AI News in December 2025: Disney, DeepSeek, Runway and Meta

GPT-5.2 is the big model release of the week. I covered it separately and found it strong at coding, knowledge work and mathematics. That is my assessment, not an independent evaluation of every task; access and model availability described in this December 2025 video are historical.

Disney and OpenAI announced a partnership to bring selected Disney, Marvel, Pixar and Star Wars characters into Sora. I think fan-created stories around familiar characters could attract a large audience, though preventing inappropriate uses of licensed characters will be difficult. That is a concern, not a demonstration that any particular safeguard has failed.

The announced three-year license covered more than 200 animated, masked and creature characters, with selected fan-inspired videos intended for Disney+. The agreement excluded actors’ likenesses and voices. Disney also planned to use OpenAI services and make a $1 billion equity investment; the announcement said the transaction remained subject to definitive agreements and approvals. See [the companies’ announcement](https://openai.com/index/disney-sora-agreement/).

Around the same time, Disney sent Google a cease-and-desist letter. As reported by Variety, Disney alleged that Google copied its copyrighted works without authorization for AI development and distributed infringing outputs. These are Disney’s allegations, not a finding by a court. I view the OpenAI deal and the dispute with Google as evidence of different approaches to licensing.

TIME named the architects of AI its 2025 Person of the Year. The cover includes Mark Zuckerberg, Lisa Su, Elon Musk, Jensen Huang, Sam Altman, Demis Hassabis, Dario Amodei and Fei-Fei Li. I notice several other prominent figures are absent, and the composition recalls the famous photograph of construction workers above Rockefeller Center.

Adobe announced integrations for Photoshop, Express and Acrobat in ChatGPT, including natural-language image editing. I think placing familiar tools inside a conversational interface can bring them closer to users. My broader claim that ChatGPT could become a default gateway to the internet is a prediction, not a measured outcome.

The Information reported that DeepSeek was developing a future model using several thousand restricted NVIDIA Blackwell chips, citing six people familiar with the matter. The report alleged the chips were smuggled into China via equipment sent to data centers elsewhere and disassembled. These are reported allegations; this video does not independently verify the sourcing or quantities.

I speculate that the report raises questions about domestic chip capabilities and DeepSeek’s efficiency claims. Neither the reported chip sourcing nor the video establishes that Huawei chips are categorically insufficient or that DeepSeek falsified its numbers. DeepSeek has published technical methods that others can examine; the hardware allegation and the quality of those methods are separate questions.

I point to DeepSeek’s work on efficient attention and inference, and look forward to a future model release. My suggestion about a possible Lunar New Year launch is based on contemporary reporting and remains speculation rather than a confirmed schedule.

Runway introduced Gen-4.5, a text-to-video model, and its examples show complex scenes, reflections, animals, movement and expressive faces. I find the visual quality impressive relative to other tools. Curated promotional examples do not establish reliable physics or human realism in every generation.

Runway’s displayed Artificial Analysis text-to-video chart lists Gen-4.5 at 1,247 Elo as of November 30, 2025, ahead of the models shown. I notice that the truncated vertical scale visually exaggerates differences. The result applies to that particular leaderboard and date, not every video task or later ranking.

Bloomberg reported that Meta was considering a more commercially focused model, code-named Avocado, that might be offered as a closed model. I would be disappointed by a retreat from open-weight releases. But the report describes a possibility, and a company can charge for services built on open-weight models without necessarily closing the weights.

The report also describes changes in Meta’s investments and the use of third-party models during development. I interpret the reported move toward AI spending as a reassessment of earlier AR and VR investment, but that is my interpretation. Llama’s weights were released under a license with conditions, so “open weight” is a more precise description than assuming unrestricted open source.

Berman says Meta is reportedly drawing on third-party models such as Gemma and GPT-OSS in training. He sees the company’s infrastructure and distribution as strengths, even if he considers its then-current model lineup less competitive. The reported internal development details are not independently verified in the video.

Rivian, whose vehicle I own, presented plans for a third-generation autonomy computer and custom silicon with an in-house neural network engine. It also discussed adding lidar alongside cameras and radar. I am excited by the possibility of improved driver assistance, but a product roadmap is not proof that a vehicle can safely drive itself without supervision.

Tesla has favored cameras rather than lidar, while Waymo uses lidar among its sensors. I see Rivian’s multimodal approach as an interesting alternative, but the sensor choice alone does not establish comparative safety or autonomy. I look forward to seeing how the announced system develops.