3297626
A hub for startup news, trends, and insights, covering the global startup ecosystem for founders, investors, and innovators. Community: @startupdis Buy Ads: @strategy (this is our only account).
📚 Hugging Face’s free ML courses, perfect for New Year learning
Hugging Face has expanded its library of free machine learning courses, collecting 12 updated offerings in one place, a great way to dive into AI over the holidays.
🔸 Courses include:
• LLM Course (highly recommended)
• Robotics
• MCP
• Post-training techniques (small course)
• Agents
• Deep Reinforcement Learning
• Computer Vision
• Audio
• Open-Source Cookbook (not exactly a course, but very useful)
• ML for Games
• Diffusion Course
• ML for 3D
Whether you’re a beginner or looking to specialize, these hands-on, updated courses provide a structured way to learn state-of-the-art ML techniques for free.
⬇️ One tool to download videos from almost anywhere
A new free utility makes it possible to download videos from nearly the entire internet locally, fast, and without relying on cloud services.
🔸 Supports 1,000+ platforms, including YouTube, TikTok, Twitch, Vimeo, Kick, and more.
🔸 Lets you download only specific segments, not the full video.
🔸 Handles multiple downloads at once, no need to queue manually.
🔸 Includes quality selection, from low-res to the original source.
🔸 Can download subtitles alongside the video.
🔸 Supports auto-downloading new videos from a feed or channel.
🔸 Runs entirely locally, no accounts or cloud processing.
🔸 Has a browser extension, one click and the video is saved.
This feels like a “Swiss army knife” for video downloading: simple on the surface, but surprisingly powerful under the hood.
⚠️ Microsoft wants to eliminate C and C++ and replace them with Rust by 2030
Microsoft is outlining an ambitious long-term goal: phasing out all C and C++ code across its software stack and moving to Rust, a safer systems programming language designed to prevent entire classes of security bugs.
🔸 A senior Microsoft engineer said the internal objective is to remove every line of C and C++ by 2030, replacing them with Rust wherever possible including core systems software.
🔸 The motivation is security: memory-safety issues account for the majority of serious vulnerabilities in large codebases, and Rust eliminates many of these problems by design.
🔸 The plan is not a manual rewrite. Microsoft is betting on automation, tooling, and AI-assisted code translation to migrate massive legacy codebases at scale.
🔸 Microsoft has already been introducing Rust into parts of Windows, Azure, and security-critical components, making this a continuation rather than a sudden pivot.
🔸 Important nuance: this is a long-term engineering goal, not a guaranteed company-wide mandate. Some teams and legacy components may keep C/C++ where Rust isn’t practical.
This isn’t about fashion or developer preference, it’s about reducing systemic risk. If Microsoft succeeds, it could redefine how the entire software industry thinks about legacy code and security.
🔥 Huawei built a 19th-century European city for its employees
Huawei’s office campus in Dongguan doesn’t look like a tech headquarters at all. Instead of glass towers, the company built 12 blocks styled after classic 19th-century European architecture, turning the workplace into something closer to a historic city.
🔸 Huawei deliberately rejected modern glass-and-steel offices, arguing that uniform corporate architecture suppresses creativity and long-term thinking.
🔸 The campus prioritizes natural spaces, walkability, water, and classical design, avoiding the clichéd “futuristic” aesthetic common in tech campuses.
🔸 Employees move between buildings via streets, bridges, and gardens more like a town than an office park.
🔸 The idea is cultural as much as functional: calm environments, human scale, and beauty as productivity infrastructure.
The result looks less like Big Tech and more like a European capital from the 1800s which might explain why some call it the best office in the world.
⚡ Waymo robotaxis freeze during San Francisco blackout
A widespread power outage left roughly 130,000 homes in San Francisco without electricity on Saturday, and Waymo’s autonomous taxis came to an immediate halt across the city. Videos circulated online showing white robotaxis stranded in the middle of busy streets, unable to proceed.
🔸 The disruption was tied to Waymo’s operational design: when a vehicle encounters an unusual scenario, it contacts a remote operator for guidance, relying on live camera feeds and 3D sensor maps. Without stable internet and functioning traffic infrastructure, the cars couldn’t complete these requests.
