AI model constitutes a copyright-relevant reproduction and, >> in addition, that making them available on the European Union market may >> infringe the right of making available to the public.” >> The composer and spokesperson for the Copyright Initiative, Matthias >> Hornschuh, comments: >> >> “There would be a new, profitable licensing market on the horizon, but no >> remuneration is flowing, while generative AI is preparing to replace those >> whose content it lives from in its own market. This jeopardizes >> professional knowledge work and cannot be in the interests of society, >> culture or the economy. All the better that the authors of our tandem study >> provide the technological and copyright basis for finally turning the legal >> consideration of generative artificial intelligence from its head to its >> feet.” >> Dornis, Tim W. and Stober, Sebastian, Copyright and training of >> generative AI models - technological and legal foundations >> >> (September 4, 2024). >> >> https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4946214 >> >> (in German) >> >> Please find below the downloads: >> >> The abstract can be found below and here (in English and German language): >> < >> https://urheber.info/media/pages/diskurs/ai-training-is-copyright-infringem… >> > >> The executive summary: < >> https://urheber.info/media/pages/diskurs/ai-training-is-copyright-infringem… >> > >> The presentation in Berlin will take place at the end of September. >> >> <https://urheber.info/diskurs/ai-training-is-copyright-infringement> > >

Sept. 5, 2024

Re: [nexa] AI Training is Copyright Infringement

by GC F

Studio commissionato da una lobby di parte che "sorprendentemente" fornisce prova che serve gli interessi di quella parte ("we now have proof that" - sic!). Al di là dei contenuti specifici, e siamo in molti con posizioni differenti o perlomeno più caute (e infatti, "This study challenges the prevailing European legal stance" e aggiungerei "internazionale"), ma qual'è il valore scientifico visto il contesto? Poi, aggiungerei, il fatto che l'intero rapporto sia redatto in tedesco non aiuta neppure l'accesso alla comunità internazionale per vagliare premesse, sviluppo del ragionamento e conclusioni. A una preliminare lettura tramite traduzione automatizzata di ToC e intro, non vendo riferimenti importanti alle questioni salienti, eg quali "dicotomia idea/espressione", forse la grundnorm del diritto d'autore, potenziale liceità della copia digitale intermedia per usi trasformativi, distinzione input e output (e se questo output finale violi o meno i diritti autoriali quale opera derivata), distinzione tra responsabilità delle piattaforma che addestra la macchina a produrre "infinite" potenzialità lecite e illecite e responsabilità dell'utente finale che fornisce alla macchina quei "suitable prompts" che conducono a generare materiali illeciti, circonvenendo tra l'altro gli strumenti tecnologici a tutela dei diritti autoriali che le piattaforme generative hanno implementato nella creazione dell'algoritmo. Comunque, "we now have proof"...e allora ci dimenticheremo dei tanti dubbi che la "prevailing legal stance" si pone... On Thu, Sep 5, 2024 at 10:02 PM Daniela Tafani <daniela.tafani(a)unipi.it> wrote: > Press Release: A computer scientist and a legal scholar shed light on the > black box of processing steps in AI training - for the first time on this > scale. > > The presentation of the interdisciplinary study “Copyright & Training of > Generative AI - Technological and Legal Foundations” took place today in > the European Parliament. > > In spring, the Copyright Initiative commissioned Prof. Dr. Tim W. Dornis > (University of Hannover) in collaboration with Prof. Dr. Sebastian Stober > (University of Magdeburg) with a tandem expert opinion on the technological > and legal aspects of training generative AI models. Their interdisciplinary > research provides urgently needed new insights into the technically > necessary intermediate steps in the training of generative artificial > intelligence. For the first time on this scale, a computer scientist and a > legal scholar are jointly creating evidence regarding the processing steps > in AI training. During the event, many open questions about protected > materials were answered in a well-founded, reliable manner and in line with > the current state of the art. > > The work of Prof. Dornis and Prof. Stober focuses on the copyright > assessment of the processing of protected material in AI training: > > “As a closer look at the technology of generative AI models reveals, the > training of such models is not a case of text and data mining. It is a case > of copyright infringement – no exception applies under German and European > copyright law,” says Prof. Dornis. Prof. Stober explains that “parts of the > training data can be memorized in whole or in part by current generative > models - LLMs and (latent) diffusion models - and can therefore be > generated again with suitable prompts by end users and thus reproduced.” > Axel Voss, MEP and host of today's event in the European Parliament, > expressly thanks the scientists Dornis and Stober and is pleased that > > “the study not only proves that the training of Generative AI models is > not covered by text and data mining, but that it also provides further > important indications and suggestions for a better balance between the > protection of human creativity and the promotion of AI innovation.” > “This study is explosive because it proves that we are dealing with > large-scale theft of intellectual property. The ball is now in the > politicians' court to draw the necessary conclusions and finally put an end > to this theft at the expense of journalists and other authors,” > commented Hanna Möllers, legal advisor to the DJV and representative of > the European Federation of Journalists (EFJ). > > Katharina Uppenbrink, Managing Director of the Initiative Urheberrecht, > emphasizes: > > “It is a groundbreaking result if we now have proof that the reproduction > of works by an AI model constitutes a copyright-relevant reproduction and, > in addition, that making them available on the European Union market may > infringe the right of making available to the public.” > The composer and spokesperson for the Copyright Initiative, Matthias > Hornschuh, comments: > > “There would be a new, profitable licensing market on the horizon, but no > remuneration is flowing, while generative AI is preparing to replace those > whose content it lives from in its own market. This jeopardizes > professional knowledge work and cannot be in the interests of society, > culture or the economy. All the better that the authors of our tandem study > provide the technological and copyright basis for finally turning the legal > consideration of generative artificial intelligence from its head to its > feet.” > Dornis, Tim W. and Stober, Sebastian, Copyright and training of generative > AI models - technological and legal foundations > > (September 4, 2024). > > https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4946214 > > (in German) > > Please find below the downloads: > > The abstract can be found below and here (in English and German language): > < > https://urheber.info/media/pages/diskurs/ai-training-is-copyright-infringem… > > > The executive summary: < > https://urheber.info/media/pages/diskurs/ai-training-is-copyright-infringem… > > > The presentation in Berlin will take place at the end of September. > > <https://urheber.info/diskurs/ai-training-is-copyright-infringement>

