uri-eliabayev talks with eran-hoffman, Chief R&D Officer at Dream, an AI company providing cyber defense (and now intelligence and other products) to governments and critical infrastructure. Government customers demand top quality and explainability, and often air-gapped on-prem deployment where fixing a bug can mean flying in. So research POCs built on big models must be re-engineered for production. Example: a model that predicts an attacker’s next steps took three days instead of five hours once moved to large customer networks. They fixed it by splitting tasks across small fine-tuned models under an orchestrator with an expert-knowledge base (“the sous-chef”), then optimized inference from drivers to model configuration. Every component must pass an eval built on a 500–1,000-item expert ground truth expanded with Claude. Skills hit a ceiling, after which they fine-tune, sometimes automatically at the customer site. On threats: AI makes attacks a commodity and zero-days a matter of hours, so defense must be faster and increasingly autonomous (“don’t bring a knife to a gunfight”), with fine-grained containment recommendations rather than “disconnect everything”.
Show: hidden-layers
Concepts: ai-cybersecurity · multi-agent-orchestration · small-language-models · llm-evals · skill-engineering