At NRF, Google announced the open standard UCP (Universal Commerce Protocol), co-developed with Shopify, Etsy, Wayfair, Target, Walmart, and others, enabling AI shopping agents to collaborate across product discovery, order payment, and post-sales support via a unified interface. Based on UCP, the US will soon offer "direct checkout" in Search AI Mode and Gemini apps, using pre-filled Google Pay and Google Wallet shipping information to complete purchases, with future support for PayPal. Simultaneously, the brand launched the "Business Agent" virtual shopping assistant and is testing Direct Offers ads that let merchants provide exclusive discounts within AI conversations.

At NRF 2026, Google Cloud released Gemini Enterprise for Customer Experience (CX), integrating shopping guidance and customer service into a unified intelligent agent solution, offering configurable pre-built agents claimed to be deployable within days. The product supports multimodal inputs like images, video, and voice, and, with user consent, can execute multi-step tasks. It also provides a "Customer Experience Agent Studio" for building, testing, and deploying personalized support agents at scale, covering 40+ languages and empowering human agents to handle complex tickets. Kroger, Lowe's, Woolworths, and Papa Johns announced adoption, with Google stating customer data is not used for model training.

Regulators in Indonesia and Malaysia, citing Grok's generation of non-consensual sexualized deepfake content on X, have successively announced temporary access restrictions/blockades on the chatbot, demanding X provide explanations regarding risks and remediation. Reports indicate such content involves women and minors, triggering regulatory intervention on "digital security and human rights" grounds. In response, xAI has limited image generation and editing features to paid subscribers of its $8 per month plan. Meanwhile, EU and UK regulators continue to pressure the company, arguing a "paywall" is insufficient to address the platform's systemic risks.

LG AI Research has disclosed technical details of K-Exaone, developed for South Korea's "homegrown AI foundation model" initiative: through an improved hybrid attention mechanism, memory requirements and computational load are reduced by approximately 70% compared to the previous version, allowing it to run on A100-level GPU environments and lower deployment costs. In the government-led project's first evaluation, K-Exaone scored an average of 72 points across 13 tests, ranking first among five participating institutions; reports also state it entered the top ten, ranking seventh, on a global AI benchmark leaderboard. LG stated it will offer an API via its official website, free of charge until January 28.

Data and AI company Databricks has raised over $4 billion in a new funding round, valuing the company at $134 billion. The company plans to use the funds for product development, AI research, employee liquidity support, and strategic acquisitions. Its business focus is combining generative AI capabilities with enterprise proprietary data, disclosing over 20,000 global customers. The company also provided key operating metrics: Annual Recurring Revenue (ARR) reached $4.8 billion, a 55% increase year-over-year. In the race to integrate enterprise data platforms with GenAI, this massive funding round will further expand R&D and M&A opportunities for leading vendors.

Andreessen Horowitz (a16z) announced it has raised over $15 billion across five new funds to support artificial intelligence and "strategic technologies." Disclosed allocations include $6.75 billion for a growth fund, $1.7 billion specifically for AI infrastructure, and $1.12 billion for areas like defense, housing, and supply chain. a16z also stated its assets under management now exceed $90 billion. In a cycle of overall VC fundraising slowdown, large capital continues to concentrate with a few top-tier firms, signaling sustained, long-term capital supply for directions like compute, model toolchains, and security.

The French Ministry of Defense has reached an agreement with Mistral AI to provide AI models, software, and services to the armed forces, defense agencies, and related public sectors. Reports indicate the solution will be deployed on sovereign infrastructure controlled by France and fine-tuned using defense-specific data to meet operational and administrative needs. The collaboration is positioned as a move to strengthen "technological sovereignty," reducing reliance on external AI and cloud services for critical military and governmental scenarios. For Mistral, this represents a deployment case in a high-sensitivity, government-level scenario, also providing significant public sector endorsement and an application gateway for European native large model suppliers.

The Register reports that a group of industry insiders from major US AI companies have launched the "Poison Fountain" project, attempting to use "data poisoning" to counter large model training: they call on website operators to add links to "toxic" content containing incorrect code, etc., within their sites, aiming for AI crawlers to ingest it and degrade model quality. The project went live about a week ago, offering both regular HTTP and dark web .onion access, and encourages caching and forwarding. The report mentions inspiration from related Anthropic research: a small number of malicious documents can significantly lower model performance. The action also sparks discussion on training data credibility, synthetic data pollution, and "model collapse" risks.

A newsletter claims that OpenAI's GPT-5.2 solved Erdős Problem #397 (related to a central binomial coefficient equation), which had perplexed the mathematics community for about 30 years. After the model generated a proof, it used the "Aristotle" tool to perform formal verification using the Lean language. The newsletter cites Terence Tao's view: these proofs resemble original derivations more than reproducing literature but are problems solvable using standard techniques. The newsletter also provides a capability comparison: GPT-5.2 scores 77% on competition-level math tasks but only 25% on open-ended research tasks requiring "true insight." If accurate, this indicates LLMs are moving from problem-solving towards verifiable proof generation workflows.