Castillo-Madrid describes Grok safety approach and controls
Castillo-Madrid says safety is core to Grok from pre-training through deployment, describes control, prevention, mitigation, analysis, and correction, says models are trained to refuse dangerous activity, evaluated for misuse, monitored after deployment, and not released if insufficiently robust.
Just as we have done with Space Launch and Human Spaceflight, SpaceX SI puts safety first.
As a mother, I know the risks New Yorkers worry about are often those close to home, protecting our children, preventing harmful content, and stopping misuse of everyday SI tools.
And I can assure you that safety is core to how we build Grok, from pre-training through deployment.
Our approach starts with a simple idea.
A super intelligent system built to seek the truth and understand the universe as it actually is will be more safe, reliable, and useful to the people who rely on it.
Our safety work is how we make sure that holds up in practice by focusing on control, prevention, mitigation, analysis, and correction.
Concretely, that effort begins well before a model is released for public use.
We train our models to refuse assistance with dangerous activity, evaluate it for risks including misuse of our terms of service, attempts to get around
safeguards, and loss of control.
We monitor systems after deployment and we do not release models we believe are insufficiently robust against misuse by bad actors or follow meaningful human control.