Unlike simple classification (one pass), generative AI performs recursive inference. Each new token (word, pixel) requires a full pass through the model, turning a single prompt into a series of demanding computations. This makes inference a major, ongoing driver of GPU demand, rivaling training.

The perceived intelligence of video generation models is often an illusion. The heavy lifting is done by a large language model that rewrites simple user prompts into highly detailed scenes. The video diffusion model itself is less intelligent, simply executing these detailed instructions literally.

The computational requirements for generative media scale dramatically across modalities. If a 200-token LLM prompt costs 1 unit of compute, a single image costs 100x that, and a 5-second video costs another 100x on top of that—a 10,000x total increase. 4K video adds another 10x multiplier.

The user experience of early AI video tools is plagued by severe rate limits, a direct result of immense compute costs. This 'come back later' experience is a retention killer, contrasting sharply with the 'endless scroll' of successful platforms like TikTok. This economic reality is forcing AI labs to shift scarce compute resources from viral consumer apps to more valuable enterprise workflows.

LLMs excel at 'spatial aesthetics'—arranging elements on a static page. For video, they must learn 'temporal aesthetics,' where information is revealed over time without requiring eye movement. This is a key training challenge for creating compelling AI-generated motion content.

Traditional video models process an entire clip at once, causing delays. Descartes' Mirage model is autoregressive, predicting only the next frame based on the input stream and previously generated frames. This LLM-like approach is what enables its real-time, low-latency performance.

The next leap in video generation won't come from monolithic models but from AI agents. These LLM-driven agents will use a suite of tools—including diffusion models, video editors like FFmpeg, and image editors—to iteratively create and refine complex, long-form videos.

The quality of generative visuals has leaped from blurry blobs to near-photorealistic films in a few years. Yet, the core technology—a diffusion process of adding and then removing noise—has remained consistent. Progress stems from optimizations and architectural improvements, not a complete paradigm shift.

The primary challenge in creating stable, real-time autoregressive video is error accumulation. Like early LLMs getting stuck in loops, video models degrade frame-by-frame until the output is useless. Overcoming this compounding error, not just processing speed, is the core research breakthrough required for long-form generation.