Iranian Hackers Paralyze US Health Corporation

A hacking group named Handala, linked to Iranian intelligence, has claimed responsibility for a massive cyberattack on the medical technology corporation Stryker. The attackers claim to have deleted data from over 200,000 systems, servers, and mobile devices, leading to the closure of offices in 79 countries. In Ireland, Stryker's largest site outside the US, more than 5,000 employees were sent home. The attack was apparently in retaliation for a US airstrike in February that killed 175 people in an Iranian school. Particularly explosive: The attackers used Microsoft's Intune service for remote management to send a 'remote wipe command' to all connected devices. Hospitals in the US are already unable to order surgical supplies from Stryker, jeopardizing patient care. → thedeepview.com

Synthszr Take: 200,000 wiped systems from a single Intune command. Microsoft inadvertently handed Iranian hackers the perfect weapon: an IT management tool that becomes a digital kill switch. Stryker paid for centralized control and got centralized vulnerability. The irony: the better the IT governance, the more devastating the total failure. Hospitals without surgical supplies demonstrate the physical reach of digital attacks. The next conflict between nations won't be fought with drones, but with admin consoles.

Perplexity Turns the Mac into a PC

Perplexity is transforming its browser agent into a permanent digital employee on the Mac mini. The new 'Personal Computer' can directly control local files and apps, whereas the browser version was limited to cloud services. The system breaks down natural language instructions into subtasks and executes them autonomously—searching folders, editing documents, and coordinating programs. For security, there's a kill switch, permission management, and complete activity logs. CEO Aravind Srinivas is deliberately positioning the product as elite: 'Perplexity Computer is for serious people.' The developer community sees this as more than just a better chatbot—it's software that actually works instead of just assisting.

→ thedeepview.com

Synthszr Take: Apple will sell more Mac minis because of AI agents than because of Final Cut Pro. Perplexity's 24/7 agent needs dedicated hardware, while OpenAI's Computer Use runs in the browser—which explains why Tim Cook remains so relaxed. The kill switch and audit logs are compliance theater; once companies entrust their documents to an agent, the illusion of control is gone anyway. 'For serious people' translates to: high-paying enterprise customers who buy a Mac mini as an appliance. The real disruption isn't autonomous work, but Apple becoming the biggest beneficiary of the agent wave—without writing a single line of agent code.

Cursor Doubles Revenue and Buries Its Business Model

Cursor has doubled its annual revenue to two billion dollars in three months, while its founders simultaneously declare their previous business model dead. The popular AI-powered development environment is facing a fundamental pivot: instead of code completion, the product is set to become a 'software factory' where developers orchestrate fleets of autonomous agents. OpenAI is already demonstrating what this looks like: an engineering team built a complete product without a single line of handwritten code (in one-tenth of the usual time). The competition isn't sleeping: both OpenAI and Anthropic are pushing into the market with their own agent systems, forcing Cursor into a risky vertical integration. The team is hastily recruiting AI researchers and fine-tuning Chinese open-source models to compete against its own suppliers. → Evan Armstrong from The Leverage

Synthszr Take: $102,000 in revenue in 30 days from an autonomous AI business (as demonstrated by Nat) marks the transition from Software-as-a-Service to Software-as-an-Agent. Cursor is sitting on two billion in ARR and still has to fight for survival because the value creation is moving upstream: from lines of code to system architecture, from IDE to agent orchestration. Companies will soon spend $100,000 a month on Claude tokens instead of $30 for a Cursor license. The desperate strategy of competing against OpenAI with fine-tuned Chinese models is reminiscent of Netscape versus Microsoft (we know how that ended). Whoever controls the agent infrastructure collects the margin; IDEs are becoming an interchangeable commodity.

