))
7th & 8th December - AI News Daily - OpenAI Accelerates GPT-5.2 Launch as NeurIPS Spotlights Evaluation Rigor
Manage episode 523208239 series 3670986
🌍 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: NeurIPS 2025 emphasized attention limits, compositional generalization, and rigorous evaluations. OpenAI fast-tracked GPT-5.2 while Google launched Gemini 3 Pro and Deep Think. Security audits found 30+ critical vulnerabilities in AI coding tools and a Gemini CLI exploit. The EU probed Meta over WhatsApp chatbot restrictions. Blue Origin's BlueGPT cut lunar hardware design time by 90%.
New Tools: Paper Trails launched as a research tracking platform. Memtrack introduced agent memory testing. Speechmatics open-sourced real-time diarization. OpenThoughts-Agent achieved small-model state-of-the-art on Terminal-Bench. Google Vertex AI Studio streamlined model development. NotebookLM Mobile added infographics and handwritten-note analysis.
LLM Updates: GPT-5.2 launches Dec 9 with improved speed. Gemini 3 Pro offers strong multimodal understanding. Gemini 3 Deep Think opened to Ultra subscribers. DeepSeek V3.2 improved long-context efficiency by 40%+. Rnj-1 (8B) reported near-state-of-the-art performance. Apple STARFlow-V unveiled flow-based video generation.
Research: NeurIPS honored attention and compositional work. New jailbreak via word associations bypassed controls. Vision systems used test-time compute for detail reasoning. LLMs transformed histopathology diagnosis. Common Crawl underpins many 2025 papers.
Industry: EU probes Meta's WhatsApp restrictions. UK ruling in Getty v Stability AI found limited infringement. Microsoft added Anthropic's MCP to Windows 11. AI drives hyper-personalized political campaigns. Security audits revealed Copilot and Amazon Q vulnerabilities.
Tutorials: Guides on long-context failures, agent memory patterns, multi-agent context engineering, LangChain's Deep Research internals, MoE router stability, and modality fusion.
Showcases: AxiomProver solved most Putnam 2025 problems. Energy Buddy used LangGraph for WhatsApp OCR routing.
Discussions: Calls for OpenAI breakthrough matching o3-preview. Missing ingredients hinder unified intelligence. Enterprise agents rely on simple workflows. Academia faces compute crisis. Drone swarms shift warfare economics.
🌍 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.
153 episodes
Share
