Weekly AI Update: January 1-11 2026

The first 11 days of 2026 brought transformative developments in AI: DeepSeek published its updated R1 architecture revealing training secrets ahead of a February V4 launch targeting coding dominance, NVIDIA unveiled its Rubin platform with 10x cost reduction claims at CES 2026, and Trump’s executive order attempting to preempt state AI laws triggered immediate legal pushback. OpenAI reported that over 40 million people now use ChatGPT daily for healthcare questions, while the FTC reversed its first AI enforcement action under the new administration’s AI Action Plan.

🧠 Big Tech News

NVIDIA unveils Rubin platform with 6-chip AI supercomputer at CES 2026

NVIDIA launched its Vera Rubin platform on January 5 at CES, featuring six new chips delivering 10x reduction in inference token costs and 4x fewer GPUs needed for training compared to Blackwell. The system enters production in H2 2026 with Microsoft, AWS, Google Cloud, and CoreWeave as first deployers.

OpenAI reports 40M+ daily users asking health questions via ChatGPT

OpenAI’s January 2026 report “AI as a Healthcare Ally” revealed that over 40 million people worldwide use ChatGPT daily for health-related questions, accounting for 5%+ of all platform messages. Approximately 200 million of ChatGPT’s 800M weekly users engage with health topics weekly, with 70% of conversations occurring outside clinic hours.

OpenAI planning new audio model launch by end of Q1 2026

Reports emerged January 1 that OpenAI will release a new audio model by March 31, 2026, featuring more natural-sounding speech and improved real-time interactions. The model uses a new architecture beyond transformers, led by former Character.AI researcher Kundan Kumar.

⚖️ Politics & Legal Affairs

California companion chatbot safety law takes effect January 1, 2026

California’s new law requiring AI companion chatbots to implement safety protocols for minors took effect January 1. The law mandates protections against suicidal ideation content and requires disclosure to minors that they’re interacting with AI.

FTC sets aside Rytr AI enforcement order under Trump AI Action Plan

On January 8, 2026, the FTC reversed its 2024 final consent order against Rytr LLC, marking the first FTC action implementing the Trump Administration’s AI Action Plan. The Commission found the original allegations did not support Section 5 violations and that the order unduly burdened AI innovation. Source:

New York finalizes RAISE Act chapter amendments for January 2026

Following Governor Hochul’s December 2025 signing, lawmakers committed to approve chapter amendments in January 2026 replacing compute-cost thresholds with a $500M revenue requirement for covered frontier model developers. The law takes effect January 1, 2027.

42 state attorneys general demand AI safeguards from tech giants

42 state attorneys general issued a joint letter to Google, Meta, and Microsoft pushing for AI safeguards against sycophantic and delusional outputs, with specific protections for children including independent third-party audits reviewable by regulators.

🔬 Research & Development

DeepSeek expands R1 whitepaper by 60 pages, reveals training secrets

DeepSeek quietly released a v2 whitepaper on January 9 expanding documentation by over 60 pages, disclosing the full three-stage “Dev” training pipeline and admitting that Monte Carlo Tree Search (MCTS) and Process Reward Models failed for general reasoning tasks.

Startups & Investments

Reports: Meta acquiring Manus AI for $2-3 billion

Multiple sources reported in early January that Meta is acquiring Singapore-based Manus AI, valued at $2-3 billion, to strengthen advanced AI capabilities across Meta AI and WhatsApp. Manus grew from $500M valuation in April 2025 to over $100M ARR by mid-December.

Industry predictions: 2026 as “show me the money” year

Venture capitalists surveyed in early January predict 2026 will require enterprises to demonstrate concrete AI ROI. Menlo Ventures’ Venky Ganesan: “2026 is the ‘show me the money’ year for AI. Enterprises will need to see real ROI in their spend.”

