Google finally realized 90% of the world doesn't speak English and maybe that's bad for business. NotebookLM now supports 80 languages for its video generation feature, which means you can finally get those AI-generated summaries without everything sounding like a Google Translate fever dream.

The video feature launched in July 2025 as English-only because of course it did. Non-English speakers have been stuck with garbage basic summaries while English users got the full experience with proper video presentations. About time Google fixed this.

Here's what actually changed: the audio summaries now work properly in other languages instead of giving you the abbreviated crap they served up before. You used to get maybe 2 minutes of basic overview if you were lucky enough to not speak English. Now you can actually get detailed breakdowns in French, German, Spanish, Japanese, and 76 other languages.

The rollout started August 25 and should hit everyone within a week. Whether the voice actually sounds natural in other languages or like robot garbage remains to be seen - Google's track record with non-English AI voices is mixed at best.

For anyone who's been waiting for NotebookLM to support their language since day one, this is actually useful. No more feeding everything through Google Translate first just to get a decent summary of your research documents. The multilingual support includes the full video generation pipeline, not just basic text summaries.

The NotebookLM app also got updated with these language features, and early users report mixed results on social media. Some languages like Spanish and French work well, while others still sound robotic.

Google's AI language models have been improving, but generating natural-sounding speech in 80 different languages is no joke. Implementation quality varies significantly by language family.

NotebookLM can finally make videos in other languages without sounding like complete garbage. You feed it research materials in one language, and it generates video presentations in any of 80 languages without losing the actual meaning.

It's not just running everything through Google Translate and praying. The AI actually understands what your documents are about, figures out the key concepts, then rebuilds the explanation in ways that make sense for each language. No more getting a German video that sounds like it was written by someone who learned German from a phrase book.

The audio overviews also got better - they used to give non-English users the cliffnotes version while English users got the full explanation. Now everyone gets the same detailed content, which is what should have happened from day one.

Universities with international students can stop hiring separate content teams for each language. I've been using NotebookLM since launch and the English-only thing was annoying as hell - half my research materials aren't in English anyway. My university tried this with Korean language materials - worked great for engineering docs, terrible for literature analysis. Universities are already testing this for their international programs.

For businesses, this means no more paying translation agencies to localize training content. Your marketing team can feed English materials into NotebookLM and get decent training videos in Spanish, French, or whatever languages your employees actually speak. The voice quality is inconsistent - some languages sound great, others sound robotic - but it beats hiring voice actors for every language.

This follows the usual pattern where Google finally realized 90% of the world doesn't speak English and maybe that's bad for business. OpenAI and Microsoft are doing the same thing - everyone's racing to support more languages because English-only AI tools are leaving money on the table.

Google is rolling this out globally by early September, which suggests they're confident it won't completely fall apart. I tested it with Japanese research papers and it actually understood the context instead of just translating word-by-word. But Czech? Sounds like a drunk text-to-speech engine. Early beta users say the major languages work well, but expect the less common ones to sound like garbage.

NVIDIA dropped Jetson Thor today - basically server-class AI compute that fits in a robot. The specs are actually impressive for once instead of the usual NVIDIA marketing hype about "transforming the world."

Here's what matters: 2,070 FP4 teraflops and 128GB of memory in a 130-watt power envelope. That's 7.5x more compute than Jetson Orin without melting your robot's chassis. Finally, robots that can think and move at the same time without needing cloud connectivity to do basic math.

The power efficiency gains are actually meaningful - 3.5x better than Orin means your robot won't die after 20 minutes of operation. Anyone who's worked with Jetson Orin knows it drinks power like a data center and runs hot enough to fry eggs.

Agility Robotics is putting it in Digit v6, which makes sense since their current robots spend half their time waiting for the cloud to tell them not to walk into walls. Boston Dynamics is also jumping on it for Atlas, which probably needs the compute power just to keep those creepy dance moves synchronized.

The real win here is ditching cloud dependency for AI inference. Previous Jetson boards could barely run one model without choking, forcing everything through AWS with lovely 200ms latency. Thor can allegedly run multiple AI models simultaneously, though NVIDIA's performance claims always need real-world testing before you believe them. Remember when they promised Jetson TX2 would change everything? Still waiting.

