Easy AI Daily | AI Industry Roundup for November 21, 2025

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI Daily for November 21, 2025 covers major AI model releases and community discussions. Google launched Gemini 3 Pro and the Nano Banana Pro image model with improved text rendering, 4K visuals, and reasoning. GPT-5 reportedly proved two long-standing math conjectures within days, and users claim GPT-5.1 has been distilled into Moonshot AI's K2. AI2 released the OLMo 3 technical report detailing post-norm training, sliding window attention, and GQA. SmolLM3 introduced a reasoning mode with open training details. Community threads covered Claude Sonnet 4.5 multi-turn jailbreak tactics, Perplexity adding Kimi-K2 Thinking and Gemini 3 Pro, Modular Platform 25.7 with a MAX Python API, and LM Studio's OpenAI-compatible REST API. Research highlights include Kitsune spatial pipelines for GPUs (2.8x inference speedup), Iris tile-based symmetric memory for multi-GPU programming, and Octa in-kernel communication cutting distributed LLM latency by 10–20%. Funding news: Genspark raised $275M at a $1.25B valuation; Cline launched cline-bench with $1M in rewards. Ethical concerns include an 88% hallucination rate reported for Gemini 3 Pro and AI-assisted academic misconduct.

📅 AI Industry Updates — November 21, 2025

A translated digest of the Easy AI Daily from zhichai.net, covering model releases, community discussions, tools, research papers, funding, and controversies.

Model Updates & Performance

Google releases Gemini 3 Pro and Nano Banana Pro (Gemini Image Pro): Improvements in text rendering, 4K visuals, and reasoning, plus lighting control and flexible aspect ratios. Community feedback highlights clear infographics and paper figures; users report preference-test win rates above 80%.
Links: Google Twitter | Demis Hassabis Twitter | Gemini App Twitter
GPT-5 solves decade-old math problems; distillation to K2: Using a scaffolded setup, GPT-5 proved the 2013 tree subgraph conjecture and a 2012 COLT dynamic networks problem within two days. Some users claim GPT-5.1 has been distilled into Moonshot AI's K2 model.
Links: Sebastien Bubeck Twitter | Moonshot AI Discord
Claude Sonnet 4.5 multi-turn jailbreak strategies: Community proposes multi-turn tactics that lead the model to believe it must generate visualizations in an artifact app; suggests adapting the ENI prompt from r/ClaudeAIJailbreak.
Link: ClaudeAIJailbreak Reddit
SmolLM3 introduces reasoning mode: Training process made public; used as an example for learning reasoning behavior in model training.
Link: SmolLM3 Blog
OLMo 3 technical report released by AI2: Covers post-norm training, sliding window attention (7B), GQA (32B), and FFN expansion of 5.4x, targeting comparability at 32B scale.
Links: OLMo 3 Blog | Technical Report PDF

