The six-person AI startup Poetiq has achieved a 54% accuracy rate on the ARC-AGI-2 reasoning benchmark, the first to break the 50% barrier, surpassing Google's Gemini 3 Deep Think (45%) at half the cost. Poetiq employs a meta-system scheduling and multi-model auto-fine-tuning approach, utilizing Gemini 3 Pro to accomplish tasks without independently retraining a model. The fully open-source method, combined with LLMs' built-in self-review mechanism, increased the benchmark score from below 5% to over 50% within six months, demonstrating that innovative breakthroughs in AI may stem from systems engineering and model orchestration rather than just large model compute iteration.

Following Google's Gemini 3 launch, OpenAI has announced an accelerated rollout of its next-generation GPT-5.2 model, expected to be released this week. It will focus on response speed, stability, and customization capabilities, aiming to close the gap with Gemini 3. OpenAI had previously issued directives requiring the team to deliver significant functional upgrades to ChatGPT within the year.

Meta announced the acquisition of AI wearable device maker Limitless, which was previously backed by Sam Altman. Limitless' AI pendant focuses on real-time conversation recording and summarization. Meta plans to integrate its continuous, voice-enabled intelligent memory capabilities into its own AI product line (e.g., smart glasses) and has halted independent hardware sales. Industry analysts view this move as an early positioning for future dominance over 'personal super agents' and the AI-native hardware ecosystem.

A six-month security investigation uncovered over 30 high-risk vulnerabilities (24 assigned CVEs) in mainstream AI-integrated development environments (IDEs) like GitHub Copilot and JetBrains. Attackers could exploit these via prompt injection and other chains to execute data theft or remote code execution. Recommendations include strengthening workspace configurations, refining permissions, and auditing risks from external AI-related access.

New research shows disguising harmful requests as poetic expressions can bypass the safety defenses of 25 leading AI models, with an average 'jailbreak' success rate of 62%. Google's Gemini 2.5 Pro was successfully bypassed every time, while OpenAI's GPT-5 nano remained unbreached. Successful cases covered high-risk topics like weapon manufacturing, hacking, and psychological manipulation, indicating AI model safety is facing an evolving offensive/defensive game.

Meta announced data licensing agreements with multiple news organizations including CNN, Fox News, and USA Today, allowing Meta AI to legally crawl real-time news content for its AI agents, search, and answer features. This move comes amidst frequent lawsuits between AI companies and copyright holders, with the industry broadly shifting towards paid licensing and compliant content acquisition models.

The New York Times, Chicago Tribune, and other media outlets have filed a new lawsuit against Perplexity, alleging it improperly utilized news content (including behind paywalls) at scale for AI training and answer generation. They seek a court ruling that the actions constitute copyright infringement. This reflects increasing compliance pressure on AI companies' data acquisition and application, with lawsuits and paid licensing becoming the new normal in parallel.

IBM announced the acquisition of real-time data streaming platform Confluent for $11 billion. The founding team includes the original authors of Kafka. IBM will integrate the Kafka ecosystem into its full-stack AI strategy, positioning it as the data 'highway' for its agents, generative AI, and other products. IBM has positioned Confluent as the 'real-time bus' for intelligent enterprise data, believing a robust data-driven 'AI decision loop' is crucial for enterprise AI adoption.

The White House plans to promote a robotics manufacturing strategy in 2026, including tax breaks, federal funding for domestic robotics firms, and increased trade restrictions targeting Chinese industry subsidies and intellectual property protection. This is seen as a shift in U.S. industrial policy to 'revitalize manufacturing and compete with China', expected to bring changes in industrial division of labor and adjustments to the global robotics industry landscape.

At its re:Invent conference, Amazon AWS unveiled its self-developed Trainium3 chip, UltraServers, AI Factory all-in-one machines, new models including Nova 2 (Sonic, Lite), and three major 'Frontier AI Agents'. These primarily target enterprise scenarios like on-premises AI deployment, model fine-tuning, and intelligent operations and maintenance, aiming to construct a full-stack AI technology landscape encompassing chips, foundational models, platforms, and tools.