Using a large language model to police another is computationally expensive, sometimes doubling inference costs and latency. Ali Khatri of Rinks calls this like "paying someone $1,000 to guard a $100 bill." This poor economic model, especially for video and audio, leads many companies to forgo robust safety measures, leaving them vulnerable.
By monitoring a model's internal activations during inference, safety checks can be performed with minimal overhead. Rinks claims to have reduced the compute for protecting an 8B parameter model from a 160B parameter guard model operation down to just 20M parameters—a "rounding error" that makes robust safety on edge devices finally feasible.
Instead of maintaining an exhaustive blocklist of harmful inputs, monitoring a model's internal state identifies when specific neural pathways associated with "toxicity" are activated. This proactively detects harmful generation intent, even from novel or benign-looking prompts, solving the cat-and-mouse game of prompt filtering.
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.
While a general-purpose model like Llama can serve many businesses, their safety policies are unique. A company might want to block mentions of competitors or enforce industry-specific compliance—use cases model creators cannot pre-program. This highlights the need for a customizable safety layer separate from the base model.
Universal safety filters for "bad content" are insufficient. True AI safety requires defining permissible and non-permissible behaviors specific to the application's unique context, such as a banking use case versus a customer service setting. This moves beyond generic harm categories to business-specific rules.
Current AI safety solutions primarily act as external filters, analyzing prompts and responses. This "black box" approach is ineffective against jailbreaks and adversarial attacks that manipulate the model's internal workings to generate malicious output from seemingly benign inputs, much like a building's gate security can't stop a resident from causing harm inside.
A comprehensive AI safety strategy mirrors modern cybersecurity, requiring multiple layers of protection. This includes external guardrails, static checks, and internal model instrumentation, which can be combined with system-level data (e.g., a user's refund history) to create complex, robust security rules.