🌍 AI News in EU & Sweden

European firms quietly adopt Chinese open models

Industry reports indicate European and American apps are increasingly shipping on top of Chinese open models including DeepSeek, Qwen3, and Kimi K2, with the lag between Chinese releases and Western frontier shrinking from months to weeks.

🧠 AI in Healthcare & Education

OpenAI has introduced ChatGPT Health, a dedicated health and wellness space inside ChatGPT that lets users securely connect medical records and data from apps like Apple Health, MyFitnessPal, and Function for more contextual answers

🤖 Robotics

NVIDIA declares “ChatGPT moment for physical AI is here”

At CES on January 5, NVIDIA CEO Jensen Huang announced “The ChatGPT moment for physical AI is here—when machines begin to understand, reason and act in the real world.” The company showcased partnerships with Boston Dynamics, Caterpillar, Agibot, and others.

🎮 Hardware

NVIDIA Rubin platform promises 10x inference cost reduction

NVIDIA’s Vera Rubin platform unveiled January 5 features six co-designed chips (Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet) delivering 10x reduction in inference token costs and 4x fewer GPUs for training MoE models versus Blackwell.

📊 Market Insights & Investment Trends

Industry consensus: 2026 is AI’s “show me money” year

Multiple VC firms and analysts surveyed in early January converge on theme that 2026 requires demonstrable ROI. Menlo Ventures: “Enterprises will need to see real ROI in their spend.” AT&T CDO: Agentic solutions face accuracy challenges across multi-step processes.

🧠 Adoption Trends & Consumer Behavior

Only 9% pay for multiple AI subscriptions, “winner take most”

Andreessen Horowitz December data (influential in January predictions) shows only 9% of consumers pay for more than one AI subscription, with ChatGPT’s 36% DAU/MAU nearly doubling Gemini’s 21%, signaling consolidation.

🧠 Research Paper of the Month

DeepSeek R1 Updated Whitepaper: Revealing the Training Recipe

Released: January 9, 2026 (arXiv v2) | 60+ pages of new documentation

DeepSeek expanded its R1 whitepaper from the original sparse documentation to a comprehensive 60+ page technical report revealing the complete training methodology that enabled the model to match OpenAI’s o1 performance at a fraction of the cost. This transparency marks a significant departure from typical frontier lab secrecy.

Key revelations:

Three-stage Dev pipeline: The paper introduces intermediate models (Dev1, Dev2, Dev3) showing exactly how each training stage affects performance, with detailed analysis of self-evolution where the model learns to reflect on and improve its own outputs.

Monte Carlo Tree Search (MCTS) and Process Reward Models (PRM) failed: DeepSeek explicitly admits these “hottest research directions in the industry” didn’t work for general reasoning tasks. MCTS works for chess or rigid math proofs but struggles with open-ended reasoning where steps are ill-defined—a finding that challenges Google DeepMind’s current investment strategies.

Chain-of-thought without explicit training: The model developed emergent reasoning patterns through reinforcement learning rather than supervised fine-tuning on reasoning traces, achieving step-by-step problem solving organically.

Superior performance claims: The updated paper emphasizes R1 not just matching but exceeding OpenAI o1-1217 on challenging mathematics, coding, and STEM reasoning tasks, with expanded evaluation across multiple benchmarks.

Real-world implications:

This level of transparency could reshape AI development practices industry-wide. If MCTS and PRM—approaches requiring massive compute—don’t actually improve general reasoning, labs may pivot toward simpler reinforcement learning methods. DeepSeek’s willingness to publish failures challenges the current culture of secrecy and could accelerate global progress by helping researchers avoid dead ends.

The paper also confirms that despite U.S. export controls on advanced GPUs, Chinese labs are finding algorithmic efficiencies that challenge assumptions about hardware requirements for frontier AI—a finding with significant geopolitical implications for AI policy.

🧠 Tools to Try

Claude Code – Anthropic’s deep-context agentic AI coding too

Kling 2.6 – Featuring Motion Control for precise character animation

Shopify SimGym – Simulate buyer behavior with AI shoppers