David Robinson, who left OpenAI, expressed his concern about the current culture of the artificial intelligence industry, as detailed in an editorial published in The Atlantic this week. His opinion is relevant because Robinson was responsible for drafting the safety reports attached to the major releases of AI models developed by OpenAI, the creator of ChatGPT.
For Robinson, the problem facing the sector is very serious, to the point where simple regulation—that is, creating standards for training AI models—is not a complete solution. He argued that artificial intelligence requires protections equivalent to those of nuclear facilities.
AI requires nuclear-level safeguards
Robinson stated that, considering the existing risks, cutting-edge laboratories must operate with the same caution as nuclear power plants or busy airports. This implies having multiple layers of redundancy and meticulous, time-consuming planning aimed at preventing a human error, although predictable, from resulting in a disaster.
He criticized the Silicon Valley environment, describing it as operating with 'extreme confidence' and 'perpetual sprints.' According to Robinson, this approach leads the industry to build increasingly larger and more advanced models under 'unrestricted optimism,' ignoring or minimizing potential problems.
Additionally, Robinson suggested that it is crucial for AI companies to develop a sense of modesty and broaden their vision beyond the restricted technological ecosystem dominated by the 'move fast and break things' mentality.
This warning from Robinson comes after a series of safety professionals and researchers leaving their positions at prominent AI companies. Jacob Coxon was one of the first to leave Anthropic. His colleague, Evan Hubinger, commented on X/Twitter that AI 'could kill us all by the end of the decade.'
Hubinger added his personal belief that there is more than a 10% chance of AI causing the death of everyone in the next ten years, acknowledging Anthropic's efforts but pointing out that there is still no clear plan to solve superintelligence alignment.
Following Coxon's departure, other warnings emerged from Robert O’Callahan, Bilal Chughtai, and Josh Engels, both from Google DeepMind, as well as Joe Benton, from Anthropic. The news that AI companies need nuclear plant rules, according to the former OpenAI employee, was initially published in Olhar Digital.