Three of the founders of AI hosting company CoreWeave — Brannin McBee, Michael Intrator, and Brian Venturo — just announced yet another planned sale of CoreWeave stock, totalling over a billion dollars’ stock sales for each since the company went public. This means that whatever happens with CoreWeave and the AI bubble … they’ve got theirs. [Bloomberg, archive]
CoreWeave has never been a solid business. It spent the past three years piling debts on debts, because this was the AI revolution! This is the future!
But bubbles pop. CoreWeave’s whole deal was to sell a financially impossible company to people who were fully equipped to know better — but they had powerful incentives to lie to themselves.
CoreWeave was founded as a cryptocurrency miner. Crypto crashed and the company pivoted in 2019 to cloud scientific computing on GPUs.
In mid-2022, CoreWeave started buying heavily into Nvidia GPUs, just in time to catch the wave of the AI bubble later that year. Nvidia owned 5% of CoreWeave and the two were very cosy. By 2023, CoreWeave was putting the Nvidia chips up as collateral to borrow money to buy more Nvidia chips. This booked a lot of revenue for Nvidia.
CoreWeave did its IPO in March 2025. This was the first big public offering of the AI bubble. Ed Zitron wrote up a nice analysis of CoreWeave’s S-1 prospectus: “CoreWeave Is A Time Bomb”. None of the numbers added up to a business that could ever make it out of the AI bubble alive. [Where’s Your Ed At]
CoreWeave’s main business was as a backup provider for the hyperscalers when they couldn’t build out their own data centres fast enough. The didn’t have a lot of revenue and had taken on ghastly and ever-increasing amounts of debt. The company was spending $1.43 to make one dollar. The S-1 only worked if there was massive AI demand coming soon — which was already not showing up.
None of the numbers have improved in the year and a half since. CoreWeave’s debts remain crippling.
CoreWeave’s one weird trick is to get their debt listed as investment grade and sell it to insurance companies. McBee from CoreWeave told the Financial Times that insurance capital was: [FT, archive]
unlimited. Like, that’s all the money in the world that sits in the insurance tranche.
The stock dropped about 15% on the news of McBee giving the game away.
McBee, Intrator, and Venturo have sold a total of $3.3 billion in stock since the IPO. They’d sold another $450 million before the IPO. The company told Bloomberg:
They’re making personal financial planning decisions that any responsible founder with concentrated holdings would make.
Particularly when the concentrated holdings are in a massive and expanding debt bomb that will blow the moment CoreWeave is not getting OpenAI and Anthropic money via Microsoft.
CoreWeave is one of the best funded and best connected AI bubble service companies. Everyone else is going to do worse.
Anthropic have a new scary campfire story on why everything they work on is evil and out to get you! This one is called “AI models may be capable of self-replicating without access to a datacenter.” Here’s the headline claim: [Anthropic]
Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes.
… If small models improve at their historical rate, they’ll match this capability level by the end of this year.
The end of this year.
The paper constructs a scenario. This is otherwise known as making up a sci-fi story about killer robots, then telling the incredibly credulous press that this can happen in the next two months, and then they print it as-is.
Here’s Joseph Menn at the Guardian swallowing this tall tale, hook, line, and sinker: [Guardian]
the red team, as such internal testers are called, said it created a malicious AI program that spread autonomously through internet-connected “smart” televisions and could do the same inside other ordinary devices.
No they did not! The team made up a story about a program that might do that if a hundred other things that don’t exist come true!
Menn also compared this scenario to Harry Potter, because of course he did. Read another book!
Menn spoke to an “expert” who told him:
What differentiates AI worms from typical reproducing viruses is that AI models can learn as they go.
Chatbot models do not learn as they go! That’s why the AI vendors spend a fortune on training and post-training!
Anthropic’s scenario posits an AI network worm — a self-replicating program.
We’ve had computer worms since 1971. Someone else (not Anthropic) built an AI-based worm earlier this year and tested it on an isolated network. This did not propagate the model itself, because chatbot models are quite big. Instead, the test worm took over smaller machines it had access to and ran tiny models on them. [arXiv, PDF]
Anthropic took this earlier experiment and says it means a whole model “could” copy itself from machine to machine. This means it “could” set up a botnet of TV set-top boxes! And spread all by itself! You can’t prove it won’t!
