One of the most powerful ways to make open-weight models safer is simply to remove dangerous information (e.g., anthrax papers) from their pre-training data. This is not yet common practice because developers are extremely reluctant to modify their expensive and proven pre-training recipes.

Leading US models have safety features that block analysis of hacking tools and logs. This forces cybersecurity teams, like Hugging Face after a breach, to use less-restricted Chinese open-source models for essential forensic analysis, creating a security paradox.

This advanced safety method moves beyond black-box filtering by analyzing a model's internal activations at runtime. It identifies which sub-components are associated with undesirable outputs, allowing for intervention or modification of the model's behavior *during* the generation process, rather than just after the fact.

A novel safety technique, 'machine unlearning,' goes beyond simple refusal prompts by training a model to actively 'forget' or suppress knowledge on illicit topics. When encountering these topics, the model's internal representations are fuzzed, effectively making it 'stupid' on command for specific domains.

Research on bio-foundation models like EVO2 and ESM3 shows that strategically excluding key datasets (e.g., sequences of viruses that infect humans) dramatically reduces a model's performance on dangerous tasks, often to random chance, without harming its useful scientific capabilities.

Chinese labs use 'smart distillation,' a sophisticated technique where a frontier model acts as a 'teacher' to guide a smaller model's judgment and data labeling. This is viewed as a legitimate and efficient catch-up method, distinct from simply copy-pasting answers.

A novel AI safety technique called gradient routing trains mixture-of-experts models to isolate dangerous knowledge (e.g., bioweapons, cyber exploits) into specific "expert" modules during pre-training. These dangerous experts can then be completely removed ("ablated") before deployment, creating an inherently safer model.

Techniques created to make AI safer and more aligned with human intent, such as Reinforcement Learning from Human Feedback (RLHF), have turned out to be the very methods that significantly enhance model performance and usability. Safety work is capability work.

Instead of treating a complex AI system like an LLM as a single black box, build it in a componentized way by separating functions like retrieval, analysis, and output. This allows for isolated testing of each part, limiting the surface area for bias and simplifying debugging.