[Digital Today reporter Jinju Hong] Demis Hassabis (데미스 허사비스), CEO of Google DeepMind, said artificial general intelligence (AGI) could emerge within a few years. He proposed establishing an independent standards body in the United States to evaluate advanced AI models before they are released. He argued that because AGI is expected to have an impact comparable to the discovery of electricity or fire, there is an urgent need to build safety systems that match the pace of technological progress.

On July 14 local time, blockchain media outlet Decrypt reported that Hassabis wrote on X, formerly Twitter, that "AGI is within 10 years, probably just a few years away." He defined AGI as AI that can understand, learn and carry out a wide range of tasks at a level equal to or beyond humans.

Hassabis said AGI could have a bigger ripple effect than the internet or the mobile revolution. "AGI is much closer to a technology like the discovery of electricity or fire," he said, adding it has the potential to fundamentally change science, industry and society as a whole.

He stressed that safety issues are as important as technological progress. Even with today's advanced AI models, cyber security threats are becoming a reality, he said, warning risks could expand into biology, nuclear technology and national security.

He argued that as AI evolves to make judgments on its own and improve its performance, safeguards are essential to ensure humans retain control. "To control increasingly autonomous, recursively self-improving systems, you need strong safeguards," Hassabis said.

To that end, he proposed establishing a "Frontier AI Standards Body" in the United States. He envisioned a public-private cooperation model similar to FINRA, a private self-regulatory organisation that oversees the U.S. securities industry, with industry and experts participating together under government oversight.

The body would evaluate the performance and risks of advanced AI models before they are made public. He suggested AI companies fund its operations and that independent technical experts and people from the open-source community take part in the evaluation process.

Hassabis said safety verification should eventually become mandatory before the release of the most powerful AI systems. "AI's rapid progress requires a testing regime that is dynamic and adaptable, yet strict," he said, arguing the United States can be the first to establish such international standards based on its technological and economic strength.

The proposal also aligns with a series of forecasts in the AI industry that AGI could arrive sooner than expected. Dario Amodei, CEO of Anthropic, said earlier this year that human-level AI could emerge within 1 to 5 years, and Hassabis also mentioned in June the possibility that AGI could become a reality before 2030.

In the United States, discussions on AI safety regulation are continuing. Sam Altman, CEO of OpenAI, proposed a licensing system for high-performance AI models and an independent safety audit scheme at a U.S. Senate hearing in 2023. U.S. President Donald Trump also signed an executive order last month to create a voluntary safety framework for reviewing advanced AI models before release. In the same month, Amodei also argued that AI safety regulation at the level of the Federal Aviation Administration is needed.

Hassabis again stressed there is not much time for the international community to establish common safety standards before AGI becomes a reality. "The future is not yet written," he said, adding, "In the time remaining before AGI arrives, we must decide the direction of the technology for all of humanity." He added, "If appropriate safeguards are put in place, AGI can serve as a catalyst for a new golden age of science and human progress."

https://t.co/PTeDiv1b6L