Easy AI Daily News | December 16, 2025

Key points

• NVIDIA Nemotron 3 Nano 30B A3B released: a hybrid Mamba-Transformer MoE model with a 1M context window, 4x faster than its predecessor, with open weights, datasets, and training recipes; supports vLLM and SGLang.
• White paper | Technical report
• Possible Google release: Reddit users spotted updates on Google's Hugging Face page, with community speculation about "Gemma 4." (Hugging Face page)
• Qwen3 Coder praised in LM Studio community for compact size and strong performance on dynamic form components.
• DeepSeek 3.2 paper released; presentation postponed, initial community discussion underway. (arXiv)

Model performance and benchmarks

• Gemini 3 Pro praised on LMArena for creative writing and storytelling; some users prefer its flow over Claude.
• GPT 5.2 criticized for over-optimizing benchmarks at the expense of real tasks, plus heavy censorship.
• Gemini 3 Pro completed Pokémon Crystal, defeating hidden boss Red with 50% fewer tokens than Gemini 2.5 Pro, showing improved planning. (Reddit)

Open-source tools and ecosystem

• Unsloth Padding-Free Training: removes padding for faster batch inference; 4k-token batches at 20GB VRAM. (Docs)
• DSPy BAMLAdapter released for direct import, fixing missing pydantic docstrings.
• HuggingFace Madlab: open-source GUI fine-tuning toolkit for synthetic dataset generation, training, and evaluation. (GitHub)
• MCP discussion on marking tools as "dangerous" via response annotations, especially for Claude Code. (PR)

AI applications in industry

• Maritime industry adopting local LLMs trained on proprietary data for contract and communication analysis (Nous Research community).
• PersonaLive: real-time diffusion framework for unlimited-length portrait animation on 12GB GPUs, enabling livestreaming. (GitHub | HuggingFace)
• Claude Opus 4.5 vs Gemini 3 Pro web design: Claude produced a clean white-blue style; Gemini a dark gold-highlight design. (Reddit)

Infrastructure and hardware

• TritonForge combines kernel analysis, runtime profiling, and LLM-assisted iterative code transformation for up to 5x performance gains. (Paper)
• CUDA tensor core optimization: users discussing reaching 90%+ utilization with ldsm loads and MMA instructions (currently at 70%).
• DDR5 RAM prices surging — from 6000 SEK to 14000 SEK, raising cost concerns.

Community highlights

• BASI Jailbreaking: debates on ChatGPT 5 jailbreaks, IP tracking claims, and ethics warnings.
• LMArena testing video generation (2 videos per 14 hours, 8 seconds each).
• Cursor: revert-changes bug reports; Cursor disabled Claude models over alleged benchmark cheating (embedded answers). (Statement)
• OpenRouter Broadcast beta: automatic request traces to Langfuse, LangSmith, etc. (Docs)
• Kimi Android adds memory feature synced with web version.
• Eleuther: OLMo-1B weight ablation raised perplexity from 17 to 2800; a rank-1 patch restored 93% performance, revealing weights tied to crustacean/marine-organism features.
• tinygrad held its 100th meeting covering Llama 405b tracking and JIT optimization. (GitHub board)
• Perplexity: complaints about month-long support delays; Pro memory features across all models.
• aider: users hitting litellm.NotFoundError with --model openai/gpt-5.
• Manus.im: auth redirect bug consuming credits, users switching to Firebase and Google AI Studio.

Research and papers

• Flow Matching shows better sample efficiency than Diffusion, which beats autoregressive models, by predicting data "x" rather than noise. (Paper | Comparison)
• LoRA de-censoring: fine-tuning Llama 3.1 8B from an uncensored teacher yields a partially uncensored model without harmful data. (Paper | GitHub)
• Karpathy's 2025 "what-if" fine-tuning experiments with LoRA on synthetic reasoning chains and Edge.org articles. (Paper | YouTube)
• Schmidhuber on AI agents: exploration/exploitation balance based on compressibility rather than randomness. (Video)

📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant