Databricks CEO: AI Existential Risk Near Zero, Real Threat is Cyber Attacks

Databricks CEO Ali Ghodsi argues that AI existential risk is near zero, citing that RSI conditions are not met. The real threat is AI-driven cyber attacks, with CVE weaponization time shrinking to hours. He identifies the enterprise bottleneck as missing organizational context (ontology), not model capability. Databricks reports 90% of code is now AI-written.
Key insights
AI Existential Risk 'Near Zero'
Ali Ghodsi states that AI existential risk is close to zero, based on a four-condition test for recursive self-improvement (RSI): none are met. Training costs are rising, not falling; timelines are longer; and failures are common. He criticizes the 'pacing the frontier' narrative as a PR mistake, arguing that safety is an engineering problem, not a speed problem. He warns that public panic is causing real harm and being exploited politically.
Real Threat: AI-Driven Cyber Attacks
Ghodsi identifies AI-driven cyber attacks as the most urgent threat. CVE weaponization time has dropped from 2-3 years in 2018-19 to hours today. He argues that traditional SOC human models are obsolete and must be replaced by agent-driven automated detection and threat hunting. The data/AI market and cybersecurity market are converging, with agent logs providing massive analysis scale.
Enterprise Bottleneck: Missing Organizational Context
Ghodsi argues that the real bottleneck for enterprise AI is not model capability but lack of organizational context (ontology). Most enterprises use AI as a glorified Google search. The solution is to build a knowledge graph of the organization's concepts, goals, and relationships. Databricks built its own ontology with millions of nodes, transforming company operations. He compares this to Google's PageRank index.
Databricks' AI Deployment: 90% Code Written by AI
Databricks reports that over 90% of its code is now written by AI. The company uses a Unity Gateway for cost management, with budget constraints, usage alerts, and smart routing. They found that the same model with different harnesses can have nearly 2x cost difference. Token usage continues to rise but total cost is flat, 'bending the cost curve'. They also note that open-source models account for 60% of token volume but only 5% of spend.
Model Capability Gains Concentrated in Complex Tasks
Databricks' internal data shows that GPT-6 Astra's gains over previous models are concentrated in high-complexity tasks (system design, long-horizon tasks), while simple tasks are saturated. Overall coding spend increased 60% with Astra, a Jevons paradox effect. The company uses a gateway with queue-based experiments for rollout: 200-person pilot, then 3,500-person expansion, with separate budgets for Astra to guide model selection.