from the fair-trade-ai-training-data dept

As many of the AI stories on Walled Culture attest, one of the most contentious areas in the latest stage of AI development concerns the sourcing of training data. To create high-quality large language models (LLMs) massive quantities of training data are required. In the current genAI stampede, many companies are simply scraping everything they can off the Internet. Quite how that will work out in legal terms is not yet clear. Although a few court cases involving the use of copyright material for training have been decided, many have not, and the detailed contours of the legal landscape remain uncertain.

However, there is an alternative to this “grab it all” approach. It involves using materials that are either in the public domain or released under a “permissive” license that allows LLMs to be trained on them without any problems. There’s plenty of such material online, but its scattered nature puts it at a serious disadvantage compared to downloading everything without worrying about licensing issues. To address that, the Common Corpus was created and released just over a year ago by the French startup Pleias. A press release from the AI Alliance explains the key characteristics of the Common Corpus:

Truly Open: contains only data that is permissively licensed and provenance is documented

Multilingual: mostly representing English and French data, but contains at least 1[billion] tokens for over 30 languages

Diverse: consisting of scientific articles, government and legal documents, code, and cultural heritage data, including books and newspapers

Advertisement

Extensively Curated: spelling and formatting has been corrected from digitized texts, harmful and toxic content has been removed, and content with low educational content has also been removed.

There are five main categories of material: OpenGovernment, OpenCulture, OpenScience, OpenWeb, and OpenSource:

OpenGovernment contains Finance Commons, a dataset of financial documents from a range of governmental and regulatory bodies. Finance Commons is a multimodal dataset, including both text and PDF corpora. OpenGovernment also contains Legal Commons, a dataset of legal and administrative texts. OpenCulture contains cultural heritage data like books and newspapers. Many of these texts come from the 18th and 19th centuries, or even earlier.

OpenScience data primarily comes from publicly available academic and scientific publications, which are most often released as PDFs. OpenWeb contains datasets from YouTube Commons, a dataset of transcripts from public domain YouTube videos, and websites like Stack Exchange. Finally, OpenSource comprises code collected from GitHub repositories which were permissibly licensed.

The initial release contained over 2 trillion tokens – the usual way of measuring the volume of training material, where tokens can be whole words and parts of words. A significant recent update of the corpus has taken that to over 2.267 trillion tokens. Just as important as the greater size, is the wider reach: there are major additions of material from China, Japan, Korea, Brazil, India, Africa and South-East Asia. Specifically, the latest release contains data for eight languages with more than 10 billion tokens (English, French, German, Spanish, Italian, Polish, Greek, Latin) and 33 languages with more than 1 billion tokens. Because of the way the dataset has been selected and curated, it is possible to train LLMs on fully open data, which leads to auditable models. Moreover, as the original press release explains:

By providing clear provenance and using permissibly licensed data, Common Corpus exceeds the requirements of even the strictest regulations on AI training data, such as the EU AI Act. Pleias has also taken extensive steps to ensure GDPR compliance, by developing custom procedures to enable personally identifiable information (PII) removal for multilingual data. This makes Common Corpus an ideal foundation for secure, enterprise-grade models. Models trained on Common Corpus will be resilient to an increasingly regulated industry.

Another advantage for many users is that material with high “toxicity scores” has already been removed, thus ensuring that any LLMs trained on the Common Corpus will have fewer problems in this regard.

The Common Corpus is a great demonstration of the power of openness and permissive copyright licensing, and how they bring benefits that other approaches can’t match. For example: “Common Corpus makes it possible to train models compatible with the Open Source Initiative’s definition of open-source AI, which includes openness of use, meaning use is permitted for ‘any purpose and without having to ask for permission’. ” That fact, along with the multilingual nature of the Common Corpus, would make the latest version a great fit for any EU move to create “public AI” systems, something advocated on this blog a few months back. The French government is already backing the project, as are other organizations supporting openness:

The Corpus was built up with the support and concerted efforts of the AI Alliance, the French Ministry of Culture as part of the prefiguration of the service offering of the Alliance for Language technologies EDIC (ALT-EDIC).

