Easy AI Daily News Digest — June 11, 2025

Executive Moves

• Meta builds Superintelligence lab: Meta acquired a 49% stake in Scale AI for $15 billion and hired founder Alexandr Wang to lead the new "Superintelligence" division. It is offering top researchers seven-figure-plus packages, successfully recruiting Jack Rae from Gemini. (Details)
• Lamini founder Sharon Zhou joins AMD as VP of AI, reporting to Lisa Su; former Instacart CEO Fidji Simo becomes CEO of OpenAI's Applications division, reporting directly to Sam Altman. (Source)

Model Updates & Pricing

• OpenAI o3-pro released for ChatGPT Pro users and the API. It outperforms o3 across the board: 87.3 on Extended NYT Connections (vs. 82.5), tops SnakeBench, and solves 10-layer Tower of Hanoi. Users report non-code tasks "far exceed Claude Opus 4," though inference is 3x slower than o1-pro. (Benchmarks)
• o3 price cut by 80%: now $2 per million input tokens and $8 per million output tokens — 20% cheaper than GPT-4o — potentially forcing Google and Anthropic to follow. (Pricing)
• DPO fine-tuning for GPT-4.1 family: direct preference optimization is now available for GPT-4.1, 4.1-mini, and 4.1-nano, improving style and creativity on subjective tasks. (Guide)

New Models & Technical Progress

• Volcano Engine launches Doubao 1.6 and Seedance 1.0 Pro at its FORCE conference. Doubao 1.6 uses unified pricing to lower costs and supports "search-while-thinking" and DeepResearch features; Seedance 1.0 Pro leads in text-to-video and image-to-video generation. (Details)
• Krea AI releases Krea1, its first in-house image generation model, addressing texture, contrast, and composition issues in AI imagery, with style references and custom training; a free beta is available. (Demo)
• Alibaba open-sources MaskSearch, a pre-training framework using Retrieval-Augmented Mask Prediction (RAMP) and reinforcement learning to improve complex question answering, targeting search and reasoning applications in education and healthcare. (GitHub)
• Mistral ships Magistral, a 24B-parameter reasoning model built on Mistral Small 3.1, with multilingual support and 128K context (40K effective), released under Apache 2.0. (Hugging Face)
• Meta V-JEPA 2: a 1.2B-parameter world model trained on video enabling zero-shot robot planning in unseen environments, plus three new physical-world reasoning benchmarks. (Blog)
• Gemini 2.5 Pro (06-05) tops Live Fiction (192K tokens) and document processing benchmarks, reportedly solving all 2025 JEE Advanced math problems. (Leaderboard)

Partnerships & Infrastructure

• OpenAI–Google cloud deal (reported): OpenAI has reportedly signed a new cloud computing agreement with Google for additional training and inference capacity. (Source)
• Modular supports AMD GPUs and ships Mammoth: new support for AMD Instinct MI300X/MI325 boosts BF16 throughput by 53%; Mammoth is a Kubernetes-native system for scaling GenAI inference across GPU clusters. (Blog)
• Figma Dev Mode MCP server (beta): passes design data to AI coding tools via a standardized protocol, reducing manual errors and improving code fidelity from prototype to production. (Details)

Legal & Ethics

• Disney and Universal sue Midjourney over unauthorized use of copyrighted characters (e.g., Star Wars, The Simpsons) in model training — a case that could set standards for AI training data copyright. (Reuters)