0:12
Porting DOOM to the Casio Loopy
So, you know how the Casio Loopy was this super niche console from the '90s that was basically the opposite of what mainstream gamers were into back then? Well, someone's been working on porting DOOM to it, and it's actually pretty cool. They've got a GitHub repository up where you can check it out, and they've even made a cartridge called the FloopyDrive that lets you play it on the console. The thing is, the Loopy's got a pretty capable processor, but it's really limited by the amount of ROM space it has, so you can only get a few levels of DOOM on it. Still, it's pretty impressive that they got it working at all, and it's even got some nice features like support for the console's built-in printer.
0:57
KDE Plasma 6.8 will finally fix the worst part of Spectacle
Hey, I just read about the upcoming KDE Plasma 6.8 update and I'm excited to share it with you. Apparently, Spectacle, their built-in screenshot tool, has been a bit of a letdown for some users. It's been struggling to capture the entire screen when you're using multiple monitors, which is a pretty basic feature.
The developers have been working on optimizing Spectacle, and it seems they've finally found a solution. They've been experimenting with a new approach to handle the window manager and the screenshotting process, and it's supposed to make a huge difference. This change under the hood will allow Spectacle to capture the entire screen, even when you're using multiple monitors, without any issues.
It's amazing how something as simple as a screenshot tool can be so finicky, but it's great to see the KDE team addressing this issue. I've used Spectacle before and I can see how frustrating it must be to have it fail you when you need it most. This update should make a big difference for anyone who uses multiple monitors.
The best part is that this change is a result of community feedback, which is exactly how open-source software should work. The developers are listening to their users and making improvements based on their needs. It's a great example of how collaboration and community involvement can lead to better software.
I'm looking forward to trying out KDE Plasma 6.8 and seeing how Spectacle performs with the new update.
0:57
10 Million Cubans Suffer Nationwide Blackout - For The Second Time This Week
Cuba went dark again on Friday, a second islandwide outage in just a few days, leaving almost the entire ten‑million‑strong population without electricity. The first blackout hit earlier this week, and now the grid has flickered back on only partially.
What’s striking is how the same aging infrastructure is being stretched to its limits: a mix of deteriorating power plants, chronic fuel shortages, and the lingering effects of an external energy blockade have left the system unable to recover before the next surge. Back‑to‑back nationwide cuts are rare in the Caribbean, which makes this double hit a clear signal of deeper strain.
Officials say crews are working to restore service area by area, but the pattern suggests the grid’s fragility isn’t a one‑off glitch—it’s a systemic problem that will need more than quick fixes to keep the lights on.
2:02
What to know about Vinod Khosla, the Silicon Valley legend whose family is buying the Seahawks
Vinod Khosla, a 71-year-old billionaire venture capitalist, is leading a group that's buying the Seattle Seahawks from the estate of Paul Allen for a reported $9.6 billion. What's interesting is that the Khosla family's involvement goes beyond just Vinod - his wife Neeru is the controlling owner, and their son Neal is expected to have a significant leadership role. Neal's described himself as an "obsessive sports fan" who likes to bring a quantitative and analytical lens to the game.
The Khosla family's approach to ownership is likely to be driven by Vinod's philosophy, which focuses on hiring the right people and getting out of their way. He's known for his emphasis on talent and his willingness to take risks. In fact, he's said that the single most important decision a startup founder will make is the team they build. This could be a departure from the Allen estate's more hands-off approach under Jody Allen.
One thing that might be a challenge for the Khosla family is building a connection to the city and the Seahawks fans. Vinod's lived and worked in Silicon Valley for decades, and the family's ties to the Bay Area are strong. They've invested in several Seattle-based companies, but they don't have the same level of roots as the Allen family.
Vinod's also known for his brutal honesty and his willingness to take unpopular positions. He's been involved in a high-profile dispute over a private beach in California, and he's been criticized for putting his own interests above the public's. However, he's also demonstrated a willingness to take risks and push boundaries - qualities that might serve him well as an NFL owner.
