📅 AI Industry Digest — June 17, 2025
*Translated and summarized from the zhichai.net Easy AI Daily.*
New Model Releases
- MiniMax-M1 open-source LLM: MiniMax AI released a 456B-parameter open-source model supporting 1M-token input and 80k-token output, using an efficient "lightning attention" mechanism and the CISPO GRPO variant. Weights and technical report are available. Hugging Face
- Hailuo 02 video model: Also from MiniMax, benchmarked against ByteDance's Seedance. Generation is currently slow (~20 min/video); no weights or API yet. Link
- Kimi-Dev-72B: Moonshot AI's open-source coding model scores 60.4% on SWE-Bench Verified, surpassing DeepSeek R1, with RL-optimized real-repo patch generation. Hugging Face
- Anthropic: A multi-agent system (Claude Opus 4 lead, Sonnet 4 subagents) improved task completion by 90.2% over a single Opus 4 in internal evals — but token usage and agent orchestration need optimization. Engineering blog
- Prompt injection risk: Columbia University research found AI agents are fooled by malicious links in 100% of cases, risking data leaks or phishing emails; Karpathy demonstrated a Reddit-scenario attack. DeepLearningAI
- ALE-Agent (Sakana AI): Coding agent ranked 21st out of 1000 in an AtCoder heuristic contest; strong at NP-hard optimization. The ALE-Bench dataset is open-sourced. Announcement
- Windsurf acquisition rumors: Unconfirmed reports that OpenAI and Microsoft are in talks to acquire Windsurf, possibly for code-generation tool integration. Source
- Google Veo 3: Now available to AI Pro/Ultra subscribers in 70+ markets; text-to-video, outperforming its predecessor. Announcement
- Qwen3 on Apple Silicon: Alibaba released Qwen3 in MLX format with 4/6/8-bit and BF16 quantizations. Release
- Gemma 3n on mobile: Under 10B parameters with LMArena score above 1300; runs on mobile devices, a new edge-computing option. Review
- Hunyuan 3D 2.1: Tencent open-sourced its PBR 3D generation model; online demo on Hugging Face. Link
- RunwayML Gen-4 References: Generates new scenes from existing video, speeding VFX workflows. Demo
- MiniMax-M1 quantization: Community tests show quantized MiniMax-M1 needs ~240GB VRAM for 65k context vs 700–800GB for FP8; Unsloth offers optimization. Unsloth Discord
- DeepSeek architecture course: 29-video tutorial building DeepSeek from scratch, covering attention, MoE, and quantization. YouTube
- AI wrapper startups: Reddit discussion argues defensibility requires vertical data, UX, or tooling integration; Vercel vs AWS as a case study. LocalLLaMA
- UK universities: Nearly 7,000 students caught using AI to cheat, highlighting detection-tool limits and reform needs. The Guardian
- German enterprises: IFO data shows 40% of German companies use AI and 18.9% plan to; productivity gains are notable, though cultural resistance persists. ifo.de
- Unsloth DeepSeek-R1: 69.4% accuracy, 426s per case, ~40% faster than the API version. Unsloth Discord
- Cursor editor: Reports of failed command execution and ~10–15% credit waste, especially on Windows. Cursor Discord
- Hugging Face ethics: Members refusing AI-generated feedback sparked debate on data provenance and attribution. Discord
- LlamaIndex: LlamaExtract users report data loss when parsing documents; official advice is to check formatting and re-upload. Discord
- Torchtune: DTensor cross-mesh errors during Llama4 Maverick fine-tuning; community suggests raising NCCL_TIMEOUT or disabling fused optimizers. Discord
Multi-Agent Systems & Safety
AI Agents & Competitions
Industry News
Model Updates & Performance
Open-Source Community & Technical Discussion
AI Applications & Policy
Community Notes & Technical Issues
*Source: Easy AI Daily (zhichai.net)*