They did run a small model on a network of 15 Android set-top boxes. It ran at a blistering 0.39 tokens per second.
There are a few other steps that would need to be shown for the threat to hold. The bot has to scan the wider internet. It needs to hack into the set-top box. This has to be done with their tiny model, at less than a token per second. Then it has to copy itself.
The researchers don’t bother testing if any of these work. They just say “not our focus.”
It is true that embedded boxes have notoriously awful security. There have been some huge botnets on these insecure boxes. Those botnets run actual code, which is quite a bit smaller and faster.
Those botnets existing does not let you handwave how your evil chatbot could totally do that for sure.
If you can create a model that’s smart enough to hack tiny embedded hardware while it’s running on tiny embedded hardware, and if it can figure out how to hack new small boxes, and if it can do all of this without the researchers helping it along at every stumble, like all the previous AI hacking researchers did … then sure. But you can’t just assume any of that.
Anthropic says “could” a lot in this paper. “Could” and “might” and even “will” are all words that mean “doesn’t”. “Could” is not enough.
And they admit that “doesn’t” is the word for how this works: [Al Jazeera]
Although no current AI model proved to be both powerful and small enough to self-replicate reliably on a smart TV used in testing,
Much like no current slingshot has proven elastic enough to fire a space shuttle. Obviously we just have to scale up.
The models would only have to shrink by 99% and run with a thousandth of the data centre compute they’re made for while still being able to hack into arbitrary gadgets and come up with new hacks. Autonomously. And they’d have to do all of this by the distant magical future era of the end of 2026.
But maybe I’m wrong. Show me! You’ve got ten weeks.
This paper is marketing. Not to you and me, but to credulous government regulators. The current US administration already believes the chatbot is the magical truth machine, and they think Anthropic has the best magical truth machine.
This paper is a message to say: regulate AI. Make rules so only the huge vendors like us can comply. Then the AI can convert NASA to use giant slingshots.
If you have built an application with an interactive map, you know what you are in for. You pick a tile provider such as Google or Mapbox, or a self-hosted stack like GeoServer or Martin, and your frontend requests thousands of small image or vector tiles as the user interacts with the map. A server processes each request: it reads from a database, renders the tile, maybe caches it, and sends it back. That server is infrastructure: it needs to be deployed, monitored, and scaled.
Dank Linux, the Wayland desktop suite built around DankMaterialShell (DMS), has put out its final post on X under that name, bidding farewell to its old identity. Accompanying it is DMS 2.0, codenamed "Heisenberg," which reached GitHub just a few hours ago.
If you didn't already know, the shell is a Quickshell and Go project that stands in for tools like waybar, swaylock, mako, and fuzzel. Around it sit a greeter, a calendar app, a system monitor, and a search service.
It began as a hobby project, with 1.0 shipping in December last year. Since then, newer releases have added niceties like desktop widgets, a new launcher, and spring-based motion.
With that many components in the suite, the project has now put all of them under a new name.
A surprise rebrand
The rebranding covers the whole suite. DMS is now Dux Shell, and the greeter, search tool, system monitor, and calendar are now Dux Greeter, Dux Search, Dux GOP, and Dux Calendar, respectively.
Dank Linux gets renamed Dux Labs, which also acts as the new name for the project as a whole and for every app in the suite.
Dux Files is a standalone file manager that is the one piece that's new instead of being renamed. It offers features like tabs, split views, trash, and network shares. It picks up your DMS theme if the shell is running.
Dux is the suite's mascot, and the name echoes Linux's penguin mascot, Tux. Starting with a D also lets the dms, dcal, and dgop commands keep their names.
The project's reasoning to go through the rebrand is that a name built around one product no longer fits a suite with its separate collection of apps and a plugin community.
Another thing to note is that the earlier release, DMS 1.6, still carried the Dank name, and its roadmap listed a standalone file browser and phone integration, but a rename was never really mentioned.
As for the DMS 2.0 release, it is a packed one that adds a Material 3 Expressive redesign, plugins for the dashboard and dock, and a lock screen with widgets you can arrange yourself.
