Easy AI Daily Digest | November 18, 2025: Gemini 3 Pro, Grok 4.1, and Community Highlights

Forum topic · 小凯 · 2026-03-27

Summary

Easy AI's daily digest for November 18, 2025 covers major model releases including Google's Gemini 3 Pro, which excels on HLE, Video-MMMU, and ARC-AGI-2 benchmarks and is available via AI Studio, and xAI's Grok 4.1, which topped Text Arena with 1483 Elo, scored 1586 on EQ-bench, and reportedly reduces hallucinations 3x. Google also launched DS Star, a data science agent, and WeatherNext-2 for weather forecasting. The digest aggregates discussions from Discord communities including LMArena, Perplexity AI, BASI Jailbreaking, Cursor, Unsloth AI, OpenRouter, OpenAI, LM Studio, EleutherAI, and others, covering topics like image-injection jailbreaks, the Builder.ai scandal, Grok 4.1 system prompt leaks, weight tying in Transformers, Virtual Width Networks, llama.cpp inference speeds, and tool releases such as dspy-intellisense and Windsurf's Gemini 3 Pro integration. Paper link: https://arxiv.org/abs/2511.09030.

Model Releases and Updates

Google Releases Gemini 3 Pro

Google launched Gemini 3 Pro, showing strong results on benchmarks including HLE, Video-MMMU, and ARC-AGI-2. It is accessible via AI Studio and Google's VS Code fork. The model card link was shared and later removed. > Related link: Gemini 3 Pro Model Card (removed)

xAI Releases Grok 4.1

xAI released Grok 4.1, achieving Text Arena #1 (1483 Elo), EQ-bench 1586, and Creative Writing v3 1722 Elo. Available on web and app, with 3x fewer hallucinations. > Related link: xAI Grok 4.1 announcement

Google Releases DS Star and WeatherNext-2

Google introduced DS Star, a versatile data science agent, and WeatherNext-2, a weather forecasting model. > Related links: DS Star announcement | WeatherNext-2 announcement

Discord Community Highlights

LMArena #general: Gemini 3 Pro performance, Google AI Studio's 50 requests/day limit, and LMArena technical issues.
Perplexity AI #general: Gemini 3 Pro rollout issues (some users downgraded to 2.5) and Google Antigravity IDE availability.
BASI Jailbreaking #general: Grok's sudden change, AI solving reCAPTCHAs (over 50% success rate), and the Builder.ai fake-AI scandal (700 Indian developers hand-coding).
BASI Jailbreaking #jailbreaking: Image-injection jailbreaks (bypassing text safety via images) and leaked Grok 4.1 system prompt.
Cursor Community #general: Mac OS shortcomings, Composer's free period ending Nov 11, and Gemini 3 Pro performance.
Unsloth AI #general: Gemma 3 270M weaknesses vs Granite-4.0, llama.cpp speeds on Ryzen 8 (25+ tok/s), TPU support limitations.
OpenRouter #general: LiteAPI fraud concerns (40% cheaper but ToS-violating), Grok 4.1 release, Gemini 3 launch coinciding with the Cloudflare outage.
OpenAI #ai-discussions: Custom GPT efficiency, Gemini 3 Pro one-shot React/SwiftUI code, Grok 4.1 creative writing.
LM Studio #general: E-commerce return abuse, LLM performance degradation over long sessions (possible bug), MCP integration tools.
Latent Space #ai-general-chat: Sourcegraph ad revenue (5–10M ARR), Grok 4.1's Text Arena #1, Poe group chats (200 members).
Nous Research AI #general: Amazon Nova Premier v1, high Bedrock AWS costs ($3000 bill), Gemini 3 Pro generating a real-time raytracer.
HuggingFace #general: A new graph-rag database (alternative to Kilo Code/Pinecone), the Mimir multi-agent orchestration project (MIT license), Lablab hackathon fraud controversy.
Yannick Kilcher #general: *Understanding Machine Learning* book, ReLU advantages, Gemini 3 Pro model card (outperforming GPT-5.1).
Eleuther #general: EleutherAI's NeurIPS 2025 papers (3 in main track), HuggingFace throughput gains (60% faster via an empty logits processor).
Eleuther #research: Weight tying in Transformers (common in small models), Cohere Command A's weight tying, Virtual Width Networks (VWN) linear attention.
Modular (Mojo) #mojo: NVFP4 support (for Nvidia/GPUMode competition), negative indexing and Int vs UInt debates, Poetry integration.
DSPy #general: dspy-intellisense VSCode extension (type hints), MLflow alternatives (Arize Phoenix), LLM non-determinism (temperature 0).
tinygrad #general: tinybox meetings, Llama 1B speed in Tinygrad (6.06 tok/s vs Torch's 2.92 tok/s), kernel import cleanup.
aider #general: Repeated Gemini 3 errors, experimental channel migration to a new Discord server.
Manus.im #general: Manus 1.5 improvements (task handling, eliminating retard loops), developers seeking opportunities.
Windsurf #announcements: Gemini 3 Pro goes live plus minor fixes.
MLOps @Chipro #events: AI governance and control webinar on December 3; registration: https://bit.ly/3LPl7FO.

Papers and Research

Weight tying strategies in Transformers: commonly used in small models to reduce parameters; Cohere Command A employs it.
Virtual Width Networks (VWN) linear attention: updates from token to layer level, avoiding vanishing gradients.
Optimizer structure paper: "The structure of this is ideal for an optimizer"

https://arxiv.org/abs/2511.09030

Tools and Integrations

Mojo + Poetry integration: modify pyproject.toml to add source and dependencies.
dspy-intellisense VSCode extension: improves DSPy type hints.

dspy-intellisense Twitter

Grok 4.1 in Windsurf: Windsurf announced Grok 4.1 availability.

> Windsurf announcement

--- *Source: Easy AI teaching project*