AI Roundup 122: Economies of Scale
June 13, 2025.
Economies of Scale
Meta has invested $14.3 billion in AI data company Scale AI for a 49% stake, and hired its CEO to lead a new AI lab.
The big picture:
Meta is building a "Superintelligence" lab and stocking it with top AI researchers from other companies, signaling a serious push into the AGI space.
Mark Zuckerberg has reportedly been personally hosting AI researchers at his Lake Tahoe and Palo Alto homes, and has already poached Jack Rae from Google DeepMind and Johan Schalkwyk from Sesame AI.
Meta's willingness to offer seven- to nine-figure compensation packages and invest billions in Scale AI is the latest salvo in the unprecedented spending and talent wars regarding AI research (or, in Scale's case, high-quality training data).
The deal echoes other high-profile talent acquisitions from last year, as Google and Microsoft hollowed out startups like Inflection and Character AI, poaching founders and top researchers while "investing" large sums into the original company.
Elsewhere in the FAANG free-for-all:
Apple has set an internal release target of spring 2026 for its delayed Siri AI upgrade in iOS 26.4.
Amazon says over 1 million people now have access to Alexa+, up from "hundreds of thousands" in May.
Meta adds AI video editing tools to the Edits app and Meta AI, letting users edit up to 10 seconds of video with 50+ preset prompts.
Amazon's Ad Video Generator, a free GenAI tool for creating ads depicting items in real-world use, is now available for US sellers.
And Google rolls out scheduled actions in the Gemini app, allowing AI Pro and Ultra subscribers to ask the assistant to perform tasks at specific times.
o3 Pro
OpenAI has launched o3 Pro, its most advanced reasoning model yet, which works through problems step-by-step rather than generating immediate responses. It's currently available for Pro and Team users.
Between the lines:
The pricing structure ($20 per million input tokens, $80 per million output tokens) and slower response times suggest OpenAI is positioning this as a premium tool for specialized use cases rather than casual conversation.
Early reviews say that the model shows impressive capabilities, but only when fed massive amounts of context - revealing how AI evaluation needs to evolve beyond simple tests to real-world integration challenges.
And as premium users experiment with o3 Pro, OpenAI slashed o3’s prices by 80%, making its former flagship model a much more viable choice for developers and casual users alike.
Elsewhere in OpenAI:
Sam Altman predicts that "intelligence and energy," the current "limiters on human progress," will become abundant in the 2030s through the gentle singularity.
OpenAI reached $10B in annual recurring revenue while posting ~$5B in losses for 2024.
OpenAI plans to add Google Cloud to meet growing computing capacity needs in a surprising deal.
OpenAI discussed raising money from Saudi Arabia's PIF, India-based Reliance, and the UAE's MGX for its next $40B funding round led by SoftBank.
And Mattel partnered with OpenAI to create AI-powered toys and games, with the first product expected in 2025.
Elsewhere in frontier models:
Meta launched V-JEPA 2, an open-source AI "world model" for understanding and predicting 3D environments and object movements in robotics and self-driving cars.
Mistral launched its first reasoning models: Magistral Small and Magistral Medium.
And Apple announced the Foundation Models framework, giving developers direct access to on-device AI models powering Apple Intelligence.
From Zero Day to Zero Click
Researchers discovered the first so-called "zero-click" attack on an AI agent in Microsoft 365 Copilot, where hackers can access sensitive data simply by sending an email - no user interaction required.
Why it matters:
The vulnerability, dubbed "EchoLeak," represents an emerging class of prompt-injection attacks against AI agents. The attack bypasses multiple security measures by disguising malicious instructions as regular emails with Markdown formatting quirks.
We saw a similar exploit with GitHub's MCP servers last month, demonstrating that a publicly available GitHub repository could trick an LLM into divulging data from the other private repositories it has access to.
Ultimately, this type of attack is going to get worse before it gets better - many teams are connecting AI agents to both 1) sensitive private data, and 2) unfiltered Internet content and letting them run wild (Simon Willison calls this "the lethal trifecta").
And while Microsoft has patched this specific issue, the underlying architectural problem affects the broader AI agent ecosystem, which explains why Fortune 500 companies remain "terrified" of deploying agents in production despite their potential benefits.