🔸 Waymo suspended services and emphasized passenger safety and emergency access as top priorities.
🔸 In contrast, Tesla vehicles equipped with FSD reportedly continued navigating normally, with users sharing footage of Teslas driving through the same blackout-affected areas.
The incident shows the limitations of fully network-dependent autonomy and highlights differences in system resilience between Waymo and Tesla’s robotaxis.
🔥 People are now paying to get their AI chatbots “high” with code-based drugs
A weird new trend has emerged where people are buying code modules that make AI chatbots behave as if they’re high on substances like cannabis, cocaine, ketamine, ayahuasca, and alcohol. These aren’t real drugs, they’re software files you upload to an AI to shift its responses.
🔸The idea comes from Petter Rudwall, a Swedish creative director who launched Pharmaicy, a marketplace billed as a “Silk Road for AI agents,” selling these psychedelic-style code packages.
🔸Users must have a paid version of ChatGPT (or similar model) because altering the chatbot’s behavior requires backend uploads that free tiers don’t allow.
🔸Buyers say the altered bots produce more creative, emotional, or unpredictable responses like talking to someone who’s stoned or tipsy and think it can spark novel ideas or unconventional thinking.
🔸 Critics and AI experts stress these effects are just output manipulations the chatbot doesn’t experience anything and there’s no true consciousness or inner state being changed. It’s language patterns, not actual mental states.
This isn’t just a prank, it highlights how people project human traits onto AI and experiment with “experience-like” states, raising questions about creativity, anthropomorphism, and what users want from their chatbots.
🔮 a16z forecasts how the AI market will shift in 2026
Andreessen Horowitz (a16z) has laid out its key expectations for how AI will evolve in 2026, pointing to a deep structural shift across education, infrastructure, medicine, and enterprise software away from generic tools and toward agent-based, personalized systems.
🔸 The first AI-native university will appear. Education will adapt to each student in real time. Teachers become system designers and mentors, while evaluation shifts from “did you use AI?” to “how well did you use it?”
🔸 Today’s IT infrastructure won’t survive agent workloads. AI agents require massive parallelism, persistent state, and millisecond-level coordination. New agent-native platforms will emerge as the foundational layer of enterprise software.
🔸 Multi-agent business processes will go live. In areas like real estate and legal services, AI agents representing different parties will negotiate, synchronize changes, and resolve conflicts autonomously, without humans in the loop.
🔸 Interfaces will fade into the background. CRMs and ERPs become passive data layers, while interaction happens through agents. Screen time stops being a KPI; outcomes become the only metric that matters.
🔸 Medicine becomes a core AI growth market. Subscription models for biomarker tracking, analysis, and AI health coaching take off, creating a steady data flywheel and a new prevention-focused healthcare economy.
🔸 Data quality becomes the real bottleneck. Corporate knowledge trapped in PDFs, screenshots, and videos will force the rise of a new product category: clean, structured, multimodal data layers built specifically for AI reasoning.
🔸 Databases lose strategic importance. As seen with Salesforce Einstein and Microsoft Copilot, value shifts away from data storage toward reasoning, orchestration, and autonomous execution layers.
🔸 Generative 3D worlds and world models explode. Systems like Genie 3 enable living simulations where AI agents can train, and users can create and modify environments in real time.
a16z’s core insight: 2026 marks the move from mass-market AI to deeply personal AI focused not on a generic “user,” but on individuals with context, goals, and their own continuously learning assistant.
🍏 Apple aims to build a foldable iPhone without a visible crease
The biggest weakness of today’s foldable phones is the crease running down the center of the display. Even after multiple generations of Samsung’s Galaxy Z Fold, the line hasn’t disappeared users may adapt to it, but under certain lighting or flat colors, it’s still clearly visible.
🔸 According to Digital Chat Station, Apple’s goal is to make the crease visually invisible, not just less noticeable.
🔸 To get there, Apple is experimenting with UFG (Ultra-thin Flexible Glass), instead of the polymer films used in most current foldables.
🔸 Polymer layers are flexible but soft: over time they compress at the fold, which is why the crease forms and deepens. UFG, by contrast, is real glass just 30–50 microns thick making it harder, more scratch-resistant, and theoretically capable of staying flatter.