This dataset was also made in partnership with Wikimedia Enterprise and Wikidata/Wikimedia Germany. We’re also thankful to our partner Libraries Without Borders for continuous assistance on extending low resource language support.

The corpus was stored and processed with the generous support of the AI Alliance, Jean Zay (Eviden, Idris), Tracto AI, Mozilla.

Advertisement

The unique advantages of the Common Corpus mean that more governments should be supporting it as an alternative to proprietary systems, which generally remain black boxes in terms of where their training data comes from. Publishers too would also be wise to fund it, since it offers a powerful resource explicitly designed to avoid some of the thorniest copyright issues plaguing the generative AI field today.

Follow me @glynmoody on Mastodon and on Bluesky. Originally published to Walled Culture.

Filed Under: ai, ai training, common corpus, copyright, open licenseing, open licensing, public domain, training data

Companies: common corpus, pleias

from the here-comes-the-comments dept

The polarization over any and all uses of artificial intelligence and machine learning continues. And, to be clear, I very much understand why this is all so controversial. Any new technology that has the chance to be transformative will also necessarily be disruptive and that causes fear. Fear that is not entirely unfounded, no matter your other opinions on the matter. If that’s you, cool, I get it.

I’ll start this off by pointing to the latest edition of the Techdirt podcast in which both Mike and Karl engaged in a fantastic discussion about the use of AI. I’ve listened to it twice now; it’s that good. And, while I found myself arguing out loud with the both of them at certain points during the podcast, despite the fact that neither of them could hear my retorts, it presents a grounded, often nuanced conversation, which we need much more of in this space.

And now, in what might be a subconscious attempt by this writer to commit suicide by comments section, let’s talk about that controversial demo of NVIDIA’s forthcoming DLSS 5 technology. What DLSS 5 does compared with previous versions of the technology is indeed new, but what is not new is the introduction of AI and machine learning into the equation. DLSS 2 and 3 had that already, in the form of pixel reconstruction and frame generation. DLSS 5, however, introduced what is being labeled as “neural rendering”, which uses machine learning to alter the lighting and detailed appearances in environments and, most importantly, character rendering over the engine on top of the 2D image output. Here’s the video demo that got everyone talking.

The backlash to the video was wide, immediate, and furious. There was a great deal of talk about the alteration of artistic intent, about whether this changed what the original developers were attempting to portray when they created the games, and, of course, industry jobs. I want to talk about the major complaint pillars seen across many outlets below, but this backlash also supposedly came with death threats foisted upon NVIDIA employees. I would very much hope we could all at least agree that any threats of that nature are completely inappropriate and absurd.

With that, here is what I’ve seen in the backlash and what I’d want to say about it.

Get your damned AI out of my games!

Perhaps not the most common pushback I saw in all of this, but a very common one. And a silly one, too. As I mentioned above, DLSS versions already used some version of AI and machine learning. That isn’t new. How it’s applied is certainly new, but that isn’t the same as the demand to keep AI entirely out of the video game industry.

And if that’s where you are, go ahead and shake your fist at the clouds in the sky. AI is a tool and, as I’ve now said repeatedly, the conversation we should be having is how it’s used in gaming, not if it’s used. That’s because its use is largely a foregone conclusion and it is an open question as to whether its use will be a net benefit or negative overall to the industry. Dogmatic purists on AI have a stance that is understandable, but also untenable. We’re too far down this road to turn around and go home. And if the tech were able to lower the barriers of entry to the gaming industry, acting as the fertilizer that allows a thousand indie studios to sprout roots, would that really be so bad for the gaming ecosystem?

I can appreciate the purists’ point of view. I really can. I just don’t see where they have a place in the conversation when it comes to gaming.

It overrides artistic intent!

Does it? If it did, then hell yes that’s bad. But if it doesn’t, then this concern goes away entirely.

DLSS 5 is built with options and customizable sliders for game developers. That’s really, really important here. At the macro level, a developer that has decided to use DLSS 5, or decided and customized how it’s used in their games, is exercising consent over their products. That should be obvious.

