The History of AI in 7 Experiments

If you only have a few minutes to spare, here’s what investors, operators, and founders should know about AI’s history.

• Architecting logic. In the summer of 1956, Herbert Simon and Allen Newell showed off their remarkable program “Logic Theorist” to a collection of enlightened peers, only to be met with indifference. In the years since, Logic Theorist – which was capable of proving complex mathematical theorems – has been recognized as the first functional AI program. Its use of structured, deductive logic was an example of “symbolic AI,” an approach that dominated in the following decades.
• The rules of the world. AI’s most glorious failure may be a project named “Cyc.” In 1984, Douglas Lenat began his attempt to create a program with an understanding of our world. The then-Stanford professor sought to develop this context by inputting millions of rules and assertions the AI could use to reason – including basics like “all plants are trees.” Lenat believed insufficient knowledge represented a huge barrier for AIs that was best solved with thoughtful, manual intervention. Despite decades of development and hundreds of millions in investment, Cyc has struggled to deliver meaningful results.
• Embodied AI. In the 1980s, a new school of practitioners emerged to challenge AI’s dogma. This group, led by Australian academic Rodney Brooks, argued that real intelligence wouldn’t come from assiduously designed logical frameworks but by allowing machines to take in sensory input and learn from their environment. This “embodied” approach to AI ushered in practical robots, albeit with narrow applications.
• Emulating the brain. Geoff Hinton is regularly cited as the “godfather” of modern AI. The University of Toronto professor earned that honorific through his steadfast belief that powerful intelligence would be achieved by modeling the patterns of the brain. Hinton’s contributions to “neural networks” – a structure directly based on our gray matter – paved the way for modern AI systems to flourish.
• Learning by doing. How do you create an AI capable of beating the world’s greatest Go player? In the mid-2010s, Cambridge startup DeepMind showcased the potential of a radically new learning technique called “reinforcement learning.” Rather than learning how to play chess or Go through a set of strict rules, DeepMind’s engines developed by playing the game and receiving positive or negative feedback on their actions. This methodology has driven advancements far beyond the Go board.

 
Συνέχεια εδώ