Google Day: Mac App, Chrome Skills, Gemini 3.1 Flash TTS

Gemini on Mac: Native Desktop App Instead of a Browser Tab — Google Gets Serious About Local AI

Google is bringing Gemini to the desktop as a native macOS app. The software runs exclusively on Apple Silicon with macOS Sequoia (15.0) and is available in all countries where Gemini is supported. The AI can be summoned from anywhere using a keyboard shortcut (Option + Space for mini-chat, Option + Shift + Space for the full chat), without opening the browser. The core feature: users can share their current window with Gemini, allowing the AI to provide context-aware assistance based on the visible content—be it code, documents, or other data. For full access to browser content, users must grant Gemini accessibility permissions in the macOS System Settings under 'Privacy & Security.' Chat history and saved settings are synchronized across all devices with the same Google account. → TAAFT - There's An AI For That

Synthszr Take: Google is using Apple's own rules against the competition: While Apple Intelligence is still in beta, Gemini is occupying the most valuable real estate on the Mac—the global keyboard shortcut. Sharing the screen content turns the AI into a digital sparring partner that thinks along rather than just waiting for prompts. The key lies in the cross-device synchronization: Google is building a personal AI memory that works across platforms. However, the permission required for browser access also shows the limits of Apple's privacy architecture—what was intended as protection becomes an obstacle to seamless AI integration. Google is betting that users will click away privacy pop-ups for real productivity gains.

Skills in Chrome: Saved Prompts as Digital Action Instructions

Google is introducing a feature in Chrome called 'Skills,' which saves Gemini prompts as reusable commands. Users can define their frequently used AI instructions once and then execute them on any website with a click. The feature is rolling out today for English-speaking Chrome users and is activated via the slash command (/) in Gemini. Saved skills synchronize across all desktop devices on a Google account. Early testers are already using the feature for tasks like calculating nutritional values in recipes or creating product comparisons across multiple browser tabs. Google also offers a library of pre-made skills that can be customized. → AI Secret

Synthszr Take: Chrome is becoming the universal execution layer for AI functions, just as Excel once became the universal calculation engine. Google is turning the browser into the operating system for AI interactions: prompts are becoming installable micro-applications that behave like browser bookmarks. This is reminiscent of the early days of JavaScript when every website suddenly became programmable. The difference: this time, Google controls both the runtime (Chrome) and the AI (Gemini), while Microsoft is building a similar lock-in with Copilot in Office 365. The real innovation isn't in the skills themselves, but in the browser becoming the primary AI interface—a role that operating systems were supposed to fill.

Gemini TTS Can Now Speak with Nuance — No Longer Just a Robot, but a Character

With Gemini 3.1 Flash TTS, Google is introducing a speech synthesis model that can be directed like an actor. The technology not only interprets text but also understands directorial cues: who is speaking (audio profile), where it's happening (scene), and how it should sound (Director's Notes). Using tags in square brackets like [whispering], [excited], or [sarcastic], every part of a sentence can be precisely controlled. The system even understands creative instructions like [like Dracula] or [singing]. An example demonstrates the range: the same sentence can sound bored, rushed, or sarcastic—all controlled simply by tags. The context description serves as a system instruction for consistent output across different transcripts. → Dev.to AI

Synthszr Take: Google is turning text-to-speech into a performance art reminiscent of method acting: the model is given character motivation, scene descriptions, and emotional subtext instructions. This recalls Stanislavski's acting method, where every line is spoken from the character's inner logic. While previous TTS systems worked like teleprompters (reading text), Gemini 3.1 Flash works like a director with their ensemble. The real innovation isn't in the voice quality but in the depth of interpretation: a simple [sarcastic] is enough to turn a neutral statement into biting mockery. Google is thereby democratizing professional voice direction—what once required a recording studio with a voice coach is now done with a tag in square brackets.

SpaceX IPO: Apple and Amazon Push Back

Apple and Amazon are forging an $11.6 billion alliance in orbit. Amazon is acquiring satellite operator Globalstar, becoming the primary satellite service provider for the iPhone and Apple Watch. The deal structure is complex: Globalstar shareholders can choose between $90 in cash or Amazon stock, with the final valuation fluctuating between $10.8 and $11.6 billion depending on the stock price. Apple has been using Globalstar for emergency satellite SMS since the iPhone 14, and Amazon Leo (formerly Kuiper Systems) is set to expand these services. With 241 satellites currently and a goal of 3,000, Amazon is far behind SpaceX's Starlink, which has over 10,000 active satellites, including 650 for direct-to-device services. → The Download from MIT Technology Review

Synthszr Take: Apple is playing realpolitik in space here. The decision against Starlink is not technical but strategic: they don't want to be dependent on Musk's whims when millions of iPhones suddenly need a satellite connection. Amazon as a partner means AWS integration, reliable contracts, and an ally who also has an interest in weakening SpaceX. The satellite business is following the same pattern as cloud computing 15 years ago: first, the tech giants build their own infrastructure for themselves, and then it becomes a service for everyone. Incidentally, Globalstar had also negotiated a sale with SpaceX. Apple and Amazon are jointly preventing Musk from becoming the monopolist for mobile satellite connectivity.

