Demis Hassabis notes that while generative AI can create visually realistic worlds, their underlying physics are mere approximations. They look correct casually but fail rigorous tests. This gap between plausible and accurate physics is a key challenge that must be solved before these models can be reliably used for robotics training.

Unlike LLMs, parameter count is a misleading metric for AI models in structural biology. These models have fewer than a billion parameters but are more computationally expensive to run due to cubic operations that model pairwise interactions, making inference cost the key bottleneck.

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.

While game engines can handle messy mesh topology, AI-generated models with poor structure (triangles and n-gons) are unusable for artists in tools like Blender or Maya. This necessitates a time-consuming retopology pass, adding significant hidden labor costs to the production pipeline.

Great games are defined by their concept and gameplay, not just visual fidelity. Many successful games use primitive graphics, while visually stunning games often fail if mechanics are poor. This justifies focusing on a strong underlying world model that enables robust interaction.

Despite mandated adoption and new capabilities, there's no clear evidence yet that AI prototyping tools lead to faster production or better software. The time spent building a highly-detailed interactive prototype may not be quicker than traditional methods, and the complexity requires rigorous code review.

Current multimodal models shoehorn visual data into a 1D text-based sequence. True spatial intelligence is different. It requires a native 3D/4D representation to understand a world governed by physics, not just human-generated language. This is a foundational architectural shift, not an extension of LLMs.

Testing reveals that the fastest AI tool for text-to-3D generation is the slowest for image-to-3D, and vice versa. This performance inversion means that benchmarks for one input mode are irrelevant and misleading for evaluating the other, as they are effectively different systems.

Modern AI models are moving towards extremely low-precision arithmetic (e.g., 4-bit numbers) because it's more efficient. The trade-off is analogous to image processing: you get a better result with more pixels (more computations) and fewer colors (less precision) than the other way around.