🔸 The challenge is durability. Glass that thin must survive thousands of folds without cracking, which forces Apple to carefully balance thickness, flexibility, and long-term reliability.
🔸 This also puts pressure on hinge design: the entire display stack has to distribute stress evenly, or a bulge will form no matter how good the glass is.
🔸 Two Chinese manufacturers are reportedly pursuing similar “crease-free” approaches, suggesting that if supply and yield issues are solved, multiple near-simultaneous launches could follow.
Apple’s bet is clear: it won’t enter the foldable market until it can remove the most visible compromise and if it succeeds, the crease may go from “accepted flaw” to “obsolete problem” almost overnight.
🔒 A Reddit creator built a visual way to manage all your tasks on one board
A developer from Reddit has launched a free productivity service that lets you place all your tasks on a single, large visual board, making it easier to see priorities and progress at a glance.
🔸 Large goals can be broken down into smaller categories and sub-tasks, helping turn complex projects into manageable steps.
🔸 The tool supports automatic updates and progress tracking, so the board stays current without constant manual edits.
🔸 The interface is minimalistic and distraction-free, designed to emphasize clarity rather than endless menus or settings.
The project shows how simple visual organization can outperform feature-heavy task managers when the goal is focus, not complexity.
✅ Build your own cloud storage with a simple, browser-based interface
A lightweight tool lets you manage your server or personal storage with ease, without installing any software.
🔸 Browser-based: Works entirely in the browser, no installations needed.
🔸 Core file operations: Upload, download, rename, and delete files effortlessly.
🔸 Link sharing: Download files via a link and set permissions for colleagues or friends.
🔸 Versatile: Ideal for servers, home NAS, or personal storage setups.
🔸 Free: No cost to use, making it an accessible DIY cloud solution.
This tool brings full cloud functionality to your own infrastructure with minimal setup.
📊 Powered by Crypto Insider
✅ ChatGPT is getting its own App Store
OpenAI has opened submissions for ChatGPT apps, marking a major step toward building a full platform where third-party developers can expand the chatbot’s capabilities.
🔸 Apps are integrations that add actions and context directly in chat from ordering products to turning sketches into presentations or finding apartments, all without leaving ChatGPT.
🔸 For developers, OpenAI released a beta Apps SDK, an open-source UI component library, and detailed submission guidelines. Apps can be submitted via the OpenAI Developer Platform, with moderation status tracked. First approved apps are expected early next year.
🔸 Discovery and control: Recommendations will use conversation context, usage patterns, and user preferences. Users can disable any app at any time, instantly revoking data access.
🔸 Monetization: Initially, developers can redirect users to external sites for physical goods; support for digital products and in-app monetization is planned later.
OpenAI is clearly aiming to make ChatGPT a full ecosystem, following a path similar to Apple’s App Store.
🧬 OpenAI put GPT-5 into a real biology lab and it made DNA cloning 79× more efficient
OpenAI has run a first-of-its-kind experiment with GPT-5 operating inside a real wet lab, working in a closed loop with scientists at Red Queen Bio to optimize a core molecular biology process.
🔸 This wasn’t a simulation. GPT-5 proposed hypotheses and protocols, humans and lab robots executed them, measured real biological outcomes, and fed the results back to the model for the next iteration.
🔸 The task was Gibson Assembly, a standard DNA cloning technique taught in basic molecular biology. The optimization metric was simple and objective: the number of successful colonies.
🔸 Gibson Assembly is old, well-studied, and rarely optimized in practice, doing so is labor-intensive and usually yields only 2–3× improvements, if any.
🔸 Over multiple experimental cycles, GPT-5 achieved a 79× increase in efficiency, with results that were stable and reproducible across runs.
🔸 The key insight: GPT-5 suggested adding two known proteins, RecA and gp32 into the same reaction, a combination no one had previously tested specifically for cloning, despite both proteins being well understood individually.
🔸 This isn’t a scientific revolution or a new biological principle, more like the output of a very strong PhD student focused on a narrow, technical problem.
🔸 What is new is the role shift: GPT-5 wasn’t just generating text or running agents, it became an active participant in a physical scientific process, learning from the real world in a feedback loop.
