Easy AI Daily News | November 21, 2025: Gemini 3 Pro Launch, OLMo 3 Report, and More AI Updates

Key points

Models and Performance

• Google released Gemini 3 Pro and Nano Banana Pro (Gemini Image Pro): improved text rendering, 4K visuals, and reasoning; supports lighting control and flexible aspect ratios. Community feedback highlights clean infographic and paper-figure generation, with user preference win rates reportedly above 80%.
• GPT-5 math results: After scaffolding, GPT-5 proved the 2013 tree subgraph conjecture and a 2012 COLT dynamic network problem within two days. Users claim GPT-5.1 has been distilled into Moonshot AI's K2 model.
• Claude Sonnet 4.5 multi-turn jailbreak strategies circulating in the community, framing the AI as needing to produce visualization outputs in an artifact app.
• SmolLM3 introduced a reasoning mode with a public training process.
• OLMo 3 technical report (AI2): post-norm training, sliding window attention (7B), GQA (32B), FFN expansion 5.4x, targeting comparability at the 32B scale.

Community and Platform Updates

• Perplexity AI: Pro/Max users now have access to Kimi-K2 Thinking and Gemini 3 Pro.
• Modular Platform 25.7: introduces MAX Python API, Nvidia Grace support, safer Mojo GPU programming; UnsafePointer generics no longer default.
• LM Studio: offers an OpenAI-compatible REST API (no API keys) for locally hosted LLMs; community recommends Qwen3-VL-30B BF16 for MacBook Pro M4 MAX.
• OpenRouter: reports of linker.sh tool-call failures (~1 in 10), Nano Banana 2 400 errors, and 401 errors from Singapore/Hong Kong; Gemini 3 reportedly does not yet support grounding.
• Cursor: discussion of Codex-max API availability; Gemini 3 Pro sending code instead of editing files at 150k–200k context.
• HuggingFace: Maya1 Voice Model launched on Fal; Diffusers MVP program started.
• MCP: modelcontextprotocol.io migrated from Anthropic DNS to the community; Tool Annotations proposal under discussion.

Research Papers

• Kitsune (spatial pipelines on GPUs via PyTorch Dynamo): 2.8x inference / 2.2x training speedups, off-chip traffic reduced 99%/45% — paper
• Iris (tile-based symmetric memory for multi-GPU): 1.79x performance gain — arXiv:2511.12500
• Octa (fine-grained in-kernel communication for distributed LLMs): 10–20% latency reduction — arXiv:2511.02168
• Gradient compression algorithm using sampled logits to adjust gradients, improving alignment.

Funding and Jobs

• Genspark raised $275M Series B at a $1.25B valuation, launching an AI Workspace.
• Cline launched cline-bench with $1M in rewards for hard deployed-code RL environments based on real OSS issues.
• Rivian is hiring GPU coding engineers for autonomous driving (Palo Alto, London); Modal is hiring GPU inference optimization engineers (SGLang, FlashAttention).

Controversies and Ethics

• Gemini 3 Pro reportedly shows a hallucination rate of up to 88% per the-decoder.com, higher than GPT-4o-class models.
• Gemini 3-generated images are increasingly hard to distinguish from real photos; Nano Banana Pro background quality degrades over multi-turn generation.
• Brave published an article exaggerating Comet's indirect prompt injection vulnerability; Perplexity clarified the flaw had not been exploited.
• Community debate over a teacher asking AI to "boost 20%" of work, condemned as academic misconduct.

*Source: Easy AI teaching project*