Easy AI Daily | December 9, 2025

A roundup of AI industry news from the Easy AI teaching project, covering model releases, training tools, research, agents, infrastructure, community projects, and industry moves.

Model Updates & Multimodal Progress

• Zhipu releases GLM-4.6V and GLM-4.6V-Flash — GLM-4.6V is a 106B MoE model for cloud/high-performance clusters, while GLM-4.6V-Flash is a 9B dense model for local, low-latency use. Both support 128k context and native multimodal function calling. The Flash API is free, and weights are open on Hugging Face.
• GLM-4.6V-Flash on Hugging Face | GLM-4.6V on Hugging Face
• Jina AI releases Jina-VLM-2B — A 2B-parameter multilingual VLM focused on charts, documents, and scene text. It averages 72.3 across 8 VQA benchmarks, with SOTA on MMMB (78.8) and multilingual MMBench (74.3).
• Announcement tweet
• Qwen3-4B shines locally — Reaches 70 tokens/sec on an RTX 2060 with strong coding ability, though users note an "overthinking" tendency; recommended paired with GLM-4.6V-Flash.
• DeepSeek V3.2 improves reasoning — Supports interleaved reasoning with strong RoO code performance. Pricing: $0.28/M input tokens, $0.45/M output tokens; reported as more stable than Kimi.

Training & Optimization Tools

• Hugging Face launches Claude Code skills for automated LLM training — Specify training tasks in natural language (e.g., fine-tuning Qwen3-0.6B); the skill handles data validation, GPU selection, and HF Jobs submission. Small runs cost roughly $0.30.
• HF skills training blog | LiorOnAI tweet
• Unsloth AI progress — Celebrates 10k Reddit members, fixes slow HF downloads, and releases Mistral Large 3 GGUF models for local inference.
• Unsloth Reddit | Mistral Large 3 GGUF
• DSPy TOON Adapter — Reduces token consumption, though nested-schema handling is weaker than BAMLAdapter; improved MMLU-Pro after GEPA optimization.
• TOON Adapter code
• LangChain Deep Agents evaluation tooling — Supports evaluation of long-running agents (Terminal Bench 2.0 mean 42.65%) with context-compression triggers.
• LangChain tweet

Research & Evaluation

• Google publishes Miras post-Transformer framework (NeurIPS) — Frames Transformers/RNNs as associative memory systems; outperforms Transformer, Mamba2, and DeltaNet on LM, reasoning, and long-context tasks, with 20% gains on long-text retrieval.
• Overview tweet
• AxiomProver solves 9/12 Putnam problems — A Lean-based system solved 9 problems within hours of the exam, emphasizing verifiability and formal pipelines; outperformed last year's top score.
• Announcement tweet
• NeurIPS mechanistic interpretability workshop — Chris Olah reflected on interpretability, arguing the field needs scalable tools rather than neuron-level analysis of single models.
• NeelNanda5 tweet
• MEMTRACK benchmark — Evaluates agent long-context memory via Slack/Linear/git scenarios; GPT-5 scored 60%, showing room for improvement in real tool environments.
• Release tweet

Agents & Workflows

• LangChain Deep Agents — Evaluates agent planning, filesystem, and sub-agent capabilities; Terminal Bench 2.0 mean 42.65%, with context compression support.
• Dexter 2.0 — An open-source financial research agent with planning and self-verification, built on LangChain for long-horizon financial analysis.
• Demo
• AI21 Maestro — Agent orchestration with multi-step planning, built-in verification, proprietary RAG, and execution graphs.
• Announcement
• Cursor Agent issues — Community reports of failures to create files and infinite loops; a temporary fix involves the approval button, but a permanent solution is needed.

Infrastructure & Hardware

• OpenAI Stargate drives DRAM shortage — The project may consume up to 40% of global DRAM output (~900k wafers/month) via deals with Samsung and SK hynix, pushing up DDR5 prices.
• Tom's Hardware report
• AMD 7900 XTX as budget AI GPU — Strong price/performance with good llama.cpp support; slightly weaker at ray tracing but excellent for local LLM inference.
• llama.cpp discussion
• Blackwell WGMMA incompatibility — Blackwell GPUs don't support WGMMA (sm_90a only), causing compile errors and requiring kernel rewrites; CUDA 13.1 introduces CUDA Tile to simplify programming.
• CUDA 13.1 blog
• RadixArk splits from SGLang — Focusing on AI scheduling, compilers, serving, and training pipelines to make frontier infrastructure more open.
• Announcement tweet

Community & Ecosystem

• Local LLM cluster: 8x RTX 3090 + 64-core EPYC — 192GB VRAM, 250GB RAM, ~$8k total; runs GLM-4.5 Air Q6_K at 49 tokens/sec via llama.cpp.
• Build details
• Vector database selection guide — Recommendations: HNSW (<10M vectors), Turbopuffer (large databases), pgvector (small/local), Chroma (lightweight), with criticism of some commercial DBs' efficiency.
• Comparison blog
• Key community contributors — Unsloth (fast fine-tuning), mradermacher (automated quantization), Bartowski (curated quants), and TheBloke (base models) recognized as core ecosystem contributors.
• Discussion
• picomon: AMD GPU monitoring — Open-source HF community tool, more reliable than nvtop (at some cost in precision).
• Code

Industry & Market

• Meta acquires Limitless (formerly Rewind) — Pendant users get 1 year of support; non-Pendant features like Rewind will be sunset.
• Announcement
• ARC Prize 2025 results — NVARC topped the leaderboard at 25.03%; the TRM paper won $50k; the $600k Grand Prize went unclaimed. Winning methods will be open-sourced.
• Results
• Manus.im billing issues — Users report lost credits and subscriptions auto-renewing into 2026, with slow support responses; suspected system bug.
• Sora 2 regional restrictions — Available in only 7 countries; VPN use violates the ToS and may lead to account bans.
• Supported countries

Fun & Memes

• ChatGPT "napping" meme — Screenshots of ChatGPT "pausing" mid-task amuse the community.
• Original post
• AI bug-fixing loop meme — Viral video capturing the frustration of AI repeatedly claiming a bug is fixed when it isn't.
• Video
• Ultra-detailed David Duchovny AI image — A 20-detail checklist image (PSG jersey, Parthenos sylvia, lens clouds, etc.) with verification.
• Tweet
• Nano Banana Pro fashion editorial workflow — A contact-sheet prompt workflow generating 6-frame fashion shoots with camera and styling constraints.
• Workflow

*Source: Easy AI teaching project*