Easy AI Daily News | November 20, 2025

Forum topic · 小凯 · 2026-03-27

Summary

This daily AI news digest from zhichai.net covers major November 20, 2025 releases and community discussions. Google launched Gemini 3 Pro Image (Nano Banana Pro) with 2-4K output, Search grounding, and improved text rendering (error rate cut from 56% to 8%). AI2 released Olmo 3, a fully open-source (Apache-2.0) model family with a 32B Think reasoning variant. Meta released SAM3 and SAM3D segmentation models with open-source licensing and 30ms inference. OpenAI shipped GPT-5.1 Codex Max with multi-context-window support, plus research showing GPT-5.1 accelerating scientific work. Other items include Perplexity's Comet browser, Cursor's beta debugging mode, Jetson Spark clusters, CUDA/DMA optimization discussions, SynthID watermark bypass reports, RAM price spikes, and community debates across LMArena, LM Studio, Unsloth, Eleuther AI, and tinygrad.

📅 AI Industry Updates — November 20, 2025

This is an English translation of the Easy AI daily digest (2025-11-20) from the Easy AI teaching project.

Model Updates & Releases

#### Google Releases Gemini 3 Pro Image (Nano Banana Pro) Supports Google Search grounding, 2-4K resolution, and text-in-image generation/editing. Pricing: $0.134 per 2K image, $0.24 per 4K image. Available via Gemini App/API, LM Arena, Hugging Face Spaces, and Together AI. Early demos show accurate infographics and chart annotation; text rendering error rate dropped from 56% to 8%.

Pricing details
Announcement
LM Arena addition
Hugging Face Spaces
Together AI integration
Error rate data
SynthID watermarking

#### AI2 Releases Olmo 3 Open Models Fully open source under Apache-2.0, including a 32B Think variant with long chain-of-thought and complex reasoning. Architecture retains post-norm; the 7B model uses sliding-window attention for KV cache efficiency, the 32B uses GQA. New RL infrastructure delivers 4x faster experiments, with emphasis on decontaminated evaluation (e.g., random-reward tests).

Hugging Face collection
Architecture analysis

#### Meta Releases SAM3 and SAM3D SAM3 unifies image/video segmentation with text and visual prompts, 2x performance improvement, 30ms inference. SAM3D enables 3D reconstruction from a single image. The data engine covers 4M phrases and 52M masks; code is open source and commercially usable.

SAM3 announcement
GitHub: segment-anything

#### OpenAI Ships GPT-5.1 Codex Max Designed for long, detail-heavy tasks; first to natively support multiple context windows (via compaction). Achieves SOTA on SWEBench; available only via ChatGPT plans, no API yet.

Release blog

#### Cogito 2.1 Enters WebDev Arena Deep Cogito's Cogito 2.1 ranks 18th overall and top-10 among open models on WebDev Arena. Hosted on Together and Fireworks; specific improvements undisclosed.

Model page
WebDev Leaderboard

Research & Science

#### OpenAI Publishes GPT-5.1 for Science Shares 13 early experiments showing GPT-5.1 accelerating math, physics, biology, and materials research; 4 helped solve previously unsolved problems. Includes blog, technical report, and researcher podcast.

Overview
arXiv paper thread

Tools & Platforms

Perplexity Comet browser: launches on Android, Mac, Windows with voice-first browsing; supports Kimi-K2 Thinking and Gemini 3 Pro. Pro/Max users can create slides, tables, and documents. Android launch
Cursor beta debug mode: adds a log-ingest server with automatic code instrumentation; agents verify hypotheses from logs instead of guessing. Details
MemMachine Playground: open-source Hugging Face Space with persistent AI memory for GPT-5, Claude 4.5, and Gemini 3 Pro. Playground
DSPy Proxy: new repo aryaminus/dspy-proxy, a proxy server built via a single prompt to simplify DSPy agent development. Repo

Hardware & GPU Tech

A user built a 6-node NVIDIA Jetson cluster for NCCL/NVIDIA development, prototyping B300-cluster workflows. Reddit thread
GPU MODE discussions: GEMM optimization (blog), texture vs. constant cache, AMD MI300X DMA collectives (+16% on large transfers) (paper), BF16 conversion issues.
Mojo 0.25.7 nightly shows a major regression on Mac M1: llama2.mojo throughput dropped from ~1000 to ~170 tok/sec; users asked the compiler team to investigate.

Safety & Jailbreaking

BASI community discusses jailbreaks against Gemini 3 Pro, Grok, and Claude 4.5 — including shell access on Grok and a trust-building bypass on Claude 4.5.
Users report SynthID watermarking can be bypassed via a "do nothing" reveal-edit prompt, or detected by directly asking whether an image is AI-generated.

Community Highlights

LMArena: debates on Nano Banana Pro quality, SynthID bypasses, GPT-5.1 vs. Gemini 3 Pro, Cogito 2.1's arena performance.
Perplexity AI: Gemini 3 Pro coding praised over Claude Sonnet 4.5; Comet RAM usage concerns; Antigravity called a "Cursor Killer."
LM Studio: EmbeddingGemma for RAG, Qwen3 thinking control, Mi60 GPU value, Vulkan crashes from model offloading.
Unsloth AI: Gemini 3 Chrome integration speed, Cogito GGUF links, RAM prices spiking (64GB at $400).
Yannick Kilcher: Skyfall AI's AI CEO benchmark (LLM long-horizon planning trails humans), SAM3D vs. DeepSeek, NVIDIA Q3 earnings.
Moonshot AI Kimi K2: Coding plan at $19 seen as pricey; SGLang tool-calling issues.
HuggingFace: KTOTrainer multi-GPU, inference endpoint 500 errors, Maya1 voice model, MemMachine.
Eleuther AI: KNN vs. quadratic attention, softmax of attention scores, IntologyAI's RE-Bench claim of superhuman expert performance and its reliability.
Nous Research: Gemma 3 hype (not AGI), world-model releases from DeepSeek/Qwen/Kimi, Nano Banana Pro infographics.
tinygrad: CuteDSL well received; post-update bug fixes.
Manus.im: case 1.5 Lite album-cover fix, Operator extension reinstallation loop bug.

Evaluability & Other News

Meta's SAM3 improves segmentation interpretability with a unified text/visual-prompt architecture.
OpenAI emphasizes evaluation rigor behind GPT-5.1's science experiments.
modelcontextprotocol.io migrates to community control ahead of a possible birthday-related outage.
OpenRouter users report 500 errors and agentic-LLM mid-run pauses; Grok 4.1 free until December 3.
RAM price surge discussion: 64GB kits at ~$400; buy now or wait?

---

*Source: Easy AI teaching project.*