Hassabis Says AGI Is Only Years Away – Proposes 30‑Day Pre‑Release Tests and Ongoing Audits

Frontier AI Framework and the Dawn of a New Era

Demis Hassabis, co‑founder and CEO of Google DeepMind and a 2024 Nobel Chemistry laureate, argues that artificial general intelligence (AGI) could appear in just a few years, marking a pivotal moment comparable to the discovery of fire or electricity. He describes this as "making sand think" and predicts an impact ten times that of the Industrial Revolution, accelerating at the same speed.

He believes AGI will enable breakthroughs in drug discovery, clean energy, and advanced materials, potentially ushering in an era of unprecedented abundance.

Frontier Challenges

While AI already delivers real‑world benefits, the approach of AGI brings new, severe risks. Advanced models threaten cybersecurity and could soon introduce biological, nuclear, and agent‑deception hazards. The rapid pace of capability growth outstrips our understanding, creating a high‑stakes geopolitical and commercial race.

Hassabis stresses the need for cautious optimism and calls for public policy that promotes innovation while enforcing responsible, safe behavior, fostering international cooperation, and ensuring AI deployment benefits society.

Establishing a Frontier AI Standards Body Framework

To keep pace with fast‑moving models, Hassabis proposes a new U.S.‑led standards institution modeled after the Financial Industry Regulatory Authority (FINRA). The body would be a public‑private partnership overseen by federal authorities, with a board of top technical experts and open‑source community representatives.

Funding would largely come from industry to attract world‑class talent and provide the compute resources needed for large‑scale testing.

The agency would define evaluation protocols, work with federal agencies and national labs on security‑sensitive tests, and certify models that meet a dynamic set of benchmarks as "frontier‑level".

Organizations whose models achieve the thresholds would be designated "frontier labs" and encouraged to adopt best practices such as publishing detailed model cards, maintaining strong internal cybersecurity, vetting key personnel, and allocating resources for safety research.

Initially, labs could voluntarily submit models for a maximum of 30 days of pre‑release review. After the process proves robust, the review would become mandatory for any model entering the U.S. market.

Post‑deployment, severe vulnerabilities would trigger mandatory collaboration between the lab and the standards body to remediate issues.

Evaluations would include rigorous scientific capability tests covering cybersecurity, bio‑risk, nuclear risk, and agent‑type AI deception checks. Additional safeguards such as digital watermarks on AI‑generated images and human‑readable output tokens would be required.

Benchmarks would be updated quarterly; outdated tests would be retired to prevent models from over‑fitting to stale criteria. The standards body would eventually develop its own testing infrastructure independent of labs to ensure unbiased assessment.

A third‑party audit ecosystem, potentially coordinated with the U.S. government, would help enforce evaluations and create new benchmarks.

The framework aims to balance technical rigor with innovation, scaling its oversight as new high‑risk threats are identified. If necessary, regulatory intensity could increase, including coordinated slow‑downs of development across frontier labs.

Recognition as a frontier lab would confer prestige, and the regime would apply equally to models from any country, whether open‑source or closed‑source. Smaller startups or academic projects not meeting frontier criteria would be exempt.

U.S. leadership in this effort could seed an international consensus on frontier AI standards, managing the most severe risks while ensuring global access to AI benefits.

Future Not Yet Written

Hassabis envisions AGI as a transformative tool for scientific and medical breakthroughs, driving massive productivity gains and economic growth. Realizing this potential requires a coordinated global framework, the strictest scientific methods, and the concentration of top talent to solve immediate technical challenges.

Beyond technical hurdles, he acknowledges deeper economic and philosophical questions about post‑scarcity societies, new economic models, values, and the meaning of life, urging broad societal participation in shaping the AGI era.

He concludes that the window before AGI arrives is precious; establishing dynamic testing, pre‑release review, third‑party audits, and international collaboration now will determine whether the technology advances for the benefit of all humanity.