Easy AI Daily | November 20, 2025
A digest of AI industry news for November 20, 2025, compiled from the Easy AI teaching project. Below is a structured English summary of the original Chinese-language roundup.
Model Updates and 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%
- Includes SynthID watermarking
- Links: Pricing | Announcement | LM Arena | Error rate data | Demo examples
- Fully open source under Apache-2.0 (code, data, training recipes, checkpoints)
- Includes a 32B Think variant with long chain-of-thought for complex reasoning
- Post-norm architecture retained; 7B uses sliding-window attention for KV cache optimization, 32B uses GQA
- RL infrastructure delivers 4x faster experiments; emphasis on decontaminated evaluation (e.g., random-reward tests)
- Links: Announcement reaction | Architecture analysis | HF collection
- SAM3 unifies image/video segmentation with text/visual prompts; 2x performance, 30ms inference
- SAM3D enables 3D reconstruction from a single image
- Data engine: 4M phrases, 52M masks; source code open for commercial use
- Links: SAM3 | SAM3D | GitHub
- Designed for long-horizon, detail-heavy tasks; first native multi-context-window support via compaction
- SOTA on SWE-Bench; available only via ChatGPT plans, no API yet
- Link: Release blog
- Deep Cogito's model ranks #18 overall, top-10 among open models; hosted on Together and Fireworks
- Link: Model page
- 13 early experiments shared; GPT-5.1 accelerates math, physics, biology, and materials science research, with 4 experiments helping solve previously unsolved problems
- Links: Overview | Paper thread
- Perplexity Comet browser: now on Android, Mac, and Windows; voice-first browsing; supports Kimi-K2 Thinking and Gemini 3 Pro; Pro/Max users can create slides, tables, and documents. Announcement
- Cursor beta debug mode: log-ingest server with automatic code instrumentation; agents validate hypotheses from logs rather than guessing
- MemMachine Playground: open Hugging Face Space supporting GPT-5, Claude 4.5, and Gemini 3 Pro with persistent AI memory. Playground
- DSPy Proxy: new repo
aryaminus/dspy-proxybuilt via one prompt with Gemini 3 Pro, simplifying DSPy agent development. GitHub - A user built a 6-node NVIDIA Jetson Spark cluster for NCCL/NVIDIA development and pre-B300 workflow prototyping (Reddit)
- GPU MODE discussions: GEMM optimization, CUDA texture vs constant cache, AMD MI300X DMA collectives (+16% for large transfers), BF16 conversion/TensorRT kernel gaps (DMA paper, GEMM article)
- Mojo 0.25.7 nightly regression: llama2.mojo throughput dropped from ~1000 to ~170 tok/sec on Mac M1; compiler team asked to investigate
- BASI community discusses jailbreak methods for Gemini 3 Pro, Grok (shell access obtained), and Claude 4.5 (trust-building bypass used to generate methamphetamine synthesis steps)
- Users found SynthID watermarks can be bypassed via reve-edit "do nothing" prompts, or detected by asking the model directly
- LMArena: debates on Nano Banana Pro text rendering, SynthID bypasses, GPT-5.1 vs Gemini 3 Pro, Cogito 2.1 rankings
- Perplexity AI: Gemini 3 Pro coding praised (vs Claude Sonnet 4.5); Comet RAM usage complaints; Antigravity app dubbed "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 price spike (64GB at $400)
- Yannick Kilcher: Skyfall AI's AI-CEO benchmark (LLM long-horizon planning lags humans), SAM3D vs DeepSeek, NVIDIA Q3 earnings
- Kimi K2: Coding plan pricing ($19) seen as high; SGLang tool-calling issues
- HuggingFace: KTOTrainer multi-GPU support, inference endpoint 500 errors, Maya1 voice model
- Eleuther AI: KNN vs quadratic attention, Seth's conjecture, softmax attention-score distributions; IntologyAI claims superhuman-expert RE-Bench results, debated by the community
- Nous Research: Gemma 3 hype (not AGI), upcoming world-model releases from DeepSeek/Qwen/Kimi
- tinygrad: CuteDSL well received; post-update bug fixes reported
- Manus.im: Case 1.5 Lite album-cover fix; Operator extension loop-reinstall bug
- modelcontextprotocol.io domain migrated from Anthropic to community control
- OpenRouter reported 500 errors and agentic LLM mid-run pauses; Grok 4.1 free until December 3
- RAM prices surging; 64GB kits reaching $400, sparking buy-now-vs-wait debates
AI2 Releases Olmo 3 Open LLM
Meta Releases SAM3 and SAM3D
OpenAI Ships GPT-5.1 Codex Max
Cogito 2.1 Enters WebDev Arena
Research and Science
GPT-5.1 for Scientific Research (OpenAI)
Tools and Platforms
Hardware and GPU Tech
Safety and Jailbreaking
Community Highlights
Other News
*Source: Easy AI teaching project (zhichai.net)*