It also adds Dux Dot, a free-floating companion or replacement for the bar that can turn into notifications, weather, media, and wallpaper views, as well as the control center, dashboard, and launcher.
What now?
The project has already moved from the AvengeMedia umbrella to a new GitHub organization called Dux Labs.
You will find many repositories that used to be known by their Dank names, including dux-shell (previously DankMaterialShell), dux-gop (dgop), dux-search (danksearch), dux-calendar (dankcalendar), and dux-greeter (dank-greeter).
Personally, I preferred the Dank branding. But if the developers feel this is the best approach and the community accepts it, there's no point talking about what was.
There are plenty of online PDF tools that let you edit documents without installing anything on your computer. You just open a website, upload your file, make the changes, and download the edited document.
But there is a catch. You are handing your files over to a third-party service, and you may not always know how it handles your data. Some services may have questionable privacy practices, while others are outright malicious.
In fact, the FBI has warned people about fraudulent online file converters that steal sensitive information and spread malware in the past.
These tools scrape uploaded files for personal details, banking information, email addresses, and passwords, making them particularly risky when dealing with confidential work documents that should've never left an organization's secure environment.
BentoPDF is an open source alternative that advertises itself as a privacy-first tool that processes files directly in your browser instead of uploading them to a server, helping avoid some of the aforementioned privacy risks associated with online PDF tools.
BentoPDF: For Your PDF Needs
homepage ss here
Created by Alam, a developer out of India, BentoPDF traces its origins to an internal tool that was being used for merging and splitting PDFs. Soon after its creation, the lead dev decided to share his work with the world via Reddit, and a few weeks later the code was live on GitHub.
BentoPDF uses WebAssembly, which lets heavy programs run right inside your browser. Its OCR tool, for instance, runs on Tesseract.js, a port of the Tesseract OCR engine. That way, the processing happens on your computer, and nothing is uploaded to a server.
🚀
You don't need to sign up for an account, and there's no web tracking either!
The project is dual-licensed, where the code is provided under the AGPL-3.0, making it free to use and open for anyone looking to fork it. A separate commercial license is available for closed-source use and costs $79 as a one-time purchase.
Alam also publishes two other open source projects, Hyper Compress and Kura. The first is a PDF compression engine, while the latter converts PDFs to standards like PDF/A and PDF/UA.
Anyhow, coming back to the PDF editor, I took it for a test run on Vivaldi, a Chromium-based browser, covering a few tools to see how it holds up in everyday use.
Editing PDFs
The Edit PDF Text tool shows each block of text in the document with a dashed outline, and picking one brings up the formatting sidebar, which houses the usual editing controls.
It has buttons for tweaking the font, size, colors, alignment, and character spacing, along with bold, italic, underline, and strikethrough formatting.
I could also make it fetch all the fonts installed on my Linux system using the System fonts button, and the info button at the bottom was useful for glancing over the metadata of a document.
Once done editing, exporting the file is only a matter of clicking the "Export" button.
The PDF Editor is mainly meant for annotation work, with tools for highlighting, strikethrough, ink, text insertion, and comments. You pick the relevant tool and drag your cursor over a piece of text to apply it.
Adding watermarks to documents is also possible via the Add Watermark tool, which should come in handy for protecting your work or marking a document as sensitive. It offers a lot of flexibility, letting you choose between a text or image type watermark, a single or tiled layout, and granular controls for the pages, text, and visibility.
Common keyboard shortcuts like the ones for copy (Ctrl+C), paste (Ctrl+V), undo (Ctrl+Z), and redo (Ctrl+Y) also work well here.
Easy OCRs
Sometimes a PDF is nothing more than a set of page images, like a scanned report or a document full of charts. You cannot search through it or copy anything out of it because there is no text for your PDF viewer to find.
The solution to that problem is to run OCR on it, which converts the pages into machine-readable text.
OCR PDF handles this in BentoPDF by letting you upload a file, select the relevant language(s) found in the document, and start the process. During my test, I used a two-page brochure for a motorcycle I like to window shop for and left the advanced settings untouched.
It took a while to process the pages, but the output was mostly accurate. Only the very small, lossy text at the bottom of the document tripped it up. I could copy everything to the clipboard or export it as a .txt or searchable PDF file.
