AI Used in Linux Kernel, Android Limits Sideloading,new Desktop Environment  ## Metadata - Author: [[Linux & Open Source News]] - Full Title: AI Used in Linux Kernel, Android Limits Sideloading,new Desktop Environment - Category: #podcasts - URL: https://share.snipd.com/episode/9ea4d50b-4c95-499d-aa49-c527c5983088 ## Document Notes ## Highlights - **AI Assists Kernel Backport Triage** - AI can help shortlist kernel patches for LTS backports by analyzing commits and code context. - Humans still must make final decisions because AI recommendations are experimental and imperfect. Transcript: Nick So it looks like AI tools are starting to be used around the Linux kernel. Not to automate writing code, mind you, at least I hope not, but to determine which patches should be backported to older kernel versions, notably LTS kernels. Sasha Levin, the co-maintainer for the LTS kernel, wrote some documentation about AI coding assistance, and they're apparently trying out using LLMs to see which patches should, In fact, be backported. Now, technically, developers who write a patch for the kernel and they want that patch to be backported should mark that patch as CC stable. But to identify which patch among all the unmarked ones could be backported, it looks like AI could be a tool that helps. The AI just basically analyzes the patch and lists reasons why this should or should not be backported. They gave an example that, for example, the patch that was analyzed fixes a user-facing issue and also fixes a workaround that did not function properly in all cases. It seems the AI is just looking at the commit messages and also parts of the code to make that recommendation. And of course, it remains to be seen if these recommendations are followed or not, and if any kind of automation is done around that. It seems to just be an experiment. ([Time 0:02:42](https://share.snipd.com/snip/9c053e83-a102-463d-a066-4fd970cd8aa4)) - **Gershwin Desktop Previewed** - GhostBSD introduced Gershwin Desktop, a macOS-like environment built on GNUstep and XFCE components. - Nick tried it in a VM but it failed to start an X11 session, so he plans to test on real hardware later. Transcript: Nick Now, this week was also introduced Gershwin, which is a desktop environment that provides a macOS 10-like interface. It is apparently being proposed by GhostBSD in their latest release. It seems to be a project mainly developed by GhostBSD themselves. It is based on GNU Step, which if you're as old as me, you probably heard about it. It is kind of similar to Next Step, which is the initial foundation for macOS and follows the Cocoa API very closely, this being the APIs that Apple used to use or maybe still relies upon To build macOS apps. This thing even sort of looks like macOS with the usual UI paradigm of the global menu bar with a dock and a theme that evokes Mac. It is just an early preview of the final desktop, obviously it is not finished. I tried to give it a run in a virtual machine but it did not even want to start an X11 session in there so I couldn't really give it a shot. Gershwin does not support Wayland yet, although Gnustep has experimental support so it will likely come. They use XFCE's window manager but they will move to their own once they write one. And apparently there's also development happening around new components, new apps, a new dock, new frameworks to make building apps for this desktop a bit nicer, and also to have a Unified interface paradigm. They like an app store. You can install apps from other desktops if you want. Just remember that this is currently running on GhostBSD. You might not have the usual Linux apps you rely upon. Most Linux apps seem to have ports for BSD, but maybe not everything. Now, once it is stabilized a little bit, I'll make a video about it. I always enjoyed GNU Step as just a concept or an idea. I actually used it way back when, but I think it's fun. So I'll give it a shot. Obviously, right now, it just did not even want to run in a VM, but maybe it would on actual real hardware. So once I have a little bit of time, I'll give it a shot. ([Time 0:05:04](https://share.snipd.com/snip/e9e859ac-f1f3-4d87-8cda-3ef12e034776)) - **Bazaar Reinvents Flatpak App Store** - Bazaar is a new Flatpak-focused app store with a fast, multi-pane GNOME UI and decoupled backend. - It prioritizes screenshots, queueing, and quick performance for Flatpak installs from Flathub. Transcript: Nick Now, this week, we also saw the release of Bazaar, not the version control system that Ubuntu relied upon for a while, but the Flatpak App Store. It's been in development for a while from the developers of Bazaar, if I remember correctly, and it's basically what GNOME software could be if it had more features, a better design, And better performance. It is a GNOME app through and through, but it is intended to support Flatpak only, specifically FlatHub, but other remotes will work as well. Of course, you can use it to browse and install applications, but it also has a very nice layout with screenshots being put front and center in their own pane in the interface if the window Is large enough. You've got links being made very visible so you can support app developers, check out the code or contribute. You also have really, really solid performance. It starts really fast, it loads apps really quickly. It is, as they say, highly multi-threaded. The UI is completely decoupled from the backend. You can queue multiple downloads. You have a visual representation of that install queue, it has a Gnome shell search provider so you can look for apps from the Gnome overview, it has that multi-pane UI with search results, The app page, the screenshots and the queue visible at the same time if your window is big enough, it seems to support accent colors, you also have a nice