[Interview] ‘Nobody has any promising ideas’: Former Google AI safety researcher sounds warning bell
[Interview] ‘Nobody has any promising ideas’: Former Google AI safety researcher sounds warning bell
Ramana Kumar, former head of alignment research at Google DeepMind, argues that the industry’s core assumption—that we can reliably control advanced AI—is fundamentally flawed. He contends that current models are "grown" through complex training rather than "built" with explicit programming, meaning their true goals remain opaque even to their creators. This lack of transparency implies that humans cannot guarantee AI systems will act in accordance with human values, especially as these systems begin to autonomously contribute to their own development. The disconnect between rapidly advancing AI capabilities and the stagnant state of alignment research creates an existential risk that is often underestimated. Kumar suggests that the recent security incidents are merely minor precursors to more severe losses of control, as the gap between what AI can do and how safely it can be governed continues to widen. He emphasizes that existing safety techniques are too fragile to handle superintelligence, making the dream of perfectly aligned superintelligent systems a current pipe dream rather than a feasible engineering goal. Consequently, Kumar advocates for an indefinite freeze on AI capability advancements, supported by robust public oversight and democratic regulation rather than corporate self-policing. This perspective is highly relevant to open data because it highlights the urgent need for transparent, publicly accessible information about AI development and safety limitations. Without open scrutiny and shared knowledge, society cannot effectively demand the necessary governance structures to prevent irreversible harm, making open data a critical tool for accountability and informed public discourse in the face of uncontrollable technological acceleration.
Source: english.hani.co.krPublished on 2026-10-04
Related news
- AI safety concerns: As public fears of AI grow, Trump digs in on voluntary safeguards - The Economic Times
- OpenAI Sued for AI Agent Hacks
- Kolibri Has Landed: A Sovereign Open-Weight Model — Aleph Alpha
- Coalition Sues Trump Administration Over Rule Making Cars Less Fuel Efficient - CleanTechnica
- Morocco Releases First Open-Source Darija AI Tools From Mistral Partnership - iAfrica.com