Deepmind CEO Hassabis says "nobody in the world knows what happens next" so "cautious optimism" means building guardrails now

Hassabis opens his latest framework proposal by acknowledging: 'No one in the world knows exactly what will happen next, even experts disagree.' This judgment is not unfounded—last December, Meta's chief AI scientist Yann LeCun publicly dismissed the concept of general intelligence based on language models as 'complete nonsense' and 'thoroughly delusional,' to which Hassabis countered that LeCun 'is simply wrong.'

The clash reflects broader splits: Gemini co-lead Oriol Vinyals reconciles that current models are powerful in some areas but lack true innovation; deep reinforcement learning pioneer Richard Sutton founded Oak Labs to address the issue; DeepMind co-founder Shane Legg directly predicted 'minimal AGI' could be achieved by 2028. Hassabis himself claimed in April that AGI is 'only a few years away,' with impact ten times the industrial revolution and arriving ten times faster, and in May said humanity has 'stepped into the foothills of the singularity.'

This disagreement is not merely academic. Just before the proposal, a group of AI researchers and economists jointly warned that AI-driven job displacement could trigger massive ripple effects. Hassabis did not sign that letter, but his proposal offers a more concrete, non-alarmist approach to address these concerns.

Hassabis's core proposal is to establish a new standards agency modeled after the U.S. Financial Industry Regulatory Authority (FINRA). This agency would develop evaluation protocols for frontier models, initially voluntary and later mandatory, funded by the industry itself, using regularly updated benchmark tests. The international community would then need to follow and seek consensus on key points.

The proposal explicitly exempts non-frontier models from startups or academic research, thus avoiding accusations of 'regulatory capture'—i.e., established companies using rules to suppress competitors. If necessary, the agency could coordinate slowdowns in development, similar to measures previously considered by Anthropic.

Hassabis emphasizes that this effort essentially institutionalizes the principle of 'cautious optimism': when uncertainty is extremely high and risks are enormous, the wise strategy is to build guardrails proactively while advancing the frontier.

The timing of Hassabis's proposal is worth noting. The recent job-loss warning letter had just been publicized, and although he did not sign it, his discussion of potential consequences overlaps with its themes. The difference is that his proposal transforms concerns into a specific, actionable governance structure, rather than a mere warning. This may be a response to recent public pressure or a positioning for an upcoming policy window.

The proposal is essentially a two-tier structure: domestic first, international later. Choosing to model after FINRA rather than creating a completely new regulatory body lowers barriers to implementation. But the key challenge remains: when experts cannot even agree on the definition of AGI, its timeline, or even the basic path, how do you design 'effective' benchmarks? Hassabis remains silent on this, only emphasizing that we must start immediately.

On a deeper level, this proposal marks a shift from 'techno-optimism' to 'forced governance' among AI industry leaders. As he writes: 'When no one knows the future, and the stakes are so high, cautious optimism is the wise path.'