Pricing starts at $3,499 for the developer kit, which is expensive as hell but actually reasonable for NVIDIA standards. The production T5000 modules will cost less once they ship, but expect months of availability hell. Good luck getting these before 2026 - NVIDIA's idea of "available now" is like Valve time.

The Isaac robotics platform gets major updates with Thor support, including better simulation tools and pre-trained models. The robotics community is already complaining about supply chain issues and wondering when normal developers can actually buy these things.

Early benchmarks from robotics startups show the performance gains are real, unlike NVIDIA's usual marketing fluff. Boston Dynamics is integrating it into Atlas robots, which says something about the compute requirements for modern humanoid robotics. Still better than trying to cram a server rack into a robot chassis.

Jetson Thor means robotics companies might finally be able to build robots that work in the real world. Previous Jetson hardware was either too slow for real-time AI or drank power like a data center. This might actually fix both problems.

Carnegie Mellon's AirLab is already experimenting with this for search-and-rescue robots. Their MAC-VO model works on current Jetson hardware but struggles with complex scenes. Thor should fix the performance bottlenecks.

The applications are everywhere: surgical robots that don't need cloud connectivity, farm equipment that can identify individual plants, delivery robots that work when WiFi is garbage. The key is running AI models locally without sending everything to the cloud, which is what current hardware struggles with.

The usual hardware companies are scrambling to build dev boards for this. Advantech, Aetina, and a bunch of others are working on carrier boards. Good luck getting consistent hardware availability from NVIDIA though - Jetson hardware has always been in short supply.

Camera and sensor companies are building Thor-compatible products. The interesting part is NVIDIA's Holoscan sensor bridge, which streams data directly to GPU memory instead of going through the CPU first. This should reduce latency, assuming it actually works as advertised.

The $3,499 dev kit is expensive but not insane for research hardware. Production modules are $2,999 each for orders over 1,000 units, which puts the real cost around $4,000-5,000 once you add carrier boards and everything else. NVIDIA's marketing is usually hype but these specs look legit.

Some software companies are already jumping on this: Openzeka for visual AI agents, Rebotnix for their robot platform, and others. The real test is whether anyone can actually buy these things in volume - NVIDIA has a track record of announcing hardware that's perpetually "coming soon."

Edge users discovered today that their ad blocker suddenly became garbage. Microsoft automatically swapped uBlock Origin for "uBlock Origin Lite," which blocks basic ads but lets tracking garbage through like a screen door.

Users on GitHub are pissed after their extensions got force-updated to version 2025.824.1755 without warning. One day you have proper ad blocking, next day you're seeing ads on YouTube again and wondering what the hell happened.

Microsoft finally caved to Google's Manifest V3 war on ad blockers. So much for Edge being the "better Chrome" - turns out they were just waiting longer to screw their users. While Chrome killed proper ad blocking months ago, Edge let people believe they cared about user choice. Spoiler alert: they don't.

uBlock Origin Lite is like diet ad blocking - technically blocking but not really. No dynamic filtering, no advanced rules, no customization. It's what Google wants ad blockers to be: mostly useless but technically present so they can say they support user choice.

The original uBlock Origin disappeared from Edge's store, though you can still install it in developer mode if you enjoy jumping through hoops. Developer mode installation means Enterprise IT will block it, normal users won't figure it out, and Microsoft gets to pretend they still support power users.

Raymond Hill, the uBlock Origin developer, has been calling out Manifest V3's garbage design for years. This isn't about security or performance - it's about making ad blockers ineffective while maintaining plausible deniability.

Anyone who switched to Edge specifically for better ad blocking compatibility just got played. The uBlock Origin community is livid, and Raymond Hill's GitHub shows constant complaints about Manifest V3 limitations.

Firefox still supports the real uBlock Origin because Mozilla actually gives a shit about user choice. The Chrome Web Store neutered the extension months ago, and now Edge followed suit.

Other alternatives like AdBlock Plus play nice with Google's rules, which means they don't block much. Brave browser's built-in blocking works better, but means switching browsers entirely. Time to switch to Firefox or accept that the web is going back to being unusable.