But then we get into really interesting questions of art, the actual artist, and the ownership of that art, because those last two are very different things. As Digital Foundry outlines:

It may even raise consent and other questions surrounding artistic integrity. On site and witnessing the demos in motion, concerns about this seemed less of a problem when the games we saw had been signed off by the studios that made them – the contentious assets we’ve seen, likewise. Nothing from the DLSS 5 reveal released by Nvidia has not been approved by the studios that own those games. But perhaps the issue isn’t just about specific approvals by specific developers on agreed DLSS 5 integrations, but rather the whole concept of a GPU reinterpreting game visuals according to a neural model that has its own ideas about what photo-realism should look like.

While we’ve seen endorsements from Bethesda’s Todd Howard and Capcom’s Jun Takeuchi, to what extent does that consent apply to the entire development team and other artists associated with the production? And by extension, there is also the question of whether now is the right time to launch DLSS 5 at a time when the games industry is under enormous pressure, jobs are on the line and cost-cutting is a major focus in the triple-A space. The technology itself cannot function without the work of game creators – it needs final game imagery to work at all – but the extent to which it could be viewed as a worrying sign of “things to come” cannot be overstated bearing in mind the reactions elsewhere to generative AI.

That strikes me as a valid and interesting ethical question when it comes to the use of this technology, but one that is probably overwrought. Individual artists who work on video games already have their artistic output live at the pleasure of the game developers they contract with. Those developers already can use this game art in all kinds of ways that the individual artist may not have had in mind when creating it, or indeed have even considered such possibilities. DLSS 5 is just one more version of that, with the main difference being that it involves AI making changes to game images. That’s an important thing to consider, sure, but there are cousins to this ethical question that we’ve all come to accept already. This strikes me more as part of the “all AI is bad all the time” crowd finding a foothold in something other than dogma to grab onto.

Developers and publishers own their games. If they want to use DLSS 5 in those games, there is little other than specific work for hire or other contractual stipulations with individual artists that would keep them from implementing it. If artists don’t like that, I completely understand that point of view, but that’s what contract negotiations and language are for.

Bottom line: I have been as vocal as anyone arguing that video games are a form of art for well over a decade now and I struggle to agree that an optional technology that has approved buy in from game developers and publishers equates to “overriding artistic intent”, writ large.

The faces in these examples look like shit, are “yassified”, or suffer from the uncanny valley effect!

Look, here we’re going to get into matters of opinion. I have to say that when I viewed the demo video myself, I had the opposite reaction. And, yes, this opens me up to claims that I am somehow a massive fan of AI-created pornography (this is where the yassified comments come in), or that I just want all the characters to look “hot” (I’m too old for that shit), or that my older age of 44 means I’ve lost touch with what video games should look like. Despite my genuine respect for the dissenting opinions here, allow me to say this: bullshit.

The caveat to all of this is that the demo revealed very little in the way of this technology working within these games in motion. It’s also certainly true that NVIDIA chose the best potential images to show off its new technology. If the DLSS 5 rendering sucks out loud in a larger in-motion game, or if the images it creates end up being inconsistent throughout gameplay, or if it does just end up looking shitty, then I’ll be right there with you with a torch and pitchfork in hand.

And here’s the other thing to consider with this particular complaint, combined with the previous one about artistic intent: do any of you use visual mods in your games? I do. A ton of them. For a variety of reasons. I have used them to alter the faces and models for games like Starfield and Skyrim, among many others. Do I need to feel bad for altering the artist’s intent? Do I need to apologize for incorporating mods to make characters and environments appear in a way that helps me better connect with the game I’m playing?

Because I’m not going to do either. And I don’t expect you to. Nor do I expect game developers that choose to use this optional technology to beg for forgiveness for their own output.

The hardware demands to run all of this are insane!

Fine, then you’ll get what you want and nobody will be able to use this technology anyway. But I don’t think that will be the case. NVIDIA knows what it will take to run this tech once it leaves the demo stage and goes into production. The idea that they would hype up technology that nobody can use strikes me as unlikely in the extreme.

Conclusion: everyone take a breath

This still strikes me as more of a “all AI is bad” crowd grasping at lots of other things to buttress their pushback than anything else. AI has plenty, plenty of potential pitfalls. Worried about jobs in the gaming industry and elsewhere? Me too! But if you’re not also looking at the potential upsides for the industry, then you’re engaging in dogma, not conversation.