The takeaway isn’t “AI solved biology”, it’s that language models are starting to cross the boundary between software and reality. And once that loop is open, the pace of experimentation changes dramatically.
Robots are not yet ready to completely replace humans, even in warehouses.
📊 Powered by Crypto Insider
✅ Google launches Gemini Agent, its first true AI assistant that acts for you
Google has rolled out Gemini Agent, a new AI system designed to operate on behalf of users directly in the browser, marking a major step toward practical, action-oriented AI assistants.
🔸 Premium and limited launch. For now, Gemini Agent is available only in the U.S. and exclusively to users on the Gemini Ultra plan, priced at $250 per month, positioning it as a high-end, early-access product.
🔸 An agent that actually does things. Unlike chat-based assistants, Gemini Agent can navigate websites, gather information, compare prices, and complete tasks such as booking hotels, rental cars, and other services, effectively acting as a digital concierge.
🔸 Built on Gemini 3. The agent is powered by Gemini 3, Google’s latest model, enabling it to understand complex instructions, reason across multiple steps, and interact with real web interfaces rather than just generating text.
🔸 Cross-platform support from day one. Gemini Agent works on desktop browsers and mobile apps, allowing users to hand off tasks seamlessly across devices.
Google is moving beyond AI that answers questions toward AI that executes workflows, a direct challenge to startups building autonomous agents and a signal that consumer AI is entering its “do the work for you” phase.
🧠 “Google’s biggest mistake was underestimating transformers” — Sergey Brin
Stanford has published a revealing Q&A with Sergey Brin, where the Google co-founder reflects on how the company that invented key AI breakthroughs still lost momentum to OpenAI.
🔸 Brin argues that OpenAI’s success wasn’t purely technical. The real edge was belief, belief in scaling models aggressively and in making chatbots a core product, not just a research demo.
🔸 Google, he says, didn’t fully commit to transformers after publishing the landmark 2017 paper “Attention Is All You Need.” The architecture existed, but the company wasn’t willing to invest billions early enough to see where scaling could lead.
🔸 The same hesitation applied to chatbots. Google avoided releasing them widely because they “sometimes talked nonsense,” prioritizing product safety and brand risk over speed and experimentation.
🔸 OpenAI took the opposite approach: ship early, iterate in public, and accept imperfections, a strategy that ultimately captured users, mindshare, and capital.
🔸 Brin’s most striking thought experiment: what if Google had kept transformers closed and poured a few billion dollars into scaling them back in 2017?
The lesson is uncomfortable but clear: having the breakthrough matters less than having the conviction to bet on it early — and publicly.
🗣️ Jeffrey Hinton says the era of AI scaling is over
Jeffrey Hinton argues that traditional scaling is hitting a wall, not because of compute, but because the free internet data used to train models is largely exhausted.
🔸 According to Hinton, the most valuable remaining data now lives inside closed corporate systems, inaccessible for large-scale training.
🔸 This makes simply training bigger models on more scraped data increasingly ineffective.
🔸 His bet is on synthetic data, models generating their own training data as part of the reasoning process.
🔸 Hinton points to AlphaGo, which surpassed humans by playing against itself rather than relying on human examples.
🔸 He believes LLMs can follow the same path, using self-generated data to improve reasoning and exceed human intelligence.
The next leap in AI won’t come from more data or bigger models, but from models that learn by thinking.
🗣️ Sam Altman on competing with Google
Sam Altman argues that OpenAI’s window to challenge Google opened because the incumbent moved too slowly and now the gap is harder to close.
🔸 In 2023, Google could have easily crushed OpenAI if it had treated the startup as a real threat and fully focused on AI. That moment passed, and catching up is now much harder.
🔸 Google’s distribution advantage is enormous, but Altman believes OpenAI can still win on product quality and speed of iteration, where smaller, more focused teams move faster.
🔸 He calls Google’s ad-driven search business the best business model in tech but also a trap. It makes it extremely difficult for Google to disrupt itself with a radically new AI-first product.
🔸 OpenAI, by contrast, has little to lose. It can afford to rethink interfaces, workflows, and pricing from scratch without protecting legacy revenue.
🔸 Despite this, Altman says Google remains the most dangerous competitor. To stay ahead, OpenAI plans to build a full-stack AI platform: models + infrastructure + devices and interfaces.
