Ali Ghodsi argues that discussing AI's existential risk, which he believes is near zero, is irresponsible leadership. It causes unnecessary public panic and mental health issues, and prompts misguided government regulation that could be counterproductive to actual safety goals.
The word "pacing" is a poor PR choice for AI safety. It's seen as a weak compromise that fails to appease doomsayers who want a "pause" and alienates security experts who want concrete actions, not vague slowdowns. The focus should have been on specific "safety and security" controls.
The debate on existential risk misses the present danger: AI-powered cyberattacks. AI agents can find and exploit vulnerabilities in hours, not years, a speed that human teams cannot handle. The entire security industry must rapidly shift to AI-driven, automated defense to keep up.
Databricks CEO Ali Ghodsi proposes a 4-part litmus test for dangerous recursive self-improvement (RSI). The risk is real only if new models simultaneously require less training time, fewer resources, and achieve higher intelligence, with this cycle being repeatable. Currently, the opposite is true.
The fear of runaway Recursive Self-Improvement (RSI) is tempered by economic reality. Training frontier models is becoming more, not less, expensive and complex. Each new model requires more GPUs, time, and brittle engineering, creating a logistical barrier that naturally slows progress.
Cybersecurity is no longer separate from data and AI. As companies deploy internal AI agents, these agents generate massive amounts of log data. Securing the enterprise now requires analyzing this data at scale, effectively collapsing the cyber and data/AI markets into a single discipline.
Most enterprises don't need smarter AI to see huge productivity gains. The real barrier is that models lack deep organizational context—unwritten rules, project histories, and key personnel knowledge. Successfully feeding this "ontology" into existing AI is the key to unlocking its value.
An "ontology" codifies the difference between a new hire and a five-year veteran: the implicit knowledge of relationships between people, projects, and processes. Structuring this knowledge into a graph allows an AI to navigate an organization and provide context-aware insights, moving beyond basic information retrieval.
Just as Google doesn't crawl the web for every search, enterprise AI shouldn't query individual systems live. To be fast and comprehensive, it needs an offline, pre-computed index—an "ontology"—of all company data, relationships, and permissions. This is the enterprise equivalent of Google's PageRank.
AI leaders don't slow down for safety because of a classic "tragedy of the commons." If one company pauses, competitors will race ahead to capture the market and IPO rewards. They publicly call for regulation as a way to force a collective, industry-wide slowdown they can't achieve on their own.
A sophisticated gateway that routes queries to different models based on complexity is key to managing AI costs. Simple tasks go to cheap, open-source models, while difficult ones use the frontier. This "expert pattern" allows token usage to rise while keeping costs flat.
The hype around Recursive Self-Improvement (RSI) is often misplaced. What most startups call RSI is simply an "autocatalytic" process—using AI as a tool to speed up development, like using a computer to design a better computer. This is a normal, non-existential technological advancement.