Overall, the Khosla family's involvement in the Seahawks is likely to bring a new dynamic to the team, and it'll be interesting to see how they approach ownership and decision-making.
2:21
Trying to Use a 2007 Samsung UMPC as Wii U Gamepad
The most interesting part is that the Wii U Gamepad protocol has finally been opened up in software, so any tablet that can run Linux can try to masquerade as the controller. The Samsung Q1 UMPC, a 2007 Celeron M device, gets a fresh 32‑bit Debian 12 install and the Vanilla Wii U Gamepad stack, but the real snag is the Wi‑Fi link: the protocol uses a quirky handshake that only a few adapters support, and the TP‑Link dongle they tried switched from a Mediatek to a Realtek chip, forcing a manual driver build.
Even after the Wi‑Fi hurdle, the hardware hits a wall. The Celeron M, lacking any H.264 decode block, is forced to handle the compressed HDMI stream in software, and it maxes out at 100 % CPU with noticeable frame loss. The integrated GMA 900 only accelerates MPEG‑2, so the video pipeline stalls.
Bottom line: the reverse‑engineered stack works, but the old UMPC’s modest CPU and limited GPU acceleration make it a poor stand‑in for the official Gamepad. It’s a neat proof‑of‑concept, yet the performance ceiling is set by the 2007 hardware.
3:47
Meta Removes Controversial AI Feature On Instagram After Backlash
Meta pulled the plug on an Instagram tool that let anyone remix public photos with AI, and it wasn’t even warning the owners that their images were being altered. The feature ran silently, so a picture could be tweaked and reposted without the original creator ever knowing. Users started noticing weirdly polished versions of their posts, and talent agencies—CAA among them—raised eyebrows about how the tech could be weaponized for image‑driven negotiations. Under that pressure, Meta posted a brief note saying the capability was being retired. The move feels like a quick course correction after the backlash turned into a full‑blown conversation about consent and AI.
4:39
AI Automations: The Complete Guide + Free Resources
I heard a tiny automation catch a duplicate invoice at 7:12 am, pull the vendor, number, amount and due date, compare it to a spreadsheet, and leave a short note for me. It didn’t delete anything or ping anyone; it just handed me the evidence in one spot, and I decided in under a minute.
What struck me was the whole workflow moving itself—no manual copy‑paste, no hopping between apps. The automation noticed the email, gathered the right context, ran a rule‑plus‑AI check, and produced a clean handoff ready for a human glance.
The guide boils that down to a simple test: can the system run unattended and leave something useful and reviewable? If it can, you’ve got a dependable automation, even if it’s just an inbox triage.
From there you pick a narrow task, map the steps, add safety checks, and chain the next handoff. The result is a series of boringly reliable pieces that keep your day moving without the hype of a “full‑AI employee.”
4:39
Google releases Gemini 3.0 Pro, arguably the best AI system yet
I’ve been digging into Gemini 3.0 Pro’s guts and the most striking shift is the MoE transformer that now only fires a slice of its parameters per token, so the model’s capacity can grow without the usual speed hit. Google kept the architecture closed, but the pre‑training jump they reported between 2.5 and 3.0 feels massive—no sign of the scaling wall people were betting on.
The new version pushes the context window to a million tokens and can spit out up to 64 k tokens, which means it can handle whole codebases or long‑form drafts in one go. What’s cooler is the native multimodal handling: text, audio, video, and images are processed together, so it can, for example, read a handwritten recipe and turn it into a bilingual step‑by‑step guide without stitching separate models.
Reasoning gets a boost from the “Deep Think” mode, which runs multiple solution paths at test time and picks the best one, nudging benchmark scores higher—especially on complex exams and GPQA‑Diamond. The trade‑off is a higher hallucination rate; it’s still more accurate than rivals, but it can be confidently wrong more often.
On the developer side, the model is already baked into Google’s Search, Gemini app, and AI Studio, and it’s being used for the new “Antigravity” agent platform that can code, run, and verify software autonomously. The pricing is steeper than some competitors, but the token efficiency and TPU‑optimized speed keep it competitive for large‑scale use.