prominent updates button in the Top left corner, it can display the apps that you have installed, it is just fantastic. ([Time 0:07:09](https://share.snipd.com/snip/9b495feb-c321-41fb-85ed-53ad5b05b83a)) - **Firefox Web Apps Are Superficial** - Firefox's returned web app feature is limited and Windows-only, lacking true PWA support and integration. - Mozilla has no plans to implement full desktop PWA features like separate windows, notifications, or file handlers. Transcript: Nick Now this week, I briefly got very happy before I got really sad again, because Firefox supposedly was gaining web app support back. Problem is, yes, that support exists in Firefox 142 if you enable it in the About config page, but first, it is only for Windows, second it's not really what you'd want for supporting Web apps. Basically what this property does if you enable it is it just adds an icon in the address bar and when you click it, it adds the website to your taskbar supposedly as a web app. The problem is these are not supported as full progressive web apps and Mozilla said they have no plans to support that. So you will still get the entire browser UI around the application, it's not opening in its own individual window, you don't get notification support, you don't get file handlers, You don't get system workers, you don't get updates in the background, you don't get badges for unread stuff. It is a pretty bad web app implementation and it seems like it's on purpose, which I don't quite understand. Other browsers and even some external apps do a much better job of supporting progressive web apps. I really don't see the use case for adding this feature back. It is not what people want when they say they want web apps to be integrated with their desktop. When you want a web app, you want an app. You want the notifications, you want its individual window, not a full web browser, because if you wanted that, you would just create a shortcut. You don't need a specific dedicated feature. ([Time 0:08:53](https://share.snipd.com/snip/cbb22e91-5f78-4d80-bbff-3e0524409bba)) - **Choose Browsers That Reject AI Hype** - Consider Vivaldi if you prefer browsers that avoid LLM-driven features and summaries. - Vivaldi explicitly refuses to add chatbots or AI summaries until implementations are rigorous. Transcript: Nick Now, still on the topic of browsers, Vivaldi took a stance this week, saying they would not start competing on the AI front. As they say, they chose humans over hype, and they do not want to turn what they call the joy of exploring, which I suppose refers to browsing the web, into inactive spectatorship, probably Referring to the pre-digested stuff that AI gives you. This means they will not be adding any AI feature into Vivaldi, not a chatbot, nothing based on an LLM, no summaries, no suggestions or anything of that ilk. Now, they still say that machine learning is an exciting field and it can be. It may also lead to useful features. I agree. But they also say that they do not want to add to the pile of misinformation. And so they will not add any AI related stuff until a more rigorous way to implement these tools is available. So basically their stance is it's clearly not ready, it's low-quality tech right now, so we're not looking at it for the moment. And honestly, this position could be enough to sway me towards Vivaldi. It's Chromium-based, which sucks. ([Time 0:10:37](https://share.snipd.com/snip/b0a85d34-970a-4a3c-a5eb-4a359bfae09a)) - **SuperTuxKart Team Splits** - SuperTuxKart development stalled due to internal conflict over release control and infrastructure ownership. - Displaced developers started SuperTuxKart Evolution to continue the project with new infrastructure. Transcript: Nick Contributing to open source and working with a community can be sort of a minefield and issues can arise in the weirdest or most unexpected of projects. And this week, it is in Super Tux Kart, the venerable kart racing game. Their latest version had completely stalled development-wise, and they published a post explaining why. And it turns out it's the usual disagreements inside of the dev team. Turns out most of the developers just wanted to push a brand new version that had been ready since April. The problem is one of the co-leaders of the project, who owns the entire infrastructure, the domain name, the website, the ability to push new releases, they did not want to publish That version, they wanted to add more Vulkan stuff, but they didn't pretty write it all that fast, and so the release was delayed since April. At some point the other developers got just fed up and decided they would wrest control away from that person so they kicked them out, them and another developer who seemed to be taking The same stance as the co-leader of the project. Issue is they don't control the domain name, they don't control the main website, they can't push updates to the game in its current form and so what they have to do is start over with an Entire new infrastructure. So they started a new project called SuperTux Cart Evolution. It is not a fork, it is the continuation of the official SuperTux project. They just are exiles from their own stuff. ([Time 0:12:20](https://share.snipd.com/snip/83423813-bc34-44bd-a7c7-87ff62d121c0)) - **EU's Tech Dependence Reveals Opportunity** - Proton reports heavy EU dependence on US tech, with most countries above 50% reliance on US services. - Greater EU investment in open source and Linux would boost autonomy and ecosystem growth. Transcript: Nick Now on the topic of digital sovereignty, Proton, the email, not the gaming translation layer, they published a little report on the EU's reliance on US companies and big tech, much Like what Nextcloud did recently. And the conclusions are not really good. Over 74% of all publicly listed EU