Another day, another Apple zero-day with working exploit code floating around the internet. CVE-2025-43300 lets attackers own your iPhone just by sending you a malicious image file. Any image you receive can potentially compromise your phone.

This vulnerability lives in Apple's RawCamera.bundle, the component that processes raw camera formats like Adobe DNG files. The attack is stupidly simple once you understand it - modify two bytes in a DNG file to create a metadata/data mismatch, and iOS completely breaks down trying to generate a preview.

Here's the technical breakdown: DNG metadata says "expect 2 samples per pixel" but the actual JPEG Lossless data only has 1 component. The decompression code trusts the metadata, tries to write more data than expected, and boom - out-of-bounds write that leads to code execution.

The researcher b1n4r1b01 demonstrated this by changing exactly two bytes in a Pentax K-3 Mark III DNG file:

• Offset 0x2FD00: Flip 0x01 to 0x02 (SamplesPerPixel metadata)
• Offset 0x3E40B: Change 0x02 to 0x01 (SOF3 component count)

That's it. Two bytes. Automatic image preview generation does the rest. Works through AirDrop, iMessage, email - basically any way someone can send you a file. Perfect for targeted attacks or mass exploitation.

Apple's response? "Fixed in 18.6.2" without a security advisory because admitting vulnerabilities is apparently still optional in Cupertino. Apple's "fixed in 18.6.2" probably just moved the vulnerability to a different part of the image processing stack. The fact that they stripped debug symbols from RawCamera.bundle suggests they knew this code was sketchy, yet somehow missed an off-by-one error that a fuzzer would catch in minutes.

Security researcher Matt Suiche had to build ELEGANT BOUNCER to detect exploitation attempts because Apple won't acknowledge this vulnerability publicly. Building your own detection tools because Apple won't acknowledge vulnerabilities is peak Cupertino arrogance. This continues Apple's tradition of "security through obscurity doesn't work but let's try anyway."

The CVE database shows this isn't an isolated incident - Apple's image processing has been a recurring security disaster. The community is discussing bypass techniques and additional attack vectors.

Security researchers continue finding similar vulnerabilities in image processing libraries across platforms. Adobe's DNG specification itself isn't the problem - it's Apple's sloppy implementation of the format parsing.

Anyone handling DNG files or running iOS 18.6.1 should patch immediately, assuming Apple actually fixed it this time instead of just moving the problem around. This is a fundamental architecture issue that won't be solved by security guides.

CVE-2025-43300 is a clever way to own iPhones using malicious DNG image files. The attack exploits how Apple processes TIFF containers with JPEG data - basically, when image formats get mixed together, Apple's code gets confused and writes data where it shouldn't.

The nasty part is you don't need to click anything. iOS automatically processes DNG files when generating previews, so just receiving a malicious image through AirDrop, iMessage, or email is enough to get pwned. No user interaction required - the phone just does it automatically.

This is what happens when Apple strips debug symbols to hide problematic code. Their DNG processing has edge cases that weren't properly tested, and because they deliberately make the code harder to analyze, security researchers can't easily find these bugs before the bad guys do.

The exploit is technically sophisticated - you need to understand both TIFF and JPEG file formats to craft a malicious image that causes a buffer overflow. The attacker creates a mismatch between what the file says it contains and what it actually contains, causing iOS to write data outside allocated memory boundaries. Classic buffer overflow leading to code execution.

This isn't just about personal devices getting owned. Since the exploit works through normal file sharing, it can target enterprise environments where people regularly share images. The working exploit code is on GitHub, so expect to see this used in the wild soon.

Since Apple won't admit this vulnerability exists, security researcher Matt Suiche built his own detection tool called ELEGANT BOUNCER. It scans for the specific file structure problems that trigger this exploit, so organizations can at least detect if someone's trying to use it against them.

Apple's response? "Fixed in 18.6.2" without a security advisory because admitting vulnerabilities is apparently still optional in Cupertino. No official acknowledgment, no CVE details from Apple, just a silent fix and hope nobody notices.

This follows a pattern of image processing bugs in iOS - Apple keeps making the same mistakes with complex image formats. When multiple file format standards interact, edge cases emerge that Apple's testing clearly doesn't cover. Every iOS update fixes security issues while breaking something else, then introduces new vulnerabilities in the process.