a16z Podcast: Ali Ghodsi Says AI Safety Needs Engineering, Not Fear

AI safety must transition from a marketing-driven panic toward a rigorous engineering discipline focused on concrete cybersecurity threats rather than existential dread. While a "tragedy of the commons" makes voluntary industry slowdowns unlikely, pacing is ultimately irrelevant to safety because dangerous systems can still be developed slowly. Real-world risks involve vulnerable infrastructure and automated attacks, which require rigorous monitoring and independent inspection rather than vague regulation. Current frontier models remain brittle and human-intensive, suggesting that the era of self-accelerating superintelligence has not yet arrived. For enterprises, the immediate value lies in organizational context and agent-native infrastructure—where over 90% of new databases are now agent-generated—rather than pursuing PhD-level reasoning. Success depends on balancing high-value frontier models with cost-effective open-source alternatives while implementing smarter harnesses for crisis support, drug discovery, and logistics.