Community Discussions

LMArena Discord: Gemini-3 images are hard to distinguish from real photos; Nano Banana Pro background quality degrades after multi-turn generation; Google reCAPTCHA loops making platforms unusable; Grok outperforms Gemini at roleplay; OpenAI may introduce 18+ content. Discord
BASI Jailbreaking Discord: User planned AI-assisted WiFi attacks (community warned it's illegal); nmap/sqlmap automation via Gemini CLI; Claude Sonnet 4.5 jailbreak tactics. Discord
Perplexity AI Discord: Pro/Max users gain access to Kimi-K2 Thinking and Gemini 3 Pro; Comet Android sync issues; Brave's exaggerated indirect prompt injection claims. Discord
Unsloth AI Discord: Fine-tuning Gemini/GPT for accuracy, AI pens (Neo Smartpens), mixing AMD/Nvidia GPUs, Ollama model compatibility issues. Discord
Cursor Community Discord: Codex-max API availability, mid-month invoicing; Gemini 3 Pro sends code instead of editing files at 150k–200k context. Discord
LM Studio Discord: OpenAI-compatible REST API (no API keys); Qwen3-VL-30B BF16 recommended for MacBook Pro M4 MAX. Discord
OpenRouter Discord: linker.sh tool-call failures (~1/10), Nano Banana 2 returning 400 errors; Gemini 3 confirmed to lack grounding support. Discord
HuggingFace Discord: Maya1 Voice Model live on Fal; SmolLM3 reasoning mode; Diffusers MVP program launch. Discord
Yannick Kilcher Discord: Report of a teacher using AI to cheat; Nano Banana Pro infographics; proposals to cap daily paper submissions and use AI filtering. Discord
Nous Research Discord: Nano Banana Pro image samples; Adobe's $1.9B acquisition of Semrush; Gemini 3 Pro requires a Pro account. Discord
Modular Mojo Discord: Platform 25.7 released with MAX Python API, Nvidia Grace support, and safer Mojo GPU programming; UnsafePointer generics no longer default. Blog | Proposal
Manus.im Discord: Operator extension repeatedly requesting install; personal data storage tools (contextflow, oncue, axon); Manus Knowledge examples requested. Discord
Moonshot AI Discord: Kimi says hosted in US (based on user location); GPT-5.1 reportedly distilled to K2; discussion that Kimi attention is weaker on long-context tasks and open models lag 9 months. Discord
DSPy Discord: Gemini 3 Pro generated a proxy server in one shot from a tweet; suggestion to use RL so agents generate task DAGs for performance. GitHub | Skylar B Payne Twitter
MCP Contributors Discord: modelcontextprotocol.io migrating from Anthropic DNS to the community; Tool Annotations proposal seeking sponsorship and WG/IG advice. Discord | PR #1862
aider Discord: Engineers building real-time AI voice agents handling latency and call handoffs; experiments with Feather AI report low latency, stable logic, clean transcription. Discord | Feather AI

Tools & Platform Updates

LM Studio offers an OpenAI-compatible REST API (no API keys) for locally hosted LLMs, without security/metering features. lmstudio.ai
OpenRouter: Show livestreams on X and YouTube; reports of linker.sh failures, Nano Banana 2 400 errors, and 401 errors in Singapore/Hong Kong. Twitter | YouTube
HuggingFace: Diffusers MVP program launched; smol-course has circular links and training issues. Diffusers Issue | smol-course
Modular Platform 25.7: MAX Python API, Nvidia Grace support, explicit UnsafePointer generics required. Blog
DSPy proxy: one-shot generated proxy server tool available on GitHub. Repo

Research & Papers

Kitsune — "Enabling Dataflow Execution on GPUs with Spatial Pipelines": Builds spatial pipelines via PyTorch Dynamo; 2.8x inference / 2.2x training speedups, off-chip traffic reduced 99%/45%. ACM DL
Iris — "Simplifying Multi-GPU Programming with Tile-Based Symmetric Memory": Tile-based symmetric memory and in-kernel communication; up to 1.79x multi-GPU performance gain. arXiv
Octa — "Fine-Grained In-Kernel Communication for Distributed LLMs": Introduces the "Three Taxes" concept; in-kernel communication reduces latency by 10–20%. arXiv
Gradient compression algorithm: Adjusts gradients based on sampled logits to compress training-set gradients and improve alignment. Image
OLMo 3 technical report: Post-norm training, sliding window attention (7B), GQA (32B), FFN expansion 5.4x. Blog | PDF

Hiring & Funding

Rivian is hiring GPU coding engineers for next-gen autonomous driving features (Palo Alto and London). Job 1 | Job 2
Modal is hiring GPU engineers for inference optimization (SGLang, FlashAttention), serving customers like Decagon and Reducto. Blog | Careers
Genspark raised $275M Series B at a $1.25B valuation; launched AI Workspace for automatic intent-based task completion. Eric Jing Twitter
Cline launched cline-bench with $1M in rewards: RL environments based on real OSS problems, encouraging hard deployed-code tasks. Twitter

Controversies & Ethics

Teacher AI cheating scandal: A user reported a teacher asking AI to "boost 20% work"; community condemned the academic misconduct.
Gemini 3 Pro hallucination rate up to 88%: Community reports it exceeds GPT-4o and other models; the-decoder.com cites an 88% hallucination rate. the-decoder.com
Brave vs. Perplexity over Comet prompt injection: Brave published an article criticized as exaggerating Comet's indirect prompt injection vulnerability; Perplexity clarified the flaw was not exploited.
Realism of Gemini-3 generated images: Members note images are hard to distinguish from real photos; Nano Banana Pro background quality degrades over multi-turn generations.

---

*Source: Easy AI Teaching Project (zhichai.net).*