Hard Takeoff

Hard Takeoff is a scenario from the future-oriented debate about thinking machines. The idea: a program becomes so good at improving itself that it repeats this process over and over again. Each improvement makes it more capable of finding the next improvement even faster. According to this idea, the leap from roughly human-level ability to vastly superior ability does not take decades, but months, days, or even hours. The counterpart is called Soft Takeoff: there, the same development unfolds slowly and is spread out over many years. Both are speculations about the future, not observed facts.

Why speed determines control

The debate is less about the final outcome than about the pace. With a slow development, society would have time to react. Policymakers could write laws, companies could build in safety tests, mistakes would surface early and could be corrected. In a Hard Takeoff, this reaction time would almost entirely disappear.

Another point is the distribution of power. If a single lab were the first to have a self-improving system, it could build up a lead in a short time that no one else could catch up to. Experts call this a decisive strategic advantage. In a slow development, by contrast, there would be many similarly powerful systems existing side by side, balancing each other out.

This is why concrete policy hinges on this question. Those who consider a Hard Takeoff likely call for safety rules already today, before the technology even exists. Those who consider it unlikely see such rules as premature and as an obstacle. Both sides rely on the same facts and arrive at opposite conclusions.

The imagined loop of self-improvement

The core of the argument is a feedback loop. A system that is good at programming and research could work on its own software. The improved system is even better at improving on the next round. Experts speak of recursive self-improvement, meaning a process that keeps applying to itself again and again.

For comparison: interest on a savings account grows similarly, because the interest itself earns further interest. Only that this happens over years. In a Hard Takeoff, the round would not last a year, but perhaps an hour. With sufficiently short rounds, the curve looks from the outside like a vertical wall.

There are serious objections to this picture. Software alone is not enough: a system needs computing chips, electricity, and data, and these cannot be procured within hours. Moreover, many problems become harder, not easier, with each round of improvement. Physical experiments take time, no matter how intelligent someone is. Some researchers therefore consider the leap impossible, others merely unlikely.

The term in debates and headlines

The idea became well known through books such as Nick Bostrom’s Superintelligence from 2014. Since then it has appeared regularly in interviews with AI executives and in hearings before parliaments. Often it appears in the text without a name, for instance as a warning about an intelligence explosion or about the point of no return.

For investors, the term is interesting for a sober reason. It explains why companies pour billions into safety research even though this does not generate money in the short term. It also explains why some labs hesitate when it comes to releasing their models.

A common misconception: Hard Takeoff does not mean that a machine suddenly gains consciousness or develops feelings. It is exclusively about capability and speed. A second misconception is to associate the term with today’s chatbots. No current system improves itself on its own; every new version is trained and released by humans.