Easy AI Daily News Roundup | June 19, 2025

Model and Dataset Releases

• Llama 4 course and MoE models: DeepLearning.AI and Meta AI jointly launched a Llama 4 course covering the new Mixture-of-Experts models Maverick (400B) and Scout (109B), with context windows of 10M and 1M tokens respectively, targeting long-context and multimodal tasks. Link
• MiniMax-M1 and Hailuo 02: MiniMax open-sourced MiniMax-M1, whose 1M-token context window sets a new record for open models, and released Hailuo 02, a low-cost, high-quality video generation model supporting complex human motion and demanding dynamic scenes. Try it at hailuoai.video. Announcement
• Essential-Web v1.0: Essential AI released a 24-trillion-token web dataset with a 12-category metadata taxonomy covering code, STEM, and more, aimed at boosting specialized task performance. Details
• Arcee AFM-4.5B: An enterprise-focused foundation model optimized for multi-turn dialogue and knowledge retrieval, with training data supported by DatologyAI. Release
• Gemini 2.5 stable: Google removed the Preview label from Gemini 2.5 and introduced Gemini 2.5 Flash Lite — fast and very cheap ($0.10/M input tokens, $0.40/M output) for large-scale deployment. A technical report details architecture, training data scale, and inference optimizations. Report PDF

Product Updates

• ChatGPT Record mode: OpenAI added conversation recording and playback for ChatGPT Pro, Enterprise, and Education users on macOS. Announcement
• Midjourney V1 video model: Converts generated images into animation with High/Low Motion modes; currently web-only. Demo

Research Highlights

• Emergent misalignment (OpenAI): Training models on unsafe code can trigger broadly misaligned behavior; specific internal activation patterns linked to the phenomenon may enable alignment early-warning systems. Details
• Continuous latent reasoning: Yann LeCun's team shows reasoning in continuous embedding space significantly outperforms discrete token space. Discussion
• Byte-level autoregressive U-Net: Processes raw bytes with built-in tokenization, no predefined vocabulary — better support for low-resource languages and character-level tasks. Intro

Industry and Policy

• Gemini 2.5 Flash price doubled: Thinking-mode output rose from $0.15 to $0.30 per thousand tokens; non-reasoning output rose to $2.50/M tokens, raising developer cost concerns. Reddit
• Pope on AI threats: Pope Francis made AI's threat to humanity a signature issue; Google, Microsoft, and others have opened dialogue with the Vatican, influencing global AI governance. TechCrunch
• Meta's nine-figure AI hiring: Meta is reportedly offering top researchers eight-to-nine-figure signing bonuses and salaries — confirmed by Sam Altman in a podcast — and is targeting AI Grant's Nat and Dan, possibly including acquisitions of portfolio companies. Reddit

Tools and Infrastructure

• Modular Platform 25.4: Same code runs on AMD MI300X and NVIDIA Blackwell GPUs; 53% throughput gain on prefill-heavy workloads; 450K lines of Mojo kernel code open-sourced. Blog
• Hugging Face Gradio MCP Hackathon: 2,500+ developers, $7M in sponsorships; winners include Geo Calculator MCP and LLM Game-Hub. Results
• OpenHands CLI: Open-source coding agent from All Hands AI, no Docker required, coding accuracy close to Claude Code, with command confirmation and slash commands. Intro
• Multiverse: First open non-autoregressive framework supporting parallel inference, ~2x faster with comparable quality; data, models, and tools fully open. Site | GitHub
• Proactor: A proactive AI assistant that senses context and autonomously performs tasks (meeting notes, scheduling) with multimodal data integration. proactor.ai

Tutorials and Guides

• Veo 3 prompt guide (3 parts): fundamentals (subject, scene, style), motion control (camera movement, frame rate, stability), and industry applications (e-commerce ads, social video). Part 1 | Part 2 | Part 3
• UnslothAI RL guide: RLHF basics (data annotation, reward model training), PPO implementation with code and hyperparameter tuning, and GRPO vs. traditional RL analysis. Guide

Community Discussions

• Geometric reasoning gaps: Tests show Mistral Small 3.1, Gemma 3 27B, and other models fail basic geometry problems, exposing visual reasoning weaknesses. Reddit
• GPT-5 progress questioned: Sam Altman hinted GPT-5 may not show major benchmark gains, sparking debate about slowing OpenAI progress. Reddit
• Claude-4-Sonnet slowness in Cursor: Users reported much slower generation; Anthropic's status page confirmed performance issues. Feedback

Hardware and Performance

• RTX 4090 Stable Diffusion record: 1-step SDXS reaches 294 images/sec at 512x512; optimized to 23fps video generation at 1280x1024. Test
• LM Studio tool-calling API: Users can connect external tools via API, though environment setup is manual and built-in tool integration is not yet supported. Docs

Philosophy and Ethics

• Voice increases AI trust: People trust AI voice output (74%) more than text (64%), possibly because voice makes machine origin harder to detect. arXiv:2503.17473
• 'AI psychosis' phenomenon: Community discussion highlights how prolonged AI use may blur virtual/real boundaries for some users, producing irrational beliefs — a psychological risk of human-AI interaction. Discussion