4:39
Is Nano Banana Pro the ChatGPT moment of image generation models?
Google’s new Nano Banana Pro sits on the Gemini 3 Pro backbone, using a sparse mixture‑of‑experts transformer that only lights up the parts it needs for each token. That lets it juggle a massive capacity while keeping each generation cheap enough to run in real time. Inside, the model drafts “thought images” first, refining composition before spitting out the final 1K, 2K or 4K picture, and it can chew through a million‑token context that mixes text and images.
What feels different is the live‑search hook: the system can pull current data—weather, news, even NASA updates—and bake that info straight into the visual output, complete with legible multilingual text. It also keeps a subtle SynthID watermark hidden in the pixels, while offering a verification tool for anyone who wants to check authenticity.
In practice, the model handles tricky prompts that used to trip other generators—like a gazelle chasing a cheetah or a horse riding an astronaut—and can edit a single image across multiple steps, preserving characters and adding new details. It’s already rolling out in the Gemini app, Google Slides, Ads and via an API that charges a few cents per 4K image, opening the door for professionals to turn complex ideas into polished visuals without the usual guesswork.
5:40
What we know about Yann LeCun vision for the future of AI
Last week, news broke that Yann LeCun, Turin Award winner and one of the pioneers of modern artificial intelligence, is leaving his role as Meta’s chief AI scientist by the end of the year. LeCun will be starting a new AI startup with details to come later.
There were already rumors of LeCun being sidelined after Meta CEO Mark Zuckerberg assembled his new Superintelligence Lab, headed by Alexandr Wang, co-founder and former CEO of Scale AI. However, LeCun had already made clear long before his departure that he was not satisfied with the direction that the AI community is headed.
While most efforts are going into large language models (LLMs), LeCun has been very vocal about their limitations, particularly in their ability to solve real-world problems.
“We are not going to get to human-level AI just by scaling LLMs,” LeCun told Alex Kantrowitz’s Big Technology podcast in May, calling LLMs systems with gigantic memory and retrieval ability, “not a system that can invent solutions to new problems.”
But which direction will LeCun be going after his departure from Meta? We already know some of the areas that he is interested in and that the industry has mostly ignored.
In a post on LinkedIn in which he confirmed his departure, LeCun wrote: “I am creating a startup company to continue the Advanced Machine Intelligence research program (AMI) I have been pursuing over the last several years with colleagues at FAIR, at NYU, and beyond. The goal of the startup is to bring about the next big revolution in AI: systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences.”
LeCun has been consistent on how he believes we will achieve these goals. He has been a long-time advocate of self-supervised learning, and in recent years, he has been working on “world models” that can be trained through self-supervised learning.
In short, self-supervised learning is a process in which the model trains itself without the need for data labeled by humans. LeCun has been talking about self-supervised learning for more than a decade.
“If artificial intelligence is a cake, self-supervised learning is the bulk of the cake,” LeCun said at the AAAI conference in 2020. “The next revolution in AI will not be supervised, nor purely reinforced.”
LLMs do some form of self-supervised learning, where they’re given a sequence of text tokens (or other type of data) and told to predict the next token. (Since the next token is already present in the training corpus, there is no need for manual labeling, thus it is considered some form of self-supervised learning.)
However, LeCun (as well as other scientists such as Richard Sutton, another Turing Award winner), argue that LLMs don’t learn in the same way that humans do. In particular, LLMs have only been trained on data generated by humans (i.e., text) and can’t generalize to the unpredictability of the real world. They do not learn how the world works from sensory data like children do by observing and interacting with the world, discovering and internalizing things such as gravity, depth, object permanency, etc.
So they cannot predict consequences and counterfactuals the same way that humans or even animals do. And next-token prediction does not seem to generalize well to the unpredictable nature of the real world, which is partly evident in how much training data they need to learn basic tasks (while still failing to generalize beyond their training distribution).
So, if not LLMs, then what?
