The history of AI since the creation of the Perceptron in 1958 has been a story of extreme optimism and failure cycles. This has lead many to draw the conclusion that, algorithmically, creating human level intelligence is insanely difficult, if not out right impossible.
However, the impressive feats produced by neural net like approaches in the last decade is causing the emergence of a new alternative narrative about the quest for AGI. One where intelligence is fundamentally simple and the history of AI is really a story about limited computational resources and advanced techniques used to leverage those resources.
It’s as if we’ve been doing microbiology without a microscope for half a century and prematurely concluded that it’s intrinsically theoretically hard instead of being a matter of insufficient equipment. As founder of reinforcement learning Richard Sutton recently explained, The Bitter Lesson is that compute reigns supreme in AI.
In point, did we draw the wrong conclusion about the troubled history of AI research? For instance the human brain contains 100 trillion synapses and the AI models deployed to date have yet to exceed 10 billion parameters. In retrospect, It didn't matter how smart the AI researchers of the 1960s were, they had no hope of achieving human-level AI on 1960s hardware.
Maybe the truth is that its not algorithmically all that difficult to develop human-level AI when given sufficient compute to experiment with. And, as a prediction, we will achieve human-level AI shortly after we start utilizing said compute. And most importantly, this hypothesis will be testable within the next 2–10 years.
TLDR: AI may be easy, we just had shit compute that made us think it was a hard problem.
However, the impressive feats produced by neural net like approaches in the last decade is causing the emergence of a new alternative narrative about the quest for AGI. One where intelligence is fundamentally simple and the history of AI is really a story about limited computational resources and advanced techniques used to leverage those resources.
It’s as if we’ve been doing microbiology without a microscope for half a century and prematurely concluded that it’s intrinsically theoretically hard instead of being a matter of insufficient equipment. As founder of reinforcement learning Richard Sutton recently explained, The Bitter Lesson is that compute reigns supreme in AI.
In point, did we draw the wrong conclusion about the troubled history of AI research? For instance the human brain contains 100 trillion synapses and the AI models deployed to date have yet to exceed 10 billion parameters. In retrospect, It didn't matter how smart the AI researchers of the 1960s were, they had no hope of achieving human-level AI on 1960s hardware.
Maybe the truth is that its not algorithmically all that difficult to develop human-level AI when given sufficient compute to experiment with. And, as a prediction, we will achieve human-level AI shortly after we start utilizing said compute. And most importantly, this hypothesis will be testable within the next 2–10 years.
TLDR: AI may be easy, we just had shit compute that made us think it was a hard problem.
