How Sora Highlights the Next Leap Toward AGI and Shifts AI Competition

Earlier this year the author shared a list of ten predicted trends for large models, and several have already materialized, including Gemini, Nvidia's Chat With RTX, and OpenAI's release of Sora. The following points outline the author’s perspective on why Sora is significant and what it implies for the future of AI.

1. Talent density and deep expertise win the AI race

While some claim Sora outperforms tools like Pika and Runway, the real advantage lies with companies that own core technologies, such as OpenAI. The notion that AI allows startups to operate as sole proprietors is proven naïve; strong talent and accumulated knowledge remain decisive.

2. AI amplifies human creativity rather than instantly replacing industries

Sora can generate high‑quality 60‑second videos, but the narrative, script, storyboard, and dialogue still require human input and prompt engineering. The technology is likely to become a powerful creation tool for advertising, movie trailers, and short‑form video platforms like TikTok, rather than a wholesale disruptor of those platforms.

3. Chinese large‑model capabilities still lag behind OpenAI’s GPT‑4

Although domestic models appear close to GPT‑3.5, the author estimates a gap of about one and a half years to GPT‑4‑level performance. He suspects OpenAI holds undisclosed advances—potentially GPT‑5 or self‑learning AIGC systems—meaning the China‑US AI gap may be widening.

4. Large‑language models now grasp and simulate the physical world

Unlike previous diffusion‑only video generators that treat frames as independent images, Sora combines LLM reasoning with diffusion, enabling it to understand physical concepts (e.g., a tank can crush a car but not vice‑versa). This two‑layer capability—world knowledge plus realistic simulation—represents a shift from 2‑D image manipulation to genuine physical reasoning.

5. Training on massive video data accelerates the path to AGI

By ingesting vast amounts of video—from movies to YouTube and TikTok—future models will acquire a richer understanding of the world than text‑only training allows. The author argues that this multimodal exposure could bring true AGI within a few years, not decades.

Implications for robotics and autonomous driving

The ability to model physical interactions will profoundly impact embodied AI, such as robot cognition and self‑driving systems, which currently focus heavily on perception without deep world understanding. Sora’s approach hints at a new generation of AI that can reason about cause, effect, and physics.

Overall, the author sees Sora as a proof‑of‑concept that large models equipped with multimodal, world‑model capabilities will become foundational tools across scientific domains—including biomedicine, chemistry, and mathematics—ushering in a new era of AI‑driven discovery.