5th, 6th December - AI News Daily - Google Gemini 3 Deep Think Reshapes AI Reasoning as OpenAI Accelerates GPT-5.2

Send a text

🌍 INAI • The Open AI Hub

The Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge. 2 Million+ tools, models, agents, tutorials & daily news—free for all, updated every day.

https://github.com/inai-sandy/inAI-wiki

Top Highlights: Google Gemini 3 with Deep Think strengthens reasoning; OpenAI pushes GPT-5.2 amid model rivalry. OpenAI expands with data centers in India (with TCS) and Sydney (with NextDC). ZTE launches Nubia M153, the first fully agentic AI smartphone using Doubao AI. NHS England scales Brainomix 360 stroke imaging, halving treatment times and doubling thrombectomies. EU probes Meta/WhatsApp; NYT sues Perplexity over content use. Google launches Workspace Studio for no-code Gemini agents in Gmail and Docs. AWS unveils autonomous agents and commits up to $50B for AI infrastructure.

New Tools: Moondream Aerial Segmentation for geospatial mapping; vLLM v0.12.0 adds speculative decoding; Transformers v5 RC introduces any-to-any multimodal pipeline; Qwen3-TTS debuts 49 voices in 10 languages; OpenAI Offline AI for emerging markets; Atlassian Rovo MCP Connector links ChatGPT with Jira/Confluence. Google Workspace Studio enables no-code automation; WordPress Telex AI creates interactive features; Runway Gen-4.5 and Kling Avatar 2.0 expand video generation.

LLM Updates: Gemini 3 Deep Think rolls out with stronger reasoning; GPT-5.1 Codex Max launches with Cline integration; GPT-5.2 rumored imminent; Amazon Nova 2 models and Forge tools arrive; Mistral 3 leads open coding leaderboards; Claude 4.5 Opus tops AutoCodeBench V2; DeepSeek v3.2 cuts latency. Off-policy RL advances show improved training; Intel SignRoundV2 pushes ultra-low-bit quantization; reasoning models now dominate OpenRouter usage.

Research: Meta + KAUST MoS improves multimodal fusion; radiance mesh method enables editable NeRF rendering; compact hybrid-search index cuts RAG costs 91%; Anthropic SCONE-bench evaluates smart-contract security; AI Evaluator Forum launches independent assessments; NeurIPS highlights include EPO, GEPA, OpenThoughts. NIST and US CAISI release evaluation frameworks; TokenPowerBench shows 90% of LLM energy in prefill/decode; MIT adaptive scaling cuts compute 50%.

Industry: Snowflake-Anthropic $200M deal embeds Claude in Data Cloud; US judge orders OpenAI to provide 20M ChatGPT logs to NYT; EU antitrust probe into Meta WhatsApp AI; OpenAI tests ads in ChatGPT; 7AI raises record funding for AI security agents.

Tutorials: Answer.AI SolveIt releases practical AI playbooks; Anthropic launches Model Context Protocol guide; community roadmap for training open LLMs; Gemini 3 + Agno cookbook; Andrew Ng course on coding agents; CrewAI multi-agent course.

Showcases: Moondream aerial segmentation with meter-level precision; Gradium live speech humanoid robot; Bionic Awards AI film blending multiple tools; Kling O1 high-fidelity avatars; X-VLA two-hour cloth folding.

Discussions: Advanced math key to problem-solving; human-AI co-improvement over self-improvement; AI chatbots can sway voters; RL vs. prompt optimization debates; China's open models gain OpenRouter share; Yejin Choi warns of synthetic data risks; construct-valid evaluations urged.

Support the show

🌍 INAI • The Open AI Hub

The Intelligence Atlas → the world’s most comprehensive, open hub of AI knowledge. 2 Million+ tools, models, agents, tutorials & daily news—free for all, updated every day.

https://github.com/inai-sandy/inAI-wiki