Automation and complexity barriers

Automation—of physical processes through machines and of information processing through software—has had a transformative impact on the world economy, especially since the Industrial Revolution and the computer revolution. The extreme limiting case of this trend towards increasing automation would be the development of artificial general intelligence, a development which would undoubtedly have profound societal consequences. It is therefore important to study what factors determine the pace and ultimate limits of automation.

While computational hardware is growing in capability at a readily quantifiable exponential rate, the rate of performance gain of software is much harder to assess. Some researchers (e.g., Hans Moravec) have argued that available computing is the main rate-limiting factor for advances in robotics and artificial intelligence. Others (e.g., Jaron Lanier) have argued that, on the contrary, software has run into a complexity barrier which will render rapid progress impossible despite abundant computing power.

One way to begin to resolve this dispute would be to conduct a series of case studies of the relation between hardware performance gains, various software design inputs (such as computing power, data, programmer hours, etc.) and software performance across several natural task domains (e.g., computer chess, autonomous cross-country driving, and speech recognition). This would also help address such questions as whether software progress tends to be incremental and predictable, due to the accumulation of many small improvements, or discontinuous and surprising due to the occurrence of rare breakthroughs. A better understanding of these issues would clarify the plausibility of scenarios that postulate major advances in various types of machine intelligence. It would also cast light on analogous issues that arise in the context of the future of automation of physical processes such as via molecular nanotechnology.