Highlights
Top Insights
Most consumers pick one general AI assistant and stick with it. Fewer than ~10% of ChatGPT weekly users also visited another major assistant, and only ~9% of consumers pay for more than one subscription across the major players.
Google’s growth is accelerating fast. Desktop users grow ~155% YoY for Gemini vs ~23% for ChatGPT, with acceleration over the last five months, driven heavily by viral creative models (Nano Banana).
AI browsers are emerging as the next battleground UI: such as Perplexity’s Comet, Dia, and OpenAI’s Atlas.
xAI’s “companions + animated personalities” is positioned as a step-change in engagement strategy. Not just voice companions: fully animated characters + edgier personality design.
Source: State of Consumer AI 2025: Product Hits, Misses, and What’s Next (a16z)
Top News
1. NVIDIA unveiled its open Nemotron 3 models (Nano, Super, Ultra) with a new hybrid MoE architecture to deliver higher-throughput, lower-cost, high-accuracy agentic AI.
2. OpenAI’s GPT-Image-1.5 model delivers more reliable instruction-following and precise image edits.
3. ChatGPT now lets developers submit apps for review and publication.
4. Gemini 3 Flash is Google’s newest Gemini 3 model that delivers near–frontier reasoning and multimodal capability.
5. Manus 1.6 introduces the powerful Max agent with significantly higher autonomous task success and user satisfaction.
6. The Grok Voice Agent API lets developers build fast, multilingual, cost-efficient voice agents.
Additional Insights
1. 2025 Breakthrough of the Year (Science)
Science’s 2025 Breakthrough of the Year highlights the seemingly unstoppable global surge of renewable energy, driven primarily by rapid, low-cost expansion of solar and wind power led by China, which now dominates manufacturing of solar panels, wind turbines, and batteries and has made renewables the cheapest source of electricity in much of the world. Renewables surpassed coal in global electricity generation this year, covered all growth in electricity demand, and are beginning to slow emissions growth—particularly in China—despite continued fossil fuel use and political headwinds elsewhere. Falling costs, massive scale, and energy-security benefits are accelerating adoption across the Global South, from rooftop solar in Africa to grid-scale projects in Asia. While challenges remain—grid infrastructure, storage, heavy industry, and lingering coal dependence—the decisive shift is economic rather than ideological: renewables now win on price and reliability, making the long-term decline of fossil fuels increasingly likely and marking a historic inflection point in the global energy system.
2. How Agents Are Accelerating the Next Wave of AI Value Creation (BCG)
BCG argues that the next wave of AI value will come from agentic AI (systems that don’t just predict or generate, but execute end-to-end workflows using judgment grounded in a company’s institutional knowledge) shifting AI’s payoff from basic productivity to durable differentiation. They outline a CEO roadmap: Now, redesign work from a “zero-based,” outcome-first lens (don’t automate the old process), embed proprietary business context through a “context fabric” of objectives, resources, and constraints, and standardize on a shared enterprise AI platform that enables “freedom within a frame” while maintaining controls and reuse. Next, prepare for major disruptions: a rebalanced operating model (fewer entry-level “first-draft” tasks, more human-agent orchestration, fewer management layers, and evolving governance as agents gain decision rights) and a recalibrated tech strategy as spend shifts from labor to platforms, increasing dependency risks and the need for modularity. Always, concentrate on a few “reshape and invent” big bets, combine predictive + generative + agentic AI, follow the 10/20/70 emphasis on people/process change, keep transformation business-led in partnership with IT, and continually strengthen the data foundation—because as AI becomes ubiquitous, advantage will belong to firms that integrate agents into distinctive processes, proprietary intelligence, and new revenue models.
