In a conversation summarized by a16z, Databricks' Ali Ghodsi argues that rising demands for power, GPUs and engineers, alongside costly failed training runs, stand in the way of a runaway self-improving AI loop. Cybersecurity is his closer concern: he says most organizations aren't equipped for changing AI-agent capabilities, and that the time between a vulnerability's publication and its weaponization has shrunk from years to hours. On adoption, he believes models are already smart enough for most companies but lack the context employees absorb over years. He thinks a halt in frontier-model progress would not meaningfully change the value most companies currently extract from AI.