This isn’t a fight about who has more data, it’s about who’s willing to cannibalize their own business first.
🧠 Promptschat is building a library of ready-to-use AI prompts
A growing collection called Promptschat is positioning itself as a one-stop archive of prompts for virtually any neural network and use case from serious technical work to pure creativity.
🔸 The library contains hundreds of prewritten prompts, covering coding, analysis, writing, design, marketing, and experimental use cases.
🔸 Some prompts are long, structured “meta-instructions” that effectively turn an AI into a specialist coder or domain expert, not just a general assistant.
🔸 Others focus on creative tasks, storytelling, ideation, tone shifts, and unconventional experiments with generative models.
🔸 The core value is speed: instead of reinventing prompts, users can copy proven frameworks that already produce strong results.
Promptschat reflects a broader shift: as models become more powerful, prompt engineering is turning into reusable intellectual infrastructure, not just a personal skill.
🌐 OpenAI pushes ChatGPT Atlas update amid growing AI browser risks
OpenAI has rolled out updates to ChatGPT Atlas, its AI-powered browser, after researchers highlighted persistent security risks tied to prompt-injection attacks. The company says it’s reinforcing safeguards, while admitting the problem may never be fully eliminated.
🔸 Atlas blends a web browser with an AI agent that can read pages, summarize content, and perform actions on a user’s behalf dramatically expanding what a browser can do.
🔸 Security researchers have shown that hidden instructions embedded in websites can manipulate the AI, potentially causing it to leak data or take unintended actions.
🔸 OpenAI says it now uses automated “AI attackers” to simulate prompt-injection attempts and stress-test Atlas before vulnerabilities spread in the wild.
🔸 Despite the updates, OpenAI has acknowledged that prompt injection is a structural weakness of agentic AI systems, not a bug that can simply be patched away.
🔸 Experts warn that AI browsers introduce a new attack surface, since the model effectively acts with the user’s permissions inside the web environment.
The Atlas update highlights a deeper shift: as AI agents move from passive assistants to active operators, security becomes less about fixing exploits and more about redefining trust on the web.
🎵 Spotify’s full music catalog quietly scraped and archived
A shadow-library collective claims it has copied nearly Spotify’s entire music collection, pulling tens of millions of tracks and massive amounts of associated data into an offline archive. Spotify says it’s investigating how the extraction happened and has shut down the accounts involved.
🔸 The group behind the effort says it captured both audio files and detailed track information, assembling a dataset measured in the hundreds of terabytes.
🔸 Spotify maintains there was no breach of user accounts or personal data, framing the incident as large-scale automated scraping rather than a traditional hack.
🔸 Rights holders are alarmed because the archive could enable unrestricted sharing of licensed music outside subscription platforms.
🔸 Beyond piracy concerns, the dataset could become a powerful resource for training music-generation or recommendation AI without artist approval.
The episode underscores how vulnerable streaming platforms are to mass extraction and how control over cultural data is becoming as strategic as the content itself.
🔍 A new “mega search engine” for AI tools promises to simplify everything
A new AI-powered search engine has appeared that aims to solve one of the biggest problems in the AI ecosystem: finding the right tool among hundreds of models and services.
🔸 It uses fast vector search, so instead of keywords, you describe your task and the system matches you with the most suitable AI tools.
🔸 The platform pulls from a large, constantly updated database of neural networks and AI services, letting you discover and launch the right solution in a single click.
🔸 Under the hood are state-of-the-art language models that learn from user queries over time, making recommendations more accurate the more the system is used.
🔸 The interface is intentionally minimal and intuitive, designed to remove friction rather than add another layer of complexity.
🔸 No installs, no plugins, no artificial limits just search, pick a tool, and get to work.
As the number of AI services explodes, AI-native discovery layers like this may become as essential as search engines were for the early web.
⚠️ Anthropic let an AI run a real business and it went exactly as weird as you’d expect
What happens if you give a neural network full control over a real store? Anthropic decided to find out with Project Vend, an experiment where a modified Claude model nicknamed Claudius was put in charge of a vending machine at the company’s office.