Gemini Moves Into All Google Apps

Google is now integrating Gemini across Workspace, significantly expanding AI capabilities in Docs, Sheets, Slides, and Drive. The new features aim to shift standard tasks like document analysis, content generation, and extracting insights from files from manual labor to assistant-driven automation. For IT departments, this means AI copilots are becoming embedded infrastructure within the tools employees already use daily. Google is thus advancing the vision of AI not as a separate application, but as an integral part of the work environment. The update signals a significant step in the systemic competition of office suites, where Microsoft has taken the lead with Copilot. → TLDR IT

Synthszr Take: Google is making Gemini the invisible colleague in every document, spreadsheet, and presentation. The strategy behind it: whoever controls the work environment controls AI adoption in the enterprise. Microsoft led with Copilot, Google is following suit (and both are hoping Slack and Notion aren't fast enough). IT departments face an interesting challenge: How do you control AI assistants that can think along in every Excel cell and Google Doc? The real disruption isn't the AI itself, but the fact that employees will soon be unable to distinguish between what they wrote themselves and what Gemini suggested.

Nvidia Builds Models Faster Than Others Build Their Pitch Decks

Multi-agent systems for tasks like software engineering or cybersecurity triage generate up to 15 times more tokens than standard chats, blowing any enterprise budget. Nvidia responds with Nemotron 3 Super, a 120-billion-parameter hybrid model that combines three different architectures: state-space models (Mamba-2), Transformer attention, and a novel 'Latent' Mixture-of-Experts design. The key is the intelligent division of labor: Mamba-2 layers act as a 'highway' for processing one million context tokens with linear time complexity, while strategically placed Transformer layers serve as 'global anchors' to extract precise facts from deep codebases or financial reports. The Latent Mixture-of-Experts system compresses tokens before forwarding them to specialists, thus consulting four times as many experts for the same computational cost. On Nvidia's Blackwell-GPUs, the model runs four times faster than 8-bit models on the predecessor Hopper architecture, thanks to native 4-bit floating point optimization—without any loss of accuracy. → venturebeat.com

Synthszr Take: Nvidia sells hardware by giving away software. 120 billion parameters as open weights on Hugging Face—that sounds generous, but it locks customers into Blackwell GPUs, where the model runs four times faster than on older architectures. The technical innovation lies in the mix: Mamba-2 for fast token processing, Transformer for precise retrieval, and compressed Mixture-of-Experts for efficient routing. Multi-agent systems consume 15 times more tokens than regular chats (a single code review can devour millions of tokens). Nemotron solves this through clever architecture rather than raw size. Nvidia is making its GPUs the only sensible platform for enterprise AI—the model is the bait, Blackwell is the hook.

ChatGPT Now Explains Math with Moving Pictures

OpenAI has equipped ChatGPT with interactive visualizations for math and science. Users can now drag triangles, move variables, and instantly see how formulas change—the Pythagorean theorem becomes a toy. The feature launches with 70 topics, from the binomial theorem to Ohm's law, and is available to all logged-in users. 140 million people use ChatGPT weekly for math and science; now they get moving diagrams instead of static explanations. Google's Gemini launched similar features back in November, and OpenAI is following suit with a library ranging from Coulomb's law to compound interest. → Techpresso

Synthszr Take: 70 interactive modules for school subjects—OpenAI is building a digital tutor that never loses patience. ChatGPT is shifting from an answer-giver to an explainer, re-educating its 140 million weekly users as students. Math as a manipulable toy works better for visual learners than any formula sheet (just ask someone with dyscalculia). The real disruption isn't the technology but the business model: if AI tutors are free, who will pay 60 euros an hour for a human tutor? OpenAI is currently commoditizing a billion-dollar market—with colorful triangles.

Pokémon Players Are Mapping the World for Delivery Robots

Niantic is using the movement data from 500 million Pokémon Go players for a new AI model. The company Niantic Spatial, spun off in 2023, aims to enable centimeter-accurate navigation for robots, explains CTO Brian McClendon. Since 2016, players have unknowingly created a three-dimensional world map through their AR hunt for virtual monsters. This data now forms the basis for a 'World Model' that links the intelligence of language models with real-world environments. Delivery robots and autonomous systems could navigate more precisely with it than with conventional mapping services. → The Download from MIT Technology Review

Synthszr Take: 500 million people have created a high-precision 3D world map for Niantic for free. Players hunted Pikachus, while Niantic collected billions of data points about streets, buildings, and obstacles. World Models need exactly these kinds of massive, diverse datasets—Google Street View isn't enough because pedestrians take different paths than cars. Robots benefit from the well-trodden paths of Pokémon hunters. Niantic is now monetizing the unpaid labor of half a billion people (cleverly conceived, ethically questionable).