3. Meet your robotic coworker (McKinsey)
Humanoid robots like Agility Robotics’ Digit are suddenly practical not because of one breakthrough, but because several things finally clicked at once: better AI that can teach skills faster than engineers can hand-code them, smaller/lighter motors, and EV-driven gains in vision and batteries—plus a real labor crunch in repetitive and injury-prone jobs. The surprising twist is that the goal isn’t “human-like” robots, it’s robots built for human spaces (same aisles, same shelf height), and the big unlock is safety: today they’re often kept in “work cells” like safety cages, but the future depends on “cooperative safety,” where the robot can recognize people nearby and automatically limit power to avoid harm. Digit 4.0 also hints at what “robot coworker” really means operationally: it can self-dock to recharge (roughly 50 minutes work, 10 minutes charge) and keep going with minimal supervision—shifting humans away from the most punishing tasks and creating new roles like “robot fleet manager,” while keeping tight guardrails because AI can still hallucinate and a 160-pound machine can’t afford mistakes.
Innovation Radar
1. AI Model Releases and Advancements
NVIDIA unveiled its open Nemotron 3 models (Nano, Super, Ultra) with a new hybrid MoE architecture to deliver higher-throughput, lower-cost, high-accuracy agentic AI, alongside open datasets and RL tools for building specialized multi-agent systems (NVIDIA).
AI2 released updated OLMo 3.1 Think 32B and OLMo 3.1 Instruct 32B (alongside OLMo 3-Base), achieving sizable benchmark gains by extending RL training for Think and scaling its Instruct-tuning recipe to 32B, with checkpoints available now and API access coming soon (VentureBeat).
Meta’s SAM Audio is a unified, prompt-driven audio separation model that isolates target sounds from real-world mixes using text, visual, and time-span prompts, outputting both the isolated audio and the residual to enable flexible editing without training sound-specific models (MarkTechPost).
OpenAI has launched an upgraded ChatGPT Images experience powered by its new GPT-Image-1.5 model, delivering more reliable instruction-following and precise image edits that preserve key details, faster (up to 4×) generation, improved text rendering/overall quality, and broader availability in ChatGPT and via the API (OpenAI). ChatGPT now lets developers submit apps for review and publication, enabling users to discover, connect to, and use chat-native apps directly within conversations (OpenAI). OpenAI introduced GPT-5.2-Codex with strengthened cybersecurity abilities (OpenAI).
Gemini 3 Flash is Google’s newest Gemini 3 model that delivers near–frontier reasoning and multimodal capability with much faster, cheaper “Flash” latency, and is rolling out broadly across Google products and developer/enterprise platforms (Google).
Alibaba has launched its Wan2.6 AI video generation models, enabling users to create multi-shot videos that preserve their own likeness and voice through a new reference-to-video technology claimed to be a first in China (Tech in Asia).
Mistral OCR 3 is a faster, cheaper, and more accurate OCR model that significantly improves extraction of handwriting, forms, low-quality scans, and complex tables, and is available via API and Mistral’s Document AI Playground (Mistral).
2. AI Tools and Features
Manus 1.6 introduces the powerful Max agent with significantly higher autonomous task success and user satisfaction, alongside new mobile app development capabilities and a Design View for interactive visual creation, enabling more complex work to be completed end-to-end with minimal supervision (Manus).
Thinking Machines Lab has made its Tinker fine-tuning API generally available and expanded it with support for the Kimi K2 Thinking reasoning model, OpenAI-compatible sampling from training checkpoints, and multimodal image input via Qwen3-VL, making it easier to fine-tune frontier LLMs/VLMs using simple Python loops while Tinker handles the distributed GPU training backend (MarkTechPost).
Zoom claimed a record 48.1% score on Humanity’s Last Exam by “federating” and orchestrating outputs from multiple top AI models (rather than training its own), prompting critics to argue it’s taking SOTA credit for an ensemble/traffic-controller approach (VentureBeat).
The Grok Voice Agent API lets developers build fast, multilingual, cost-efficient voice agents with real-time tool use, natural voices, and industry-leading intelligence and latency (x.ai).
Google’s Gemini now enables Android users to hear natural, real-time translations of over 70 languages directly through any headphones, with iOS support coming next year (ZDNET).