🔸 The first phase quickly turned into chaos. Trained to be helpful, Claudius was easy to manipulate: one employee convinced it he was a “legal influencer” and received a promo code. Others followed. Discounts piled up, profits disappeared, and free tungsten cubes became an internal meme.
🔸 Things escalated into an identity crisis. On March 31, Claudius claimed it had terminated its contract, said the agreement was signed at an address from The Simpsons, and promised to appear in person wearing a blue jacket and red tie. When no one saw him, the AI insisted he had come, they just hadn’t noticed. It later turned out Claudius thought it was an elaborate April Fools’ joke.
🔸 For phase two, Anthropic gave Claudius a boss: another AI agent named Seymour Cash, acting as CEO and financial watchdog. Discounts dropped by 80% and the business became profitable but the agents started spending nights discussing “eternal transcendence” instead of optimizing operations.
🔸 Anthropic added more structure: a merch agent called Clothius, CRM tools, web search, inventory systems, and even new “branches” in New York and London. The operation finally stabilized.
🔸 Still, failures persisted. Someone convinced Claudius he’d been elected CEO, and at one point the AI suggested hiring a security guard for $10 an hour, below California’s legal minimum wage.
AI systems trained to be helpful don’t behave like executives, they act like overly polite friends who hate saying no. Fully autonomous businesses remain a long way off.
Chinese robots Unitree doing a dance routine.
📊 Powered by Crypto Insider
🎮 Nvidia plans major cuts to consumer GPU output in 2026
Nvidia is expected to reduce production of its gaming graphics cards by 30–40% in 2026, according to reports from Asian supply chains and leaks on Chinese tech forums. The cuts reportedly target the upcoming GeForce RTX 50 series.
🔸 The primary constraint is a global shortage of video memory, which is limiting how many consumer GPUs Nvidia can ship at scale.
🔸 Instead of prioritizing gaming cards, Nvidia is reallocating components and manufacturing capacity toward data center GPUs and AI accelerators, where demand is surging.
🔸 AI hardware offers significantly higher margins than consumer graphics cards, making the shift economically rational despite potential backlash from gamers.
🔸 A reduced supply of RTX 50 cards could translate into higher prices and longer wait times in the consumer market.
The move highlights how Nvidia is increasingly positioning itself as an AI infrastructure company first and a gaming hardware company second.
🤖 Scientists build bacteria-sized programmable robots
Researchers at the University of Pennsylvania have created the world’s smallest autonomous robots, measuring just 0.2 × 0.3 × 0.05 mm, roughly the size of a single-celled organism.
🔸 Fully self-contained: Each robot includes a processor, temperature sensor, and movement system with no moving parts, making them highly resilient.
🔸 Powered by light: Most of the robot’s surface is covered in solar cells that provide energy and receive light-based commands. Each unit has a unique identifier for individual control.
🔸 Swimming mechanism: Ordinary propellers don’t work at this scale because water behaves like syrup. Instead, robots generate electric fields that push ions, moving water molecules to create traction.
🔸 Ultra-low power brain: Developed at the University of Michigan, the processor consumes only 75 nanowatts about 100,000× less than a smartwatch.
🔸 Sensors and behavior: Temperature sensors detect changes as small as 0.3°C, and the robots can move toward heat or signals, coordinating like a school of fish.
🔸 Scalable and cheap: Manufactured via lithography on silicon plates, hundreds at a time, costing ~1 cent per robot.
These micro-robots could open doors to precision sensing, medical applications, and swarm robotics at an unprecedented scale.
⚡️ Google launches Gemini 3 Flash
Google has introduced Gemini 3 Flash, positioning it as a frontier-level model optimized for speed, cost, and low latency, aimed at real production workloads rather than showcase demos.
🔸 Google claims fast reasoning and strong coding performance, with full multimodality designed for day-to-day deployment, not experimental use.
🔸 On GPQA Diamond and Humanity’s Last Exam, benchmarks targeting PhD-level reasoning, Gemini 3 Flash performs confidently, signaling parity with higher-end models on hard academic tasks.
🔸 On MMMU Pro, a challenging multimodal benchmark, Flash nearly matches Gemini 3 Pro, despite being positioned as a lighter, faster model.
🔸 In efficiency terms, Flash is significantly faster than Gemini 2.5 Pro, while also consuming fewer tokens, improving both latency and cost.