Will DLSS 5 be good? I have no idea and neither do you. Will DLSS 5 alter previously released games in a way that fundamentally alters how we play these games? I have no idea and neither do you. Will it negatively impact the gaming industry when it comes to the number of jobs within it? I have no idea and neither do you.

This was a tech demo. Details on how it works are still trickling out. Most recently, there has been some clarification as to the 2D rendering nature of the technology and what that means for the output on the screen. As an early demo of the technology, feedback is going to be important, so long as it’s informed and reasonable feedback.

The technology may end up being trash and hated for reasons other than “all AI is bad all the time.” If that ends up being the case, I trust the gaming market to work that out for itself. But a lot of the hand-wringing here looks to me to be speculative at best.

Filed Under: ai, developers, dlss 5, rendering, video games

Companies: nvidia

The US Federal Communications Commission (FCC) is tasked with regulating both wired and wireless communications, which also includes a national security component. This is how previously the FCC tossed networking gear made by Huawei and foreign-manufactured drones onto its Covered List, effectively banning it from sale in the US. Now foreign-made consumer routers have been added to this list, barring explicit conditional approval on said list that would exempt them during a ‘transition phase’.

As per the FCC fact sheet, this follows after determination by an interagency body that such routers “pose unacceptable risks to the national security of the United States [..]”. This document points us to the National Security Determination PDF, which attempts to lay out the reasoning. In it is noted that routers are an integral part of every day life, and compromised routers are a major risk factor, ergo it follows that only US-manufactured routers are to be trusted.

These – so far fictional – US-manufactured consumer routers would have to feature ‘trusted supply chains’, which would seem to imply onshoring a large industrial base, though without specifying how deep this would have to go it’s hard to say what would be involved. The ‘supporting evidence’ section also only talks about firmware-related vulnerabilities, which would imply that US firmware developers do not produce CVEs.

Currently there do not appear to be any specific details on what router manufacturers are supposed to do about this whole issue, though they can continue to sell previously FCC-approved routers in the US.

Although hardware backdoors are definitely a possibility, this requires a fair bit of effort within the supply chain that should generally also fairly easily to detect. Yet after for example Bloomberg claimed in 2018 that Supermicro gear had been infested with hardware backdoors, this started a years-long controversy.

Meanwhile actually verified issues with Supermicro hardware are boringly due to software CVEs. In that particular issue from 2024 two CVEs were discovered involving a lack of validation of a newly uploaded firmware image.

All of which is reminiscent of an early 2024 White House ‘memory safety appeal’ that smelled very strongly of red herring. Although it’s easy to point at compromised hardware with scary backdoors and sneaky software backdoors hidden deep inside firmware of servers and networking devices, the truth of the matter is that sloppy input validation is still by far the #1 cause of fresh CVEs each year, especially if you look at the CVEs that are actually being actively exploited.

As for this de-facto ban on new routers being sold in the US, this will correspondingly not change much here. The best defense against issues with networking equipment is still to practice network hygiene by keeping tabs on what is being sent on the LAN and WAN sides, while a government could e.g. force consumer routers to pass a strict independent hardware and software audit paid for by the manufacturer.

Speaking as someone who used to run DIY routers for the longest time built around FreeSCO and Smoothwall Linux, there’s also always the option of turning any old PC into a router by putting a bunch of NICs and WNICs into it and run SmoothWall, OpenWRT, etc.. A router is after all just a specialized computer, regardless of what the government feels that it identifies as.

When Daredevil: Born Again resurrected the Marvel hero on Disney Plus last year, one thing was clearly established: This series would be as ruthlessly violent as its Netflix predecessor. It’s delivered and then some, reintroducing Matt Murdock, Wilson Fisk and the grim themes that made the superhero crime drama so compelling. While the first season took time building its interconnected storylines up to its climax (and memorable finale), season 2 has an energetic momentum that allows the show to keep its clout and keep you on edge from the first episode to the very end. 

Debuting Tuesday night on the streamer, it’s suspenseful, graphic and intelligent, with eight well-knit episodes playing out fluidly like a really long, really good movie. 

