Easy AI Daily | November 27, 2025

A roundup of AI industry news, model releases, research, and community discussions.

Agents & Tool Ecosystem

Anthropic Releases Persistent Agent Patterns and MCP Tasks Update

Anthropic proposed persistent agent practices (state checkpoints, structured artifacts, etc.); MCP released SEP-1686 "tasks" supporting background long-running tasks; LangChain clarified its framework–runtime–harness stack, positioning LangGraph as a runtime.

• Anthropic blog summary
• MCP tasks announcement
• LangChain stack clarification

Booking.com Deploys Production Agents for Customer Messaging

Booking.com built agents with LangGraph, Kubernetes, GPT-4 Mini, and Weaviate semantic search, processing tens of thousands of messages daily with a 70% satisfaction improvement.
• Technical deep dive

Perplexity Launches Memory and Virtual Try-On

Perplexity added user-level Memory (view/delete/disable) and a virtual try-on shopping feature.
• Memory announcement
• Virtual try-on

Model Updates & Performance

Claude Opus 4.5 Shines in Benchmarks

Opus 4.5 Thinking ranked first on LisanBench and topped Code Arena WebDev; the non-Thinking version showed regressions, with community reports of Python tool overuse. Claude.ai added automatic context compression.
• LisanBench results
• Code Arena leaderboard
• Context compression update

Alibaba Open-Sources Z-Image-Turbo Text-to-Image Model

A 6B-parameter model based on the Qwen3 4B text encoder, free on ModelScope and integrated into Hugging Face Diffusers, with performance approaching Seedream 4.0.
• ModelScope page
• Reddit discussion

FLUX.2 Series Released

FLUX.2 pro/flex models joined LMArena with improved visual quality, eliminating the "plasticky" look; competitive against Nano Banana Pro.
• LMArena announcement
• Comparison images

EGGROLL Speeds Up Evolution Strategies

EGGROLL uses low-rank perturbations to accelerate evolution strategies, supporting 100k+ populations and stable pre-training of recurrent LMs in large discrete systems.
• Technical overview

dnet Tackles Apple Silicon Memory Limits

Dria's dnet uses distributed inference, disk streaming, and UMA scheduling to run over-memory models on Apple Silicon clusters, resolving OOM issues.
• Announcement

Inference & Efficiency

LatentMAS Cuts Multi-Agent Communication Tokens

LatentMAS replaces text-based communication with latent vectors, reducing tokens by 70–84% and improving speed 4–4.3x without accuracy loss.
• Paper
• Summary

Reasoning Trace Distillation Lowers Costs

Training a 12B model on gpt-oss traces reduces token usage 4x and avoids repeated inference.
• Summary
• Demo

Multimodal & Generative Models

• PixelDiT: Dual Transformer (patch-level and pixel-level) achieving FID 1.61 on ImageNet 256x256 and GenEval 0.74. Paper
• Apple STARFlow-V: Normalizing-flow video generation supporting T2V/I2V/V2V with causal prediction and flow-score matching. Paper
• FLUX.2 Pro: Richer detail vs. FLUX 1 Pro, no more "plastic" look. Comparison
• Nano Banana 2: Improved structured image generation on StructBench; community prompt resources shared. Analysis | Resources

Open-Source Ecosystem & Evaluation

• Hugging Face downloads: Chinese models reached 17.1% of downloads, surpassing US models, led by DeepSeek and Qwen; multimodal models are popular. Overview
• METR: Considered the most trusted external evaluator by practitioners. Comment
• AI Security Institute: Published a positive-but-caveated case study evaluating whether Opus 4.5 would sabotage AI safety research. Thread
• Zhihu: Used Qwen2.5-VL-72B/3B with LoRA fine-tuning for multimodal recommendations, improving MMEB-eval-zh by 7.4%. Write-up
• New benchmarks: MultiPathQA (pathology navigation), MTBBench (oncology decisions), WER is Unaware (clinical ASR). Pathology | MTBBench | WER

Reddit Highlights

• Z-Image-Turbo buzz: Users note near-Seedream-4.0 performance; 6B parameters suits local deployment. Reddit
• Opus 4.5 ports ZBar to Swift 6: Solved a long-standing bug where other models failed. Reddit
• SWE-bench chart debate: Opus 4.5 leads at 80.9%, but chart design criticized. Reddit
• AI progress charts: Thomas Pueyo's "fun toy to AGI" chart questioned for rigor. Reddit
• Memes: Ilya Sutskever scaling quotes, Grok 4.1 unhinged replies, Gemini 3 satire. Singularity

Discord Community Discussions

• LMArena: Flux 2 vs. Nano Banana Pro comparisons; SynthID prevents nerfing. Announcement
• Perplexity: Thiel/Palantir concerns, Nvidia–OpenAI bubble talk.
• Unsloth: ERNIE developer challenge, ES HyperScale CPU training, Qwen3 fine-tuning issues. Devpost
• Cursor: Haiku for docs vs. Composer-1 for code; linting red-squiggle issues.
• GPU MODE: Triton kernels, NVFP4_GEMV leaderboard, NVRAR for multi-node inference. Paper
• OpenAI: ChatGPT political bias debates, Nano Banana comics.
• LM Studio: API endpoint errors, image captioning fixes. Docs
• OpenRouter: Opus overload, DeepSeek R1 removal, model fallback bugs. Fallback docs
• Nous Research: Psyche office hours, Suno–Warner deal, Blackwell INT/FP performance.
• Eleuther: Multi-stage LLM hallucinations, SGD shuffling debates, Emergent Misalignment replication. Paper
• Latent Space: Claude Code Plan Mode upgrades, Jeff Dean's 15-year ML retrospective.
• Yannick Kilcher: Information retrieval lecture, curriculum learning debates. Lecture
• HuggingFace: Inference API concerns, RapidaAI open-source voice platform. Rapida
• Modular: MAX examples, Python-based MAX controversy, Mojo API regressions.
• tinygrad: TinyJit kernel replay, random function implementation. Tutorial
• Moonshot AI: Kimi limits, canvas over chatbots.
• DSPy: dspy-cli open-sourced with FastAPI and MCP support. Repo
• MCP Contributors: New protocol version, namespace collision discussions.
• Manus.im: API quota errors affecting ~500 users.
• aider: Benchmark updates, Opus 4.5 upgrade survey, Bedrock model errors.

---

*Source: Easy AI teaching project.*