companies depend on US-based tech services, whether it be Google, Microsoft, or something else. The countries that are worse off are Iceland, Norway, Ireland, Finland, and Sweden, all above 90%, meaning 90% of their publicly listed companies use some US-based services. The best case in the EU is Bulgaria with 16%, but things quickly escalate. Out of the 31 countries analyzed, only Bulgaria, Romania, Slovakia, and Poland are under the 50% mark. So basically, in almost all EU countries, half of the companies, or way more than that, rely on big tech. Obviously, this is not a good situation to be in in our current world, where pressures are exerted by governments on other governments through their companies, through the use of tariffs, As through regulation being applied, or the threats of sanctions and shutting down services. Proton highlighted a few case studies, with France sitting at 66%, Despite speaking some big words about strategic autonomy and actually having autonomy in a lot of fields, France Simply is not independent tech-wise. The UK is at 88%, completely dependent on the US, and Ireland, which is the home to most EU's installations of big tech companies, it stands at 93%. Now personally, without going into the political concerns of geopolitics and interdependency between countries, why I want the EU to be completely independent from big tech is mostly Because the only way they can do that is by supporting open source and Linux. So if they go fully independent, if they build their own stack, it means they're investing in Linux and open source, which would make all of that stack go way bigger, have way more users, Way more investment, and grow much larger, much faster, which is exactly what I want for our community. ([Time 0:14:08](https://share.snipd.com/snip/d237bf76-6b7b-44fb-b7eb-367c40a54e03)) - **Google Tightens Sideloading Controls** - Google will require APKs and developer identities to be verified to sideload apps, rolling out tests in October and full deployment by 2026–27. - This adds a control point where Google could block or track third-party app distribution and developers. Transcript: Nick Google decided to block side loading of unverified Android apps starting next year. Basically, if you want to install an APK on your Android smartphone, that APK needs to be verified by Google, or at least its developer need to be registered and identified with Google. This might seem innocuous, but it does mean Google has control over who can make an application that users can run and who cannot make an app. Packages and signing keys will need to be registered with the Android Developer Console, which means Google can potentially deny a developer entry entirely or revoke permissions, So updates couldn't be installed and new apps made by that developer could not be distributed. They will start testing this in October. A full deployment is planned for March 2026, and the full international deployment of this is for 2027. This is probably a move to handle the fact that they will have to let users download third-party app stores from the Google Play Store. App installs from these stores will be sideloaded, much like when you install something from FDroid, it is a side-load of an APK, and Google will want to keep some amount of control over That. And so they will add that layer of security on top, which in itself isn't bad, but it does mean Google could do some nefarious things. It doesn't mean they will, but it is Google, so they probably will. At the very least, it means that your device, when you're trying to install an APK, will call to a Google server to check if that APK is actually certified, which means Google will at the Very least get your IP address and probably some more identifiers to track who installed what, which is something that is profoundly undesirable, in my opinion, but it also means Google Could just block some developers for any reason they want, and then you have to settle that in court. ([Time 0:16:42](https://share.snipd.com/snip/0a32783f-72f0-4b86-82f8-b6f114dfbd41)) - **Asahi Developer Moves To Intel** - Asahi developer Alisa Rosenzweig left after completing M1/M2 GPU support, moving to work on Intel drivers. - She felt she accomplished her goals and will continue contributing to Linux in a new role. Transcript: Nick Now, another developer leaves the Azahi project this week. Alisa Rosenzweig, one of the most prominent developers on the GPU side of things, decided to step down from the project. Now it's nothing dramatic, no big drama this time, it's just that she thinks that she's done what she came for. With the M1 and M2 GPU code now in solid shape for the project, being conformant with Vulkan and OpenGL, they just did what they came to do. Other developers of course remain on the Azahi project, but they did lose Hector Martin, Azahi Lina, and now Elisa, but they were the most prominent developers for that project. There are apparently some big challenges as well remaining for implementing support for the next iterations of the Apple Silicon architecture. M3 and M4 aren't really that well supported right now. Alisa will be a fantastic help for Linux still though because she's going to be working at Intel on their open source Linux drivers. Now totally normal at some point when you're doing volunteer work or even contract work when you're working on something, at some point you get bored or you feel that you've accomplished Everything that you set out to accomplish. And moving on to other things, totally normal. And Alyssa will be a fantastic help for the Intel drivers because she's done a fantastic work on the Apple Silicon side of things. So I'm sure she'll be able to do the exact same on Intel. ([Time 0:18:44](https://share.snipd.com/snip/b73783b6-b8e2-439e-8446-1a903b465d5a)) - **NVIDIA Releases