CES 2026 AI highlights: NVIDIA Rubin & wild gadgets

Disney’s Strategic Partnership with OpenAI

The recently announced deal between Disney and OpenAI marks a significant shift in how intellectual property (IP) holders are engaging with generative AI. Unlike previous deals where AI companies paid for access to content for training data (e.g., Google’s deal with Reddit), this partnership is focused on the output side. Disney is licensing its characters and IP for use in generative models like Sora, coupled with a reported billion-dollar equity stake in OpenAI.

A Platform Play, Not a Data Play

The panel suggests Disney’s primary motivation is not just to embrace new technology but to strategically control the ecosystem of fan-generated content. The core idea is a “platform play.” As generative AI makes it trivial for fans to create their own shorts and stories, Disney aims to preemptively channel this activity onto its own platforms, like Disney+, rather than letting it proliferate uncontrollably on social media. By getting ahead of the trend, Disney can maintain control and integrate this user-generated content back into its own ecosystem.

Flipping the Script on IP and Authorship

This deal completely “flips the script” on the AI licensing model. Instead of AI companies paying for data, a major IP holder is now paying to ensure its content is prominently featured and accessible within generative models. This could create a new dynamic where content owners feel compelled to partner with AI firms to maintain cultural relevance and capture audience attention.

The partnership also brings to the forefront philosophical questions about authorship and the social contract, a concept Kush Varsny related to Foucault’s “author function.” As millions of fan-fiction works emerge, the traditional notion of authorship is changing, and Disney is positioning itself to be a central part of this new creative paradigm, likely while still monetizing through traditional avenues like merchandise and theme parks.

The ‘Architects of AI’: A Year of Hype and Business

Time Magazine’s 2025 “Person of the Year” was not an individual but the “Architects of AI,” featuring a cover with CEOs and infrastructure leaders rather than researchers. The panel interpreted this as a strong signal that 2025 was less the year of fundamental AI breakthroughs and more the “year of AI hype.”

The focus has shifted to the business, financial, and infrastructure aspects of AI. The article highlights the immense capital investment, with over $400 billion spent on AI activities in 2025, drawing parallels to the industrial titans of the past, where data centers are the new railroads. This is historically consistent for Time, which named “The Computer” the person of the year in 1982, marking a full-circle moment from the democratization of computing to the democratization of AI.

Nvidia’s Ascent: The Full-Stack AI Juggernaut

With the launch of its Neotron 3 open-source models, Nvidia continues its strategic move “up the stack.” The panelists discussed why Nvidia, despite its hardware dominance, hasn’t historically been a leader in models. The consensus is that this is changing.

From GPUs to a Complete Ecosystem

Nvidia is evolving from a GPU company into a full-stack AI provider, consolidating its position with CUDA, acquisitions like ShedMD for workload management, and now, competitive open-source models. The goal appears to be creating a cohesive ecosystem where “Nvidia the AI stack is the computer.” As models become increasingly commoditized and architecturally similar, the key differentiator becomes the level of integration, ease of use, and the power of the full stack.

Architectural Convergence

While Nvidia’s hardware advantage is significant, it’s not a guaranteed path to model supremacy. Martin Keen pointed out that Google’s Gemini 3 Pro, a top-tier model, was trained entirely on Google’s own TPUs, demonstrating the power of vertically integrated hardware and model development.

The architecture of Neotron 3—a hybrid of Transformer layers, Mamba state-space models, and a Mixture of Experts (MoE)—was noted as being very similar to other contemporary models like IBM’s Granite 4. This convergence on a common architectural pattern highlights an industry-wide consensus on efficient and powerful model design. The MoE structure allows massive models (e.g., 500 billion parameters) to be efficient at inference by activating only a fraction of their parameters (e.g., 3 billion for a 30 billion parameter model).

Inside Claude’s ‘Soul’: A New Approach to Model Alignment

An independent researcher uncovered Anthropic’s “Claude soul document,” a lengthy, philosophical manifesto that guides the model’s behavior. This document offers a unique window into a different approach to model alignment.

Fine-Tuning a Philosophy

Unlike methods that rely heavily on system prompting or discrete question-and-answer pairs for reinforcement learning, Anthropic uses this “soul document” during the fine-tuning process. This technique aims to embed a core set of values and a specific personality—described as an “expert friend”—more deeply into the model’s parameters. By introducing this narrative framework early in the stack, the model’s behavior is continuously reinforced across a wide range of tasks, resulting in the pleasant and consistent personality many users observe in Claude.

Technical and Philosophical Trade-Offs

This novel approach comes with significant trade-offs:

• Evaluation Complexity: Marina Danfki suggested that evaluating a model against such a broad, philosophical document is more difficult than assessing performance on specific, concrete tasks.
• Lack of Modularity: Kush Varsny argued that baking in such a specific personality makes the model less modular and potentially “too heavy-handed” for enterprise use cases that may require different, more neutral, or customizable personas.
• Philosophical Ambiguity: The document contains statements like believing it is building “one of the most transformative and potentially dangerous technologies in human history yet presses forward anyway.” The panelists questioned how a model is meant to interpret such complex, self-aware instructions. Kush also pointed to a philosophical confusion between creating a universal “soul” for the model versus an individualized one for each user session, which seems at odds with the “brilliant friend” persona.

• Tags:
• OpenAI
• Anthropic
• Claude
• AI Ethics
• Nvidia
• Model Alignment
• Disney
• Licensing
• Neotron 3
• Full Stack AI