How Long Before Superintelligence?
9 months ago
- #Superintelligence
- #Moore's Law
- #Artificial Intelligence
- Nick Bostrom discusses the potential for superhuman artificial intelligence (AI) within the first third of the 21st century.
- Superintelligence is defined as an intellect much smarter than the best human brains in practically every field, including scientific creativity, general wisdom, and social skills.
- Moore's Law is highlighted as a key factor in the exponential growth of computational power, with processor speed doubling every 12 to 18 months.
- Estimates of the human brain's computational capacity range from 10^14 to 10^17 operations per second (ops), with Moravec's estimate at 10^14 ops for the whole brain.
- The paper explores the bottom-up approach to AI development, mimicking the human brain's learning mechanisms and initial neuronal structure.
- Neuroscience advancements are deemed crucial for understanding brain architecture and learning algorithms, with a wild guess of 15 years needed to gather sufficient knowledge.
- The stagnation of AI in the 70s and 80s is attributed to insufficient hardware power, a limitation no longer present with current technological advancements.
- Once human-level AI is achieved, a positive feedback loop is expected to rapidly lead to superintelligence, with economic and military incentives driving development.
- The paper concludes that superintelligence is likely to be developed in the first quarter of the 21st century, with hardware requirements possibly met by 2004-2008 for lower estimates and 2015-2024 for upper bounds.
- Postscripts update progress, noting IBM's Blue Gene/L supercomputer exceeding Moravec's estimate in 2005 and the launch of the Blue Brain project aiming to simulate the neocortical column.