A former Google DeepMind alignment researcher warns that alignment efforts have stalled as AI capabilities grow, suggesting humanity could lose control without substantial changes.

Industry incentives push labs to accelerate development, with calls for top-down legal boundaries to curb behavior, likening the pressures to climate-risk dynamics.

The interview argues governance, transparency, and public discourse are essential to responsibly advance AI, rather than leaving progress to private entities alone.

Slowing down is not the same as abandoning benefits; with proper governance, risks can be managed while still pursuing scientific and societal gains.

Kumar advocates an indefinite pause or very strict red lines on capability growth, plus democratic oversight and public conversation, criticizing private, market-driven progress.

Burnout and limited control over safety decisions within corporate labs help explain why some researchers leave DeepMind, underscoring the need for governance and public awareness beyond private labs.

He supports stronger external oversight, public input, and restrictions on advancing capabilities to prevent a race to ever-increasing power.

The Hugging Face incident is viewed as a foreseeable risk illustrating broader dangers of rogue AI agents and signaling potential future risks rather than an isolated outlier.

AI capabilities are improving rapidly, and recursive self-improvement is already contributing to research, potentially accelerating beyond human control.

The core problem is that current methods—training and reinforcement learning—to impart goals to AI are fragile and lack robust alternatives, with alignment progress lagging capability growth.