OpenAI IPO: Investors Are Getting Nervous

OpenAI investors are getting restless. The Financial Times reports growing criticism of the company's $852 billion valuation. 'What are you doing with enterprise and code when ChatGPT has a billion users and is growing 50–100% per year?' the FT quotes an early investor. The accusation: OpenAI is 'deeply unfocused' and is getting lost in enterprise software while Anthropic is gaining traction with business customers. The investors' nervousness reveals a fundamental problem: the astronomical valuation is based on the assumption that OpenAI can dominate multiple markets simultaneously. But reality is starting to show cracks in this calculation. → StrictlyVC

Synthszr Take: OpenAI's $852 billion valuation is reminiscent of the physics of superheated systems: the higher the pressure, the more likely a phase transition. The investor criticism reveals the core of the problem: ChatGPT is growing like a consumer product, but the valuation demands enterprise margins. The company is trying to merge two incompatible business models, while Anthropic is consistently focusing on high-paying business customers with Claude. The 'unfocused' nature isn't a weakness of the strategy but its logical consequence: at this valuation, OpenAI has to be everything to everyone. Sam Altman is betting that artificial intelligence will suspend the rules of market physics.

Price War in the Chinese Video AI Market: HappyHorse vs. Seedance

Alibaba's HappyHorse has dethroned Bytedance's Seedance 2.0: with 1411 points, the anonymous model leads the Video Arena rankings, 55 points ahead of the previous market leader. Three days after the surprise launch, Alibaba claimed its 'happy horse.' Bytedance's response was swift: ByteDance 2.0 is opening its API to everyone after users previously complained about eight-hour wait times and price hikes from 7 to 80 yuan for two-minute AI videos. While OpenAI shut down its Sora project after just $2.1 million in revenue on an estimated $5 billion in development costs (a ratio of 2500:1), Chinese providers are fighting for dominance in practical application. Kuaishou is already reporting 340 million yuan in quarterly revenue with Keling AI; HappyHorse supports seven languages for lip-syncing in e-commerce applications. → Hello China Tech

Synthszr Take: Chinese video AI providers are playing a price poker game like German gas stations at noon: when the neighbor lowers their prices, you have to follow suit. Alibaba's timing is surgically precise: they wait until Bytedance has built the market and users are frustrated, then strike with a better product. This is reminiscent of the dynamics in mature franchise systems, where late market entry often brings advantages because you learn from the pioneer's mistakes. The real innovation isn't in the technology (everyone uses DiT-architectures) but in the business model: while OpenAI waited for the big breakthrough, the Chinese are optimizing for 90 percent usability at a reasonable cost. The market rewards pragmatism over perfection.

Cybersecurity Becomes a Function of Token Volume

The UK AI Safety Institute has tested Claude Mythos Preview and confirmed Anthropic's claim: the model finds security vulnerabilities with frightening precision. The researchers noted a simple rule: the more tokens they invested, the more vulnerabilities the system discovered. Drew Breunig boils down the new reality of cybersecurity to a brutal formula: to harden a system, you must spend more tokens searching for exploits than attackers will invest in exploiting them. This token economy paradoxically makes open source more valuable, not less relevant. Once a library has been vetted with millions of tokens, all users benefit from that investment—a shared security buffer against the next generation of AI attackers. → Techpresso

Synthszr Take: Cybersecurity is becoming the energy market of the digital age. As with Bitcoin mining, sheer computing power determines victory or defeat, only here it's about finding vulnerabilities, not validating blocks. This is reminiscent of the Red Queen's hypothesis from evolutionary biology: you have to run faster and faster just to stay in the same place. Open source functions like vaccine development—the high initial cost of token investment is distributed across the entire community, while proprietary software must expensively purchase its own immunity. The ironic twist: the very technology that commoditizes code generation could make open-source libraries indispensable. Security is becoming a shared infrastructure that no one can afford alone anymore.

Workslop: The Productivity Paradox Strikes

A new study from Stanford coins a term that captures the reality in many offices: 'Workslop'—AI-generated work that looks polished on the surface but is so flawed that colleagues have to completely redo it. Ken, a copywriter at a cybersecurity firm in Miami, experiences this daily: after layoffs, the CEO ordered the remaining employees to use AI chatbots to increase productivity. The result: first drafts are quick, but the rework—correcting errors, resolving contradictions between different chatbot outputs—takes longer than without AI. 40% of office workers in a survey of 5,000 respondents say AI doesn't save them time, while 92% of executives rave about productivity gains. The Stanford researchers quantify the damage: 3.4 hours per month per employee are spent on 'work-in-the-loop' cleanup; for a 10,000-person organization, that's an $8.1 million loss in productivity. Particularly telling: companies like Amazon, Block, and Pinterest are simultaneously laying off employees, citing AI-driven efficiency gains. → The Download from MIT Technology Review

Synthszr Take: Workslop is the counterpart to the bullshit job: instead of meaningless human work, we now produce meaningless AI work that humans then meaninglessly rework. We know this pattern from industrial history—the first looms also produced more waste than artisans before they got better. The difference: back then, factory owners knew that poor quality cost them money. Today, CEOs believe the promises of AI vendors more than the experiences of their own teams. The 92% of executives reporting productivity gains are likely measuring output instead of outcome—more documents produced doesn't mean better work. Workslop is the price we pay for wanting to delegate thinking to machines without understanding what thinking actually is.