Before the main credits start rolling in the first episode of the sophomore season of Daredevil: Born Again, a video from street reporter BB Urich (Genneya Walton) paints a picture of what’s happening in the city under Mayor Fisk/Kingpin (Vincent D’Onofrio). Remember when he declared martial law in last season’s finale? Well, it’s all good vibes on camera: New York Born Again is the slogan plastered on posters around Manhattan, and citizens give the mayor a thumbs up. In fact, Fisk’s face is on many of the posters, hanging around town like ornaments, letting people know what a great job he’s doing making NYC safe. 

Action happening at night tells a different story, when a black-suited Daredevil — yes, THAT black suit with two red D’s emblazoned on the chest — boards a cargo boat on the river and starts battering its armed guards. Blood splatter flies everywhere (including the screen), broken limbs crack, and the mysterious cargo is revealed: illegal weapons. The first half of episode 1 sets up the entire season with this doomed, sinking ship that Daredevil barely escapes. 

We have a politician who turns a band of law enforcement officers into his personal army that targets his enemies, everyday citizens and rebellious “vigilantes” he’s deemed as terrorists. Kingpin is in power, always ready to use fear to instill loyalty, dole out bloody assaults and put people in cages. His Anti-Vigilante Task Force (AVTF) is looking for fugitives like Karen Page and Daredevil, whom he’s publicly named as being responsible for the ship debacle. The bloodthirsty AVTF is hunting for them and anyone who detracts (or distracts) from Kingpin’s agenda. 

It’s a treat to watch Daredevil and The Punisher beat people up, and seeing Bullseye’s tricks feeds my affinity for stylish assassins, but the core of this series’ first two seasons is Matt Murdock versus Wilson Fisk, or Daredevil versus Kingpin. There are two people who are always worried about what the other one is up to. Several characters are now in the mix who shake things up this season, including Jessica Jones (Krysten Ritter), Mr. Charles (Matthew Lillard) and the governor of New York, Marge McCaffrey (Lili Taylor). Villains and heroes come from unassuming places — and so does backup.

Jessica’s leap to this reboot opens up so many questions, but this show isn’t about her. The former Defender is here to help, so there’s not much catching up we get to do about what’s been happening in her life since 2019 (when Jessica Jones aired its series finale). Believe me, there are questions. Despite that, it’s good to see someone on Matt’s side when we don’t know Frank Castle’s current whereabouts. 

D’Onofrio and Cox bring their undeniable gravitas to the screen once more, with their characters’ complicated dynamic setting the tone. Stunt-work and camera shots show off Daredevil’s nimbleness and sharp auditory skills, along with Kingpin’s brawn. 

Yet Deborah Ann Woll, Wilson Bethel (as Bullseye, aka Benjamin “Dex” Poindexter), and Ayelet Zurer (who plays Vanessa Fisk) nail scene-stealing performances that ramp up the narrative’s intensity. There’s a diner sequence involving Bullseye that I haven’t been able to get out of my head, a reminder of how the fight choreography and cinematography in this series complement each character to a tee.  

Kick-butt Karen is on a warpath, Vanessa is scheming alongside her husband, and Fisk faces new challenges to his power outside of Matt. That doesn’t mean Kingpin isn’t surrounded by loyalists like Daniel, Buck and Dr. Heather Glenn, whose story arcs take interesting turns. As the tension and action unfold throughout every episode, a lot of people end up hurt, deceived or dead. An unbelievably gruesome season finale is the R-rated delight you expect from Daredevil: Born Again.

Where season 1 dove into Fisk and Matt’s darkest natures, season 2 examines whether redemption and true justice can exist. Pay attention: even Daredevil’s armor reflects the story. The show still has its imperfections; a few minor details about the crime at the center of the plot are inconsistent. And Heather’s storyline gets a little weird, but maybe this will pay off in season 3. 

Until then, enjoy wincing at bloody scenes, DIY sleuthing and the jaw-dropping surprises that Daredevil: Born Again serves up in season 2. And if you have time to check out the Marvel shows that were originally on Netflix, I think you’ll appreciate this season, all of its Easter eggs and winks at the MCU even more.