Easy AI Daily | November 21, 2025

A translated digest of AI industry news from the Easy AI daily report.

Model Updates & Performance

• Google releases Gemini 3 Pro and Nano Banana Pro: The new Gemini 3 Pro and Nano Banana Pro (Gemini Image Pro) models improve text rendering, 4K visuals, and reasoning, with lighting control and flexible aspect ratios. Community feedback notes clean infographics and paper figures, with preference-test win rates over 80%. (Google on X | Demis Hassabis | Gemini App)
• GPT-5 solves decade-old math problems: With scaffolding, GPT-5 proved the 2013 tree subgraph conjecture and a 2012 COLT dynamic networks problem within two days. Users claim GPT-5.1 has been distilled into Moonshot AI's K2 model. (Sebastien Bubeck | Moonshot AI Discord)
• Claude Sonnet 4.5 multi-turn jailbreak: Community members propose a multi-turn strategy convincing the model it must produce visualizations in an artifact app, adapting the ENI prompt from /r/ClaudeAIJailbreak. (Reddit thread)
• SmolLM3 adds reasoning mode: Training process is public; the community uses it as an example for studying reasoning behavior. (SmolLM3 blog)
• OLMo 3 technical report: AI2 details post-norm training, sliding-window attention, and GQA, targeting comparability at 32B scale. (Blog | Report PDF)

Community Discussions

• LMArena: Gemini-3 images hard to distinguish from real photos; Nano Banana Pro background quality degrades over multi-turn generation; Google reCAPTCHA loops making platforms unusable; Grok better at roleplay, OpenAI possibly adding 18+ content. (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)
• Unsloth AI: Fine-tuning Gemini/GPT, AI pens (Neo Smartpens), AMD/Nvidia GPU combinations, Ollama model compatibility issues. (Discord)
• Cursor Community: Codex-max API availability, mid-month invoicing; Gemini 3 Pro sends code instead of editing files at 150k–200k context. (Discord)
• LM Studio: Offers an OpenAI-compatible REST API (no API keys); Qwen3-VL-30B BF16 recommended for MacBook Pro M4 MAX. (Discord)
• OpenRouter: linker.sh tool-call failures (~1/10), Nano Banana 2 400 errors; Gemini 3 does not yet support grounding. (Discord)
• Modular Mojo: Platform 25.7 released with MAX Python API, Nvidia Grace support, and safer Mojo GPU programming; UnsafePointer generics no longer default. (Blog | Proposal)
• Moonshot AI: Kimi hosted in the US based on user location; claims of GPT-5.1 distillation into K2; discussion of Kimi's weaker attention on long-context tasks and open-source models lagging ~9 months. (Discord)
• DSPy: Gemini 3 Pro generated a proxy server in one shot from a tweet (GitHub); suggestion to use RL for agent-generated task DAGs. (Skylar B Payne)
• MCP Contributors: modelcontextprotocol.io migrating from Anthropic DNS to the community; Tool Annotations proposal seeking sponsorship. (Discord | PR)
• aider: Engineers building real-time AI voice agents handling latency and call handoffs; experiments with Feather AI (featherhq.com) reporting low latency and clean transcription. (Discord)

Tools & Platform Updates

• LM Studio: OpenAI-compatible REST API for locally hosted LLMs, no API keys, no security/metering features. (lmstudio.ai)
• OpenRouter: Show livestreams on X and YouTube; reported API errors in Singapore/Hong Kong. (X | YouTube)
• Hugging Face: Diffusers MVP program launched with active community contributions; smol-course has circular links and training issues. (Issue | smol-course)

Research & Papers

• Kitsune — *Enabling Dataflow Execution on GPUs with Spatial Pipelines*: spatial pipelines via PyTorch Dynamo; 2.8x/2.2x inference/training speedups, 99%/45% off-chip traffic reduction. (ACM DL)
• Iris — *Simplifying Multi-GPU Programming with Tile-Based Symmetric Memory*: tile-based symmetric memory and in-kernel communication, 1.79x multi-GPU gains. (arXiv:2511.12500)
• Octa — *Fine-Grained In-Kernel Communication for Distributed LLMs*: "Three Taxes" framework; in-kernel communication cuts latency 10–20%. (arXiv:2511.02168)
• Gradient compression: algorithm adjusting gradients based on sampled logits to compress training-set gradients and improve alignment.
• OLMo 3 report: post-norm training, sliding-window attention (7B), GQA (32B), FFN expansion of 5.4x. (Blog | PDF)

Hiring & Funding

• Rivian hiring GPU coding engineers for next-gen autonomous driving features (Palo Alto & London). (Job 1 | Job 2)
• Modal hiring GPU inference-optimization engineers (SGLang, FlashAttention). (Careers)
• Genspark raised $275M Series B at a $1.25B valuation, launching an AI Workspace. (Eric Jing on X)
• Cline launched cline-bench with a $1M prize pool for RL environments based on real OSS issues. (X post)

Controversies & Ethics

• Teacher AI misconduct: A user reported a teacher using AI to "boost 20%" of work; community condemned it as academic dishonesty. (Discord)
• Gemini 3 Pro hallucinations: Reported 88% hallucination rate per the-decoder.com, exceeding models like GPT-4o. (the-decoder.com)
• Brave vs. Perplexity: Brave accused of exaggerating Comet's indirect prompt-injection vulnerabilities; Perplexity clarified the flaw was not exploited. (Discord)
• Realism of Gemini-3 images: Generated images increasingly indistinguishable from real photos; Nano Banana Pro background degradation in multi-turn generation. (Discord)

*Source: Easy AI teaching project.*