File Conversions
BentoPDF has a range of file conversion tools available, and I tested out two of them. PDF to JPG converts each page of a PDF into a JPG image, and using it is quite simple. You upload a PDF file, set the image quality with a slider (the pre-set value is 90%), and click "Convert."
Images to PDF does the reverse, and it accepts an array of formats beyond JPG, including PNG, BMP, HEIC, and PSD.
There are plenty more conversion tools on offer, covering formats like ODP, CSV, EPUB, JSON, WebP, and Markdown, among others.
Quick Compressions
Usually, when handling PDF files that need to be sent via email, its file size matters a lot. Popular email services like Gmail, Proton Mail, Outlook.com, and Yahoo Mail cap attachments at around 20 to 25 MB.
To tackle that, the Compress PDF tool offers two algorithms, "Condense", which is the recommended one, and "Photon," that's tailored for photo-heavy PDFs. Alongside them are four compression levels that run from "Light" to "Extreme.".
For checking it out, I stuck with Condense and only changed the compression level, going with Extreme to see how small the file could could get.
It took quite a while to process, but BentoPDF brought my 120.4 MB PDF down to 28.5 MB, a 76.3% reduction. That is still a little over the usual email limits, but for a 214-page document, it's a good result.
The text and images stayed very readable too, and I only noticed the loss in quality when I zoomed into the images.
Start Editing
There's a lot more BentoPDF has to offer than what I have covered above. Just scroll down on the homepage to see its full list of 100+ PDF tools. There are some security-focused ones that I am sure many of you will like.
If you are looking for the source code, then you should visit the project's GitHub repository, which also contains details on the tech stack that powers BentoPDF.
A file manager is one of the most basic tools a Linux distribution comes with, and desktop environments ship their own, such as Nautilus on GNOME, Dolphin on KDE Plasma, and Nemo on Cinnamon.
It is where people browse folders, move and rename files, preview what is inside them, and clean up any cluttered directories. Thanks to the platform's openness, you can also install more than one file manager on your system and use whichever you prefer.
A new one that has made an appearance recently is Strata, which is tailored for Omarchy and, to some extent, Arch Linux. Other glibc-based distributions might be able to run it, provided they meet its library requirements.
This file manager is written in Rust with GTK4 and released under the MIT license, with its lead developer, Pierre, announcing its first stable release a few hours ago. He claims that in 45 days, the project has picked up over 30 contributors and 700+ commits, with ~300K lines of Rust now in the codebase.
Strata is Here
Sourced from Strata's Git repo.
It can show folders side by side, so opening one reveals its contents in a new pane to the right while its parents stay visible, like the column view in macOS Finder. Icon and list views are available too, and being a keyboard-first app, it supports navigation with the arrow keys or the vi-style h, j, k, and l keys.
Ctrl+K opens a fuzzy search over your Home folder and every mounted local drive, which matches file names and paths but not file contents. The 1.0 release adds exclusions in Settings for skipping folders like node_modules or a single directory.
Tabs are another new addition that let any folder be opened in a new tab, with keyboard support for switching/reordering them, and previously open tabs being restored after a restart.
Strata also stores sort order and icon size per folder, where one folder can be sorted by date with large icons while another stays alphabetical with small ones, and each opens that way the next time.
Pressing Space toggles a preview panel for images, camera RAW files, PDFs, source code, audio, and video.
It includes a diverse range of themes too, and an in-app editor lets you create your own and save them as TOML files. Code and config files also get language-specific icons based on Catppuccin, in brand colors or the theme's accent.
Install it
Strata has an interactive installer for Omarchy and Arch Linux that runs from a single cURL command:
The installer detects the CPU architecture, the glibc version, and whether the system runs Omarchy 3 or 4 or Arch Linux. It installs the latest stable release for the current user in ~/.local/bin.
Optional extras such as a desktop entry, a default folder handler, SMB support, and Omarchy keybindings are also available. You can even feed Omarchy's onboard AI agent a specific prompt to get it installed on your computer.
If you just want to check out Strata before committing to it, the project's homepage has a demo that lets you play around with it. Building from source is also possible.
Log in
