CES 2026 AI highlights: NVIDIA Rubin & wild gadgets
This episode explores the strategic implications of the Disney-OpenAI licensing deal, critiques Time Magazine's "Architects of AI" focus on business over research, analyzes NVIDIA's full-stack ambitions with the Neotron 3 model release, and delves into Anthropic's unique approach to AI safety with the "Claude Soul Document".
Disney’s Strategic Partnership with OpenAI
The discussion begins with the newly announced three-year licensing deal between Disney and OpenAI. This partnership is notable because it shifts the typical generative AI licensing model. Instead of OpenAI paying for training data (like its deals with the Financial Times or Google’s with Reddit), this deal focuses on licensing Disney’s IP for use in the output of generative models like Sora. Disney is also taking a significant equity stake in OpenAI.
The panelists suggest several motivations for this move:
- Platform Control: Disney aims to get ahead of the inevitable wave of fan-generated content. Rather than letting this content proliferate on platforms like X or Blue Sky, Disney wants to create a controlled ecosystem. By integrating these tools, they can encourage fans to create and share on Disney’s own platforms, likely driving engagement back to services like Disney+.
- Embracing the Inevitable: As one panelist quotes Disney’s CEO Bob Iger, the technology’s rise is “coming anyway, so we want to get ahead of it instead of fall prey to it.” By officially licensing its IP, Disney can shape the narrative and monetize a trend that would otherwise happen in unauthorized ways.
- Flipping the Script for IP Holders: This deal could signal a major shift for other IP owners. Previously, the concern was about models being trained on copyrighted data without permission. Now, the question becomes: “If your IP can’t be generated by leading AI models, are you missing out?” This could create a rush for other major IP holders to partner with AI companies to ensure their characters and worlds remain relevant in the age of generative content.
The conversation also touches on how this blurs the lines between official and unofficial content, potentially changing the “social contract of authorship” as fanfiction becomes easier to create and distribute on official platforms.
Time Magazine’s “Architects of AI”
The panel critiques Time Magazine’s choice for “Person of the Year”: the “Architects of AI.” The cover, featuring CEOs and infrastructure providers, is seen as a reflection of the year’s narrative.
- Hype Over Substance: The consensus is that 2025 was less the “year of AI” and more the “year of AI hype.” The focus was on financial deals, business strategies, and “cults of personality” rather than fundamental research breakthroughs.
- Business-Centric View: The selection of CEOs and infrastructure leaders (like AMD’s Lisa Su) over researchers or even data workers highlights that the “architecture” being celebrated is primarily financial and commercial. As one expert puts it, “capitalism is forefronted.”
- Historical Parallel: A panelist draws a parallel to Time’s 1982 “Person of the Year,” which was “The Computer,” popularized by the IBM PC. The 2025 choice represents a similar cultural moment, marking the point where AI has become mainstream and accessible to everyone.
NVIDIA’s Neotron 3 and the Full-Stack AI Play
NVIDIA’s launch of its next-generation open-source model, Neotron 3, is analyzed not just as another model release but as a key part of its broader business strategy. The panelists discuss why NVIDIA, despite its hardware dominance, isn’t always the leader in models.
- Moving Up the Stack: NVIDIA is strategically moving from a hardware (GPU) company to a full-stack AI provider. By developing its own models (Neotron), software (CUDA), and workload management tools (through acquisitions like ShedMD), NVIDIA is creating a deeply integrated ecosystem. The goal is to become the de-facto “computer” for AI, where the entire stack is optimized and controlled by NVIDIA.
- Commoditization of Models: As model architectures converge (Neotron 3’s hybrid Transformer/Mamba/MoE architecture is noted as being very similar to IBM’s Granite 4), the key differentiator becomes the integration, ease of use, and a comprehensive ecosystem. This full-stack play is NVIDIA’s answer to the increasing commoditization of the models themselves.
- The Hardware-Software Synergy: While NVIDIA’s vertical integration is a powerful strategy, it’s not a guaranteed win. The panel points to Google’s Gemini 3 Pro, which was trained entirely on Google’s own TPUs, as a prime example of how owning the hardware and building the model on top of it can lead to state-of-the-art results.
The Philosophy of AI Alignment: Anthropic’s “Claude Soul Document”
The final segment covers the discovery of Anthropic’s “Claude Soul Document,” a manifesto used to guide the behavior of its Claude model.
- Alignment Through Fine-Tuning: Unlike typical safety approaches that rely heavily on system prompts or simple reward models, Anthropic integrates this “soul document” during the fine-tuning process. This embeds a set of narrative, philosophical principles deep into the model’s parameters, aiming to create a more consistent and inherent “personality.”
- A Different Technical Approach: This method reinforces a value structure across all tasks, rather than just providing task-specific examples of good and bad answers. The result is a model that many find more “pleasant” and coherent to interact with.
- Trade-offs and Debates: The panel discusses the trade-offs of this approach.
- Evaluation Difficulty: It may be harder to quantitatively evaluate the model’s performance against specific benchmarks when alignment is based on a broad, philosophical document rather than discrete rules.
- Modularity vs. Baked-in Personality: Baking in a specific personality (e.g., an “expert friend”) makes the model less modular. For enterprise use cases, a more neutral or customizable model might be preferable to one with a strong, pre-defined “soul.”
- The Future of Prompting: This deep, fine-tuning approach represents a move beyond simple prompting. While prompting is a way to get information into a model, the industry is moving toward more robust methods for injecting information and desired behaviors earlier in the development stack, creating more reliable and less “fragile” systems.