The commercial logic of frontier artificial intelligence is built around momentum. More computing power, more capable models, faster research cycles and larger capital commitments reinforce one another. Father Bohdan, an Orthodox priest and technology entrepreneur, has now challenged the assumption underneath that cycle: that the next increase in capability must arrive as soon as it can be built.
Writing on September 8, Father Bohdan did not call for a ban on artificial intelligence. He explicitly separated his position from shutting down systems that already work or rejecting technology. His proposal is to pause or slow the training of new, increasingly powerful models. The practical question he poses is simple: would people’s lives actually become worse if the frontier stopped advancing for a while?
For businesses already using AI, the answer could be uncomfortable for the companies racing at the frontier. Existing models are already embedded in software development, customer service, research, marketing, analysis and office workflows. A temporary halt to the next capability jump would not remove those tools. It would principally affect the laboratories, chip suppliers, cloud providers and investors whose strategies depend on continued expansion of the frontier.
The argument has become harder to dismiss as an outsider’s concern because leading researchers are raising related questions. OpenAI Chief Scientist Jakub Pachocki wrote on September 6 that no laboratory has solved alignment and monitoring well enough to continue responsibly scaling at maximum speed for much longer. He said he expects voluntary slowdowns to become common until shared safety thresholds are established.
At the same time, the technology is beginning to accelerate its own development. Fortune reported on September 8 that OpenAI’s research organisation was using 3.1 agent-workdays of AI effort for every human workday by mid-August. That is commercially significant because it means the marginal value of a stronger model may include the ability to help design the next stronger model. The race is therefore not only for customers; it is increasingly for the speed of future research itself.
This creates a classic coordination problem. A laboratory that slows on safety grounds may reduce its own risk while handing an advantage to a rival that continues. If every firm reasons that way, all may keep accelerating even if their researchers believe the collective pace is unsafe. In July, more than 1,200 employees at major laboratories, including Anthropic, Google DeepMind, OpenAI and Meta, backed a statement asking the U.S. government to help build mechanisms that could slow advanced AI development if necessary.
The tension became more visible on September 9 when Anthropic researcher Jacob Coxon was reported to have resigned from the company and the wider AI industry. He argued that no individual company could manage the risk on its own and called for government regulation or a coordinated slowdown. His position is more severe than Father Bohdan’s short public appeal, but both point to the same structural weakness: restraint is difficult when safety decisions are made inside a competitive race.
OpenAI itself has shown that temporary holds are technically and organisationally possible. In August it described a two-week pause in reinforcement-learning training on its latest deployment models while security and monitoring were strengthened. Its largest planned frontier reinforcement-learning run remained on hold during further evaluations. The company did not stop serving existing users; it separated deployment from the pace of further scaling.
For investors, that distinction matters. A regime of safety-triggered pauses would not necessarily destroy the AI market. It could shift value from raw scaling towards evaluation, cybersecurity, monitoring, model efficiency and applications built on systems that already exist. It could also make capital expenditure less predictable, particularly for infrastructure businesses priced on assumptions of uninterrupted frontier growth.
The deeper business question is whether the industry can create credible rules before a crisis creates them from outside. Voluntary restraint may work briefly, but the incentive to defect grows as models become more economically valuable. Shared thresholds, independent audits and common reporting requirements are attempts to turn safety from a competitive disadvantage into a market-wide constraint.
Father Bohdan’s challenge therefore reaches beyond ethics. If today’s AI remains useful during a pause, then the economic case for immediate escalation must be weighed against the option value of waiting. The next phase of the AI market may depend less on who can train the strongest model first than on whether companies can prove that they know when not to train it yet.