Pacing the Frontier

Companies building the currently most powerful AI systems could release every new result immediately. Some deliberately do not do so. They hold back new capabilities, give them only to small test groups, or release them in a weakened form. This stance is called “Pacing the Frontier” in English, meaning roughly: setting the pace of development at the very edge of what is possible yourself. The idea behind it is simple. If a technology grows faster than laws, safety checks, and the public can keep up with, risks arise that no one can contain anymore. Pacing is therefore not a technical procedure, but a deliberate decision about when something is released.

The dispute over the right pace

The term stems from a debate that has shaped the AI industry for several years now. On one side stand people who consider rapid release the right approach. Their argument: only when many people use a system can its weaknesses be found. Moreover, it is better for responsible companies to stay ahead rather than let others take the lead.

On the other side is concern about irreversible mistakes. Software that has once been released freely online cannot be taken back. If a model can, for instance, provide detailed assistance with cyberattacks or dangerous chemicals, a later correction is of little use. Pacing is meant to avoid exactly this point: better to wait a few months than to accept damage that remains.

For investors and observers, this is also an economic question. A company that slows down may be giving up market share to competitors that do not. Critics call this a race to the bottom: whoever is least cautious wins in the short term. That is why many call for the pace not to be left to individual companies, but to be set through rules that apply to everyone.

What slowing down looks like in practice

Pacing rarely means that nothing at all is released. Intermediate steps are more common. A new system first goes to a small group of testers who deliberately try to make it misbehave. This intentional attacking of one’s own product is called red teaming. Only when nothing serious turns up does the broad launch follow.

A second method is withholding individual capabilities. A language model, for example, may be offered publicly while the ability to act independently on the internet remains blocked. The weights, meaning the learned numerical values inside the model, are also often not released. This way, control stays with the provider, because no one can simply remove the safety restrictions.

Several large providers have written their own tiered plans for this. These specify which measured capabilities trigger which protective measures. If a model reaches a certain tier, it may only appear after additional checks. However, these plans are voluntary. A company can change them or interpret them differently if competitive pressure increases.

Pacing in headlines and products

Pacing can usually be recognized by phrasing in press releases. Sentences like “initially for selected users,” “gradual rollout,” or “for now only in the US” point to it. Delayed launch dates for new features also belong to this category. Sometimes the reason is technical overload, sometimes a deliberate safety decision.

In politics, the topic appears under other names. The European Union’s AI Act requires additional checks before market entry for especially powerful models. This is pacing by law instead of by voluntary commitment. How strongly such rules actually work is an open question and is being discussed continuously.

A common misconception is to confuse pacing with standstill. It does not mean halting research. It means keeping the gap between capability and control small. Looking closely, one can also see a business model in this: providers who can credibly demonstrate caution are more attractive to authorities, banks, and hospitals than those who simply release everything immediately.