Sept. 5, 2024

Re: [nexa] draft 0.0.9 di Open Source AI Definition (da: large language model e open washing)

by Giacomo Tesio

Ciao Antonio, grazie per la segnalazione. > - Data information: Sufficiently detailed information about the data > used to train the system, so that a skilled person can recreate a > substantially equivalent system using the same or similar data. [Data > information shall be made available with licenses that comply with > the Open Source Definition.] (aggiunto) Qui trovi la mia controproposta: https://discuss.opensource.org/t/draft-v-0-0-9-of-the-open-source-ai-defini… Giacomo

Sept. 5, 2024

Re: [nexa] AI Training is Copyright Infringement

by Stefano Quintarelli

sounds familiar... se non sbaglio fu Locke a postulare che i frutti della terra appartenevano a chi la lavorava, ponendo le basi ideologiche per l'appropriazione dei territori indiani da parte dei coloni. una massiccia appropriazione di valore da parte dei newcomers a scapito dei precedenti. ciao, s. Il 5 settembre 2024 21:02:50 UTC, Daniela Tafani <daniela.tafani(a)unipi.it> ha scritto: >Press Release: A computer scientist and a legal scholar shed light on the black box of processing steps in AI training - for the first time on this scale. > >The presentation of the interdisciplinary study “Copyright & Training of Generative AI - Technological and Legal Foundations” took place today in the European Parliament. > >In spring, the Copyright Initiative commissioned Prof. Dr. Tim W. Dornis (University of Hannover) in collaboration with Prof. Dr. Sebastian Stober (University of Magdeburg) with a tandem expert opinion on the technological and legal aspects of training generative AI models. Their interdisciplinary research provides urgently needed new insights into the technically necessary intermediate steps in the training of generative artificial intelligence. For the first time on this scale, a computer scientist and a legal scholar are jointly creating evidence regarding the processing steps in AI training. During the event, many open questions about protected materials were answered in a well-founded, reliable manner and in line with the current state of the art. > >The work of Prof. Dornis and Prof. Stober focuses on the copyright assessment of the processing of protected material in AI training: > >“As a closer look at the technology of generative AI models reveals, the training of such models is not a case of text and data mining. It is a case of copyright infringement – no exception applies under German and European copyright law,” says Prof. Dornis. Prof. Stober explains that “parts of the training data can be memorized in whole or in part by current generative models - LLMs and (latent) diffusion models - and can therefore be generated again with suitable prompts by end users and thus reproduced.” >Axel Voss, MEP and host of today's event in the European Parliament, expressly thanks the scientists Dornis and Stober and is pleased that > >“the study not only proves that the training of Generative AI models is not covered by text and data mining, but that it also provides further important indications and suggestions for a better balance between the protection of human creativity and the promotion of AI innovation.” >“This study is explosive because it proves that we are dealing with large-scale theft of intellectual property. The ball is now in the politicians' court to draw the necessary conclusions and finally put an end to this theft at the expense of journalists and other authors,” >commented Hanna Möllers, legal advisor to the DJV and representative of the European Federation of Journalists (EFJ). > >Katharina Uppenbrink, Managing Director of the Initiative Urheberrecht, emphasizes: > >“It is a groundbreaking result if we now have proof that the reproduction of works by an AI model constitutes a copyright-relevant reproduction and, in addition, that making them available on the European Union market may infringe the right of making available to the public.” >The composer and spokesperson for the Copyright Initiative, Matthias Hornschuh, comments: > >“There would be a new, profitable licensing market on the horizon, but no remuneration is flowing, while generative AI is preparing to replace those whose content it lives from in its own market. This jeopardizes professional knowledge work and cannot be in the interests of society, culture or the economy. All the better that the authors of our tandem study provide the technological and copyright basis for finally turning the legal consideration of generative artificial intelligence from its head to its feet.” >Dornis, Tim W. and Stober, Sebastian, Copyright and training of generative AI models - technological and legal foundations > >(September 4, 2024). > >https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4946214 > >(in German) > >Please find below the downloads: > >The abstract can be found below and here (in English and German language): ><https://urheber.info/media/pages/diskurs/ai-training-is-copyright-infringem…> >The executive summary: <https://urheber.info/media/pages/diskurs/ai-training-is-copyright-infringem…> >The presentation in Berlin will take place at the end of September. > ><https://urheber.info/diskurs/ai-training-is-copyright-infringement>

Sept. 5, 2024