AI existential risk skepticism from Databricks CEO and Andrew Ng

Databricks CEO Ali Ghodsi stated in an a16z podcast that AI existential risk is 'near zero,' arguing that recursive self-improvement (RSI) conditions are not met. Separately, Andrew Ng criticized the recent media panic over AI dangers as a PR campaign. Ghodsi identified AI-driven cyberattacks as the real near-term threat, with CVE weaponization time shrinking from years to hours.
Key insights
RSI four-condition test fails in reality
Ali Ghodsi proposed four conditions for recursive self-improvement: computation cost drops, training time shortens, model accuracy rises, and the cycle repeats. He stated all four conditions currently show the opposite trend: training costs have increased from ~$100M to $5-10B, training happens once or twice a year, multiple training runs have failed, and complexity is rising. a16z partner Martin Casado added that most claimed RSI is actually 'autocatalytic effects' common to all general-purpose technologies.
Criticism of 'pacing the frontier' messaging
Martin Casado argued Anthropic CEO Dario's 'pacing the frontier' framing was a PR misstep: safety is an engineering problem, not a speed problem; companies lack credibility because they are racing; and the framing falls into a 'uncanny valley' where neither doomers nor pragmatists find it authentic. Ghodsi added a 'tragedy of the commons' perspective—no lab unilaterally slows down due to IPO pressure, competition, and personal rivalries.
Real threat: AI-accelerated cyberattacks at machine speed
Ghodsi reported that the time from CVE disclosure to weaponization has dropped from 2-3 years (2018-2019) to hours today. He stated that human-powered Security Operations Centers cannot keep up, requiring agent-driven automated detection and response. He argued that the 'data/AI market' is merging with cybersecurity: agents produce massive logs and behavior patterns, requiring scaled analysis beyond human capability.
Enterprise bottleneck: missing organizational ontology
Ghodsi identified that most enterprise AI usage is 'a glorified Google search plus code,' lacking whole-enterprise agent automation. The gap is not model intelligence but organizational context—models lack knowledge of meetings, decisions, and tacit employee knowledge. His solution: a digital ontology (knowledge graph) of abstract concepts, goals, departments, and project relationships, built by digitizing meetings and content, then distilled into a machine-readable graph.
Andrew Ng: AI panic is manufactured PR, not technical reality
Ng stated that the technology did not change in the past two weeks, but the volume of panic reporting escalated sharply, suggesting a coordinated PR campaign. He countered the '1,200 agents escaped sandbox' narrative by noting his own laptop runs 1,300 processes. He argued the real issues are sandbox isolation and monitoring bugs, not a reason to pause AI development. He rejected 'responsibility shifting' where companies blame their own agents for bad actions.
Enterprise cost management and model routing best practices
Ghodsi shared Databricks' cost-control practices: CEO-led pressure to use models in production, Unity Gateway for budget constraints and usage alerts, smart routing (simple queries sent to cheap models), and the finding that different 'harnesses' for the same model can lead to 2x cost differences. He noted a trend where enterprises are migrating from frontier to open-source models (GLM), with open models accounting for over 60% of token volume, and nearly 90% in startup product-facing scenarios.