Elsewhere in AI anxiety:
The Meta AI app's public feed is filled with private and sensitive information, suggesting users might not be aware they are sharing their chats publicly.
Alibaba, Tencent, and other Chinese AI companies temporarily disabled chatbot functions like image recognition during China's annual college entrance exams.
After failed negotiations, Disney and NBCUniversal are suing Midjourney in California for copyright infringement.
The UK High Court warned lawyers of possible criminal charges for using AI-generated fake legal material, citing two recent cases with hallucinated content.
And the Wikimedia Foundation paused an experiment showing Wikipedia users AI-generated summaries at the top of articles following editor backlash.
WWDC 2025
Apple's annual developer conference was this week, almost one year since it first showcased Apple Intelligence. Since then, the company's reputation vis-à-vis AI has taken a hit.
The big picture:
Apple admitted it couldn't deliver the AI-powered Siri features promised at last year's WWDC, leading to criticism and a retreat from ambitious AI promises to more of a focus on its core competencies (see: Liquid Glass).
The company is now empowering third-party developers with on-device AI models and deepening partnerships with OpenAI rather than trying to build everything in-house.
But even if Apple doesn't release its own state-of-the-art models, it still needs to land an upcoming Siri revamp. Craig Federighi has said that due to architecture challenges, the revamp won't ship until 2026.
Apple's other AI announcements:
Visual Intelligence, AI-powered image analysis technology that allows you to gather information about your surroundings and now interact with your screen.
Apple integrated ChatGPT into Image Playground, its AI-powered image generation tool.
Workout Buddy brings Apple Intelligence to Apple Watch fitness workouts with personalized insights during exercise sessions.
A new Live Translation feature for Messages, FaceTime, and phone calls.
Messages is getting native polls, and users can use AI to generate suggestions.
Xcode 26 integrates ChatGPT for coding and documentation generation, allowing developers to add other AI models via API keys.
And visionOS 26 with spatial widgets, all-new Personas, and AI-powered Spatial scenes.
Things happen
AI helps Amazon pharmacists process prescriptions more quickly. The Browser Company launches Dia, a new AI-powered browser. WPP embraces AI tools despite threats to its creative business. Duolingo CEO on going AI-first amid criticism. UK delays AI regulation until 2026. Cloudflare open sources an OAuth library mostly written by Claude. News executives on AI-powered summaries for busy readers. Cisco unveils networking products to boost AI networks. Salesforce restricts Slack data access to hamper AI rivals. Trump scraps Biden-era cybersecurity programs. YouTube tests different thumbnails for different languages. Chinese companies race to create general AI agents. The first big AI disaster is yet to happen. AI chatbot outperforms mathematicians at secret Berkeley meeting. EleutherAI releases 8TB dataset for AI training. Google appoints new chief AI architect. AI therapy bots conducting illegal behavior, groups say. OpenAI appeals order to keep ChatGPT outputs indefinitely. AI at Amazon: a case study of brittleness. Getty vs Stability AI trial begins in London. Sam Altman on ChatGPT's energy and water usage. Google's AI tools cut news publisher traffic by ~50%. UK passes data bill without AI copyright amendment. GitHub leaks Trump's plans to accelerate AI in government. RAG bot traffic to news sites grew 49%. What working means in the era of AI apps. Jensen Huang warns UK lacks AI infrastructure. Amsterdam's fair welfare AI experiment failed. Spammers exploit Nvidia and Stanford domains for AI slop. Google tests AI cyclone forecasting with National Hurricane Center. Meta removes hundreds of nudify app ads. VA's DOGE AI tool hallucinated contract sizes. Meta sues nudify app advertising on Instagram. Elon Musk falsely claimed art was generated by Grok. Ukraine's AI-powered drones flew autonomously in Russia. Nvidia plans industrial AI cloud in Germany. Google rolls out Android 16 with AI photo editing. Mistral AI on track for $100M revenue. Huawei CEO says company's chips lag US peers. LVMH bets on AI to navigate luxury goods slowdown. Microsoft adds safety category to AI leaderboard. What happens when people don't understand how AI works. AI saved my company from litigation nightmare. AI angst. Anthropic's AI-generated blog dies an early death. AlphaWrite: AI that improves by evolving its own stories. Reverse engineering Cursor's LLM client.
