Calls to slow frontier AI often blur two very different ideas: slowing the development of the most capable new systems and stopping AI progress altogether. The first may be possible for a time. The second almost certainly is not.
The signal
Frontier training remains unusually concentrated. Advanced chips, high-bandwidth memory, networking, data centers, electricity and expert teams create bottlenecks that make the largest training runs difficult to reproduce quietly. Researchers I interviewed for IBM Think described those constraints as real leverage over the pace of frontier development.
But trained models can diffuse. Once capable weights or techniques spread, developers can fine-tune systems, give them more inference-time compute, connect them to tools and build new applications without repeating the original training run.
The risk
A slowdown that applies only to a few labs could redistribute activity rather than halt it. Open models, smaller companies and researchers in other countries would continue working. AI itself may also lower the amount of expertise or labor required for some kinds of experimentation.
What researchers say
The strongest argument for slowing the frontier is therefore about time. A pause or deceleration could create space to improve evaluations, incident reporting, model monitoring, access controls and international coordination. Whether that time matters depends on what institutions do with it.
The hardest governance question is not simply whether governments or companies can press a brake. It is whether the brake applies to the capabilities that matter, lasts long enough to change safety practices and can be verified across organizations with different incentives.
What I’m watching
- Whether major labs converge on common thresholds for slowing or delaying deployment.
- How governments verify large training runs and advanced-chip use.
- Whether independent model evaluations become routine before release.
- Whether safety measures spread beyond the handful of companies training the largest models.
AI Safety Watch 