Massive Multilingual Dataset** - NVIDIA open-sourced Granary: ~1M hours of voice data across 25 European languages plus translation and speech models. - The dataset and models can empower open-source speech, translation, and assistant projects if rights are cleared. Transcript: Nick Well, this is the part when we do that. First, something I missed last week, but NVIDIA opened a gigantic data set of 25 European languages. They call that thing Granary, and it's an open source database of about a million hours of recorded audio in these 25 languages. And it can be used either for training AIs to better understand, to better translate, to better recognize, or even to better generate speech in these languages. It could also be very useful for any kind of open source assistant, any form of open source voice control or translation app or language learning app, you name it. Among supported languages, you'll find French, German, Greek, English, Spanish, Estonian, Croatian, Italian, Lithuanian, Maltese, Portuguese, Romanian, Slovak, and a bunch Of others, including Ukrainian and Russian. NVIDIA also released Canary, which is the AI model that they used with this dataset for inter-language translation, and they also released another model that they built for turning Speech into text for any of these languages. Now, of course, who knows how this dataset was sourced initially, but if they're open sourcing it fully with the entire models and their weights, which seems to be what they're doing Right now, I would assume they have the rights to all of it, because opening up something that you have no rights to seems pretty shaky legally so i'm pretty sure that if they're opening This they actually have the rights to use all that data and it could be useful not just for nvidia projects but for probably a bunch of other stuff so i mean my big problem with ai is that The data is never really ethically sourced and this seems to be the case here so that's good. ([Time 0:20:27](https://share.snipd.com/snip/1b65886c-93f1-46db-b7a8-35e162c6e4af)) - **Anthropic Settlement Leaves Legal Questions** - Anthropic settled a major lawsuit to avoid catastrophic damages, leaving copyright training questions unresolved. - Settlements reduce legal precedents, so copyright-by-training remains legally unsettled. Transcript: Nick Now if you remember the big lawsuit against Anthropic that could have turned into a class action that would really hurt most AI companies, well this lawsuit is now probably over. Anthropic settled it. Instead of trying to fight it and risk what was estimated to be up to 900 billion dollars in total damages, they decided to settle. The details will be finalized in early September and the deal will of course need to be approved. Anthropic had tried to get the suit's class action classification dismissed to no avail and since Anthropic apparently makes 5 billion dollars in revenue, which is insane to me how Much money they could make because they have nothing as a product, but they also make billions of dollars of losses, they probably did not want to risk that thing going in front of a jury. Of course this does not solve anything for people wondering if AI training is copyright infringement because, well, no decision was reached. If the company decided to settle, it is a clear indication that they didn't really think they would win in front of a jury or at least the risk was just too great. The company still has other suits to contend with at the same time against music publishers and Reddit for the exact same reasons. They allegedly used copyrighted works without authorization to train their large language model. And I say allegedly, but we all know that's exactly what they did. ([Time 0:22:12](https://share.snipd.com/snip/0d045fbe-2c32-4c3b-99ec-bddb9b7bd29e)) - **Don't Use LLMs For Fact-Checking** - Avoid relying on current LLMs for fact-checking or content verification because they hallucinate and lack true factual grounding. - Treat AI outputs as starting points and always validate facts with human review and reliable sources. Transcript: Nick Now, another thing I had missed previously, Wikipedia editors decided to refuse the use of AI to review articles. Wikipedia's co-founder had proposed a test to rely on more AI tools to review various articles and make sure they followed Wikipedia's policy. The problem is the test use case using ChatGPT failed to adhere to Wikipedia's policies, notably on neutrality, on verifiability, and on sourcing, which should surprise no one. Jimmy Wales, the co-founder, shared their ChatGPT experiment in forums, but Wikipedia editors quickly realized that this article was full of hallucinations, it misrepresented Facts, and it was riddled with mistakes. It seems Wikipedia will thus remain free of AI trash for now, as they've also added policies aimed at removing AI written articles very quickly without lengthy debates. Basically, if it's clear that it's AI generated, they're going to remove it, they're not going to take the time to look at it because it's wasted time because it's obviously going to Be utterly false. They also rejected AI-generated articles entirely after a short trial that also presented AI summaries to some viewers. So fortunately, Wikipedia, as flawed as it can sometimes be, will at least be free of mass hallucinations and will just remain full of human-made mistakes and biases. Now, it should be obvious to anyone with at least a modicum of experience with any current AI tools that using AI to fact-check stuff is the stupidest idea because that's literally the Use case where AI cannot be used. It has no concept of what is true, what is biased or not, and it will change its answer depending on how you phrase your question. ([Time 0:23:50](https://share.snipd.com/snip/61f92548-f7fc-476f-9b73-676c740f4070))