www.lesswrong.com/posts/Lyjq6jEfJPyBgLCDy/ai-safety-guide-for-true-beginners-by-...

The sentence claims that all AI models share the Transformer architecture. That is incorrect.

Authoritative overviews of AI and machine learning describe AI as including many different model families, not just transformers. IBM’s AI overview says machine learning includes techniques such as linear regression, logistic regression, decision trees, random forests, support vector machines, k-nearest neighbor, and clustering. The same page separately explains that in generative AI, the recent rise of complex systems involved three major deep-learning model types: variational autoencoders, diffusion models, and transformers.

IBM’s decision-tree explainer also describes a decision tree as its own supervised learning algorithm with a hierarchical, tree structure consisting of a root node, branches, internal nodes, and leaf nodes. That directly contradicts the idea that all AI models use Transformer architecture.

A more accurate version would be something like: many modern large language models use the Transformer architecture, but AI as a whole includes many non-Transformer model types.

This claim presents a "safety layer" as a universal final stage inside the model itself. That is not generally accurate.

Official product documentation from major AI providers shows that safety controls are often implemented as separate systems, not as a mandatory final layer in the mathematical architecture of every AI model.

• Google Cloud says Vertex AI content filters analyze generated text and block harmful responses, and explicitly states that these filters "function independently from Gemini models" as part of a layered defense.
• OpenAI’s API documentation exposes Moderations as a separate API capability: given text and/or image inputs, it classifies whether they are potentially harmful, using dedicated moderation models such as omni-moderation-latest and text-moderation-latest.

Those sources contradict the idea that AI outputs universally pass through a built-in final "safety layer" within the model. In practice, safety can come from several places: model training, system instructions, external classifiers, application-level filters, or human review.

A more accurate statement would be: many deployed AI systems use additional safety mechanisms, but these are often separate guardrails or moderation components rather than a universal final layer inside the model architecture.