Let’s think for a moment about how the technology market was moving before November 2022. The artificial intelligence It was already present in search engines, cameras, recommendations and digital services, but it was not yet the label that brands tried to place on each product or the argument around which a good part of their presentations revolved. The industry had other visible priorities and distributed its attention among multiple fronts. Less than four years later, it’s hard to find a large manufacturer or platform that hasn’t reorganized part of its strategy around AI.
You only have to look at where the money is flowing to understand the magnitude of this race. Big technology is allocating enormous investments to chips, servers and data centers that spread across different parts of the world, while seeking to ensure the energy necessary to keep them running. Governments are not just watching, either: the United States and China support the development of infrastructure and computing capacity as part of economic and strategic competition. In the midst of this acceleration, one of its main protagonists has warned that we are advancing faster than we understand.
The warning from those who are on the front line of the race
The message does not come from outside the sector. Demis Hassabis is co-founder and CEO of Google DeepMind, one of the laboratories that promotes the development of the most advanced artificial intelligence systems. Under his direction, projects such as AlphaGo, which defeated a world Go champion, and AlphaFold, capable of predicting protein structures, emerged. In 2024, Hassabis and John Jumper jointly received half of the Nobel Prize of Chemistry for his work on protein structure prediction. His words have weight precisely because of this double condition: he helps build this technology and, at the same time, he calls for mechanisms to contain its risks.
Part of the text published by Demis Hassabis in X | Click to see the full message
Hassabis starts from a conviction that helps understand the urgency of his proposal: he believes that artificial general intelligence (AGI), defined in his text as a system capable of exhibiting all the cognitive capacities of the human brain, could arrive in just a few years. He does not present it as a certainty, but as a close possibility that would force us to prepare before technology reaches that point. Their concern covers cybersecurity, possible biological and nuclear risks and, later, increasingly autonomous systems, capable of acting with less supervision and improving their own capabilities.
In an extensive article published in X, the manager tries to support two ideas at the same time. Artificial intelligence can become an extraordinary tool for science, medicine and the economy, but that potential does not eliminate the need for establish controls and oversight mechanisms. Nor does he propose waiting for a specific threat to appear to react, because then the measures could arrive too late. Before detailing which organization and what evaluations it considers necessary, it sets out the diagnosis that serves as the basis for its entire proposal:
“We are currently locked in an extremely intense trade and geopolitical race unfolding at multiple levels. While these competitive dynamics are driving rapid advances and accelerating their extraordinary benefits, progress on the frontier of AI is outpacing our understanding of the technology. No one in the world knows for sure what will happen next, and not even the experts agree. When there is such a high degree of uncertainty and so much at stake, moving forward with cautious optimism is the sensible and correct strategy.”
The answer he proposes is to create in the United States a specialized organization in evaluating the most advanced artificial intelligence models. Their proposal takes as reference a public-private partnership or a self-regulatory entity with federal supervision, led by a board that would also include independent specialists and representatives of the open source ecosystem. This institution would define what thresholds make a system a frontier model and design assessments on cybersecurity, biological threats and other high-risk areas, as well as tests to detect attempts to circumvent safeguards or signs of deception. In a first stage, laboratories would voluntarily share their models up to 30 days before launching them.
Initial cooperation could later become a mandatory requirement. Once the protocol is validated, any model considered frontier would have to pass the evaluation before reaching the United States market. Tests would be periodically reviewed to replace outdated ones and measure new capabilities, while independent auditors would help expand the control system. The approach goes even further: if the severity of the risks justified it, the framework could be used to coordinate a slowdown in development between major laboratories.
The concern is not exclusive to Hassabis. Geoffrey Hinton has acknowledged that we do not know if we will be able to retain control of systems more intelligent than us, while Yoshua Bengio calls for more research and specific mechanisms to supervise them. In 2023, Elon Musk also signed an open letter calling for pause for at least six months the training of AI systems more powerful than GPT-4, although a few months later it announced xAI and began to compete directly in this market. Many of these voices agree in calling for caution, but there is no consensus on the probability that we will lose control.
The history of artificial intelligence still does not have a written outcome, although its first effects have already changed our relationship with technology and the decisions of those who develop it. It remains to be seen whether the body proposed by Hassabis would have a place, whether its evaluations would be truly effective, and whether laboratories and governments would agree to submit to them when they conflicted with their own interests. It may also happen that some of the risks it poses are oversized or do not materialize. For now, his proposal adds a concrete roadmap to a discussion in which there are still more unknowns than certainties.
Images | Google
In | Claude has a space in which he manipulates concepts before his final response. And it has surprised Anthropic itself