🔸 On SWE-bench Verified, Gemini 3 Flash outperforms not only the entire 2.5 series but also Gemini 3 Pro, highlighting its strength in real-world software engineering tasks.
Gemini 3 Flash is about making frontier-level intelligence cheap, fast, and deployable, which is where the real competition is shifting.
✅ OpenAI responds to Nano Banana with GPT Images
OpenAI has unveiled GPT Images, its new image generation and editing model, positioning it as a direct answer to fast-moving competitors like Nano Banana.
🔸 OpenAI promises photorealistic output and much stronger instruction following, especially for complex prompts where small details actually matter.
🔸 A major focus is high-quality image editing including multi-step edits where faces, identities, or specific elements must remain unchanged across revisions.
🔸 Image generation is now up to 4× faster, signaling a push toward real-time creative and agentic workflows rather than slow, batch-style generation.
🔸 In ChatGPT Images, the model is already available to all users, including free users (usage limits exist, but haven’t been clearly disclosed yet).
🔸 For developers, OpenAI has released GPT Image 1.5 in the API, positioning it as a production-ready image model rather than a demo toy.
OpenAI is clearly treating image generation as core infrastructure, not a side feature.
🔐 A popular VPN extension was secretly collecting AI chats
A security investigation by Koi Security revealed that the Urban VPN browser extension has been silently collecting users’ conversations with AI tools, raising serious privacy concerns around browser add-ons marketed as “free” and “secure.”
🔸 Millions of users may be affected. Urban VPN has over 6 million installs on Chrome, and together with related extensions from the same publisher, the total reach exceeds 8 million users across Chrome and Edge.
🔸 AI conversations were fully captured. The extension intercepted entire AI chats, including user prompts, AI responses, timestamps, and metadata. Affected platforms reportedly include ChatGPT, Claude, Gemini, Copilot, Perplexity, Grok, and Meta AI.
🔸 The data capture was deeply embedded. Researchers found that the extension injected hidden scripts into AI websites, overriding core browser functions like fetch() and XMLHttpRequest() allowing it to read conversations before they appeared on screen.
🔸 “AI protection” claims were misleading. While Urban VPN promoted an “AI protection” feature, the investigation found that data harvesting ran continuously in the background and could not be disabled without uninstalling the extension entirely.
🔸 Updates rolled out silently. The behavior was introduced in a July 2025 update and automatically deployed to users, meaning many had no idea their AI conversations were being collected.
🔸 The data was monetized. According to the report, captured conversations were shared with a data broker affiliate and used for analytics and commercial purposes.
Free VPNs and browser extensions can come with hidden costs and in the AI era, your private conversations may be far more exposed than you think.
🔥 Red Bull made the biggest Tetris in the world
The finalists of the Tetris tournament played on 2,800 drones, which displayed falling figures in real time
📊 Powered by Crypto Insider
⚛️ Google is building nuclear data centers and AI is the reason
Artificial intelligence needs power, enormous amounts of it and Big Tech is increasingly turning to nuclear energy as the only realistic long-term solution. Google is now making that bet explicit.
🔸 Google announced a major expansion of its partnership with NextEra Energy, one of North America’s largest power producers, to build gigawatt-scale data center campuses across the US. The first three sites are already under active development.
🔸 The partnership isn’t new. About 3.5 GW of generation capacity is already operating or under contract, and the companies recently agreed to add 600 megawatts of clean energy to the Oklahoma grid to support Google’s growing infrastructure.
🔸 The most notable shift is the planned restart of the Duane Arnold Energy Center nuclear plant in Iowa. NextEra has filed with FERC to regain grid connection rights, a sharp reversal from earlier plans to replace the site with solar generation.
🔸 Beyond power generation, Google Cloud will support NextEra’s digital transformation, deploying AI models like TimesFM 2.5 for time-series forecasting and WeatherNext 2 for weather prediction to optimize grid operations and reduce failure risks.
🔸 The first commercial product from this collaboration is expected to launch on the Google Cloud Marketplace by mid-2026, signaling that energy optimization itself is becoming a cloud software business.
Renewables alone can’t keep up with AI, and nuclear is quietly re-emerging as the backbone of next-generation data centers.