Easy AI Daily Digest | November 21, 2025

A roundup of AI industry updates compiled by the Easy AI teaching project.

Model Updates and Performance

• Google releases Gemini 3 Pro and Nano Banana Pro: The new image model (Gemini Image Pro) improves text rendering, 4K visuals, and reasoning, with lighting control and flexible aspect ratios. Community reports say it generates clear infographics and paper figures, with preference-test win rates above 80%. (Google on Twitter | Demis Hassabis | Gemini App)
• GPT-5 solves decade-old math problems: Using scaffolding, GPT-5 proved the 2013 tree subgraph conjecture and a 2012 COLT open problem on dynamic networks within two days. Some users claim GPT-5.1 has been distilled into Moonshot AI's K2 model. (Sebastien Bubeck)
• Claude Sonnet 4.5 multi-turn jailbreak: A community strategy frames the task as needing visualization output inside an artifact app, adapting the ENI prompt from r/ClaudeAIJailbreak. (Reddit)
• SmolLM3 adds reasoning mode: HuggingFace published training details; the community uses it as an example for learning reasoning behavior. (Blog)
• OLMo 3 technical report: AI2 details post-norm training, sliding window attention (7B), GQA (32B), and 5.4x FFN expansion, targeting comparability at 32B scale. (Blog | PDF)

Community Discussions

• LMArena: Gemini-3 images are hard to distinguish from real photos; Nano Banana Pro degrades backgrounds over multiple turns; Grok outperforms Gemini at roleplay; reCAPTCHA loops making platforms unusable. (Discord)
• Perplexity AI: Pro/Max users gain access to Kimi-K2 Thinking and Gemini 3 Pro; discussion of Comet Android sync issues and Brave's exaggerated indirect prompt injection claims. (Discord)
• Cursor Community: Codex-max API availability and mid-month billing; Gemini 3 Pro sends code instead of editing files at 150k–200k context. (Discord)
• LM Studio: Provides an OpenAI-compatible REST API (no API keys); recommends Qwen3-VL-30B BF16 on MacBook Pro M4 MAX. (Discord)
• OpenRouter: Reports of linker.sh tool-call failures, Nano Banana 2 400 errors, and regional 401 errors; Gemini 3 does not yet support grounding. (Discord)
• Modular: Platform 25.7 released with MAX Python API, Nvidia Grace support, safer Mojo GPU programming, and no default UnsafePointer generics. (Blog)
• Moonshot AI: Kimi confirms US hosting based on user location; users discuss K2's weaker attention on long-context tasks and open-source models trailing by ~9 months. (Discord)
• DSPy: Gemini 3 Pro generated a proxy server in one shot from a tweet (GitHub); RL suggested for agent-generated task DAGs. (Skylar B Payne)
• MCP Contributors: modelcontextprotocol.io migrating from Anthropic DNS to community control; Tool Annotations PR seeking sponsorship (PR #1862).
• HuggingFace: Maya1 voice model on Fal; Diffusers MVP program launched; smol-course has circular-link and training issues. (Diffusers issue | smol-course)

Tools and Platforms

• LM Studio: OpenAI-compatible REST API for local LLM hosting, without API keys or security/metering features. (lmstudio.ai)
• OpenRouter: Show livestreams on X and YouTube (X | YouTube).
• DSPy proxy: One-shot generated reverse-proxy tool (GitHub).

Research and Papers

• Kitsune — *Enabling Dataflow Execution on GPUs with Spatial Pipelines*: spatial pipelines via PyTorch Dynamo; 2.8x/2.2x inference/training speedups and 99%/45% off-chip traffic reduction. (ACM)
• Iris — *Simplifying Multi-GPU Programming with Tile-Based Symmetric Memory*: 1.79x multi-GPU performance gains. (arXiv:2511.12500)
• Octa — *Fine-Grained In-Kernel Communication for Distributed LLMs*: proposes the "Three Taxes" framework; 10–20% latency reduction. (arXiv:2511.02168)
• Gradient compression algorithm: adjusts gradients using sampled logits to compress training gradients and improve alignment.

Hiring and Funding

• Rivian: hiring GPU coding engineers for next-gen autonomous driving in Palo Alto and London. (Job 1 | Job 2)
• Modal: hiring GPU engineers for inference optimization (SGLang, FlashAttention). (Careers)
• Genspark: raised $275M Series B at a $1.25B valuation; launched AI Workspace. (Eric Jing)
• Cline: launched cline-bench, an RL environment based on real OSS issues, with $1M in rewards. (Twitter)

Controversies and Ethics

• Academic misconduct: a teacher reportedly asked students to use AI to "boost 20% work"; community condemned the practice.
• Gemini 3 Pro hallucination rate: reported at up to 88%, exceeding models like GPT-4o. (the-decoder.com)
• Brave vs. Comet: Brave's article on Comet's indirect prompt injection was described as exaggerated; Perplexity clarified the vulnerability had not been exploited.
• Synthetic images: Gemini-3 outputs are hard to distinguish from real photos; Nano Banana Pro background quality degrades over multi-turn generations.

*Source: Easy AI teaching project*