Published On Dec 3, 2023
Why are AIs very good at some tasks and quite poor at others? AIs seem particularly good at reasoning and abstract thought, which we might otherwise have said was the pinnacle of biological evolution. The insight is the realization that the human perspective is skewed. Humans are not actually very good at reasoning, though we think we are. This is called Moravec's Paradox.
Conscious processing and reasoning are very recent evolutionary developments, so there are still many rough edges. Multitasking, remembering important details, consciously controlling memory durability, etc are all difficult. Our unconscious mind (System 1) is much more advanced, and that is why unconscious tasks like motor control and visual processing seem so easy to us in comparison.
AI researchers often state a goal of attaining AGI, or human-level AI. However, this is an asymmetric goal. Human level reasoning is probably not very difficult for AIs. But the unconscious processing like motor control and visual cortex are challenging to replicate, because they're actually more sophisticated in terms of complexity. They took longer to evolve. Hence, it might make sense to pursue goals other than AGI in AI research.
Human- versus Artificial Intelligence
https://www.frontiersin.org/articles/...
Analyzing AGI: What Would the World of Artificial General Intelligence Look Like?
https://www.aei.org/articles/analyzin...
Moravec's paradox
https://en.wikipedia.org/wiki/Moravec...
Paper offers perspective on future of brain-inspired AI as Python code library passes major milestone
https://techxplore.com/news/2023-11-p...
Training Spiking Neural Networks Using Lessons From Deep Learning
https://ieeexplore.ieee.org/document/...
Whispers of A.I.’s Modular Future
https://www.newyorker.com/tech/annals...
ais-modular-future
#agi #intelligence #robotics #ai
0:00 Intro
0:17 Contents
0:24 Part 1: The paradox of abstract thought
0:42 Humans can use tools and language
1:01 Exponential growth from big heads
1:19 Is abstract thought the pinnacle of intelligence?
1:35 AI tackles abstract thought well
2:03 Less intellectual jobs are harder
2:30 Moravec's Paradox
3:00 Hard is easy, easy is hard
3:34 Part 2: Examining human cognition
3:54 Biases inherent in brain's structure
4:18 Thesis: Human reasoning is not mature
4:44 Humans cannot do multitasking
5:21 Humans cannot consciously control memory lifetime
6:24 Unconscious information processing
6:35 System 1 and System 2
6:46 Mastery moves a task from System 2 to System 1
7:14 Example: System 1 is very fast
7:58 Rewording Moravec's Paradox
8:22 Part 3: Redefining general AI
8:59 The human lens
9:35 Definition of AGI doesn't make sense
10:18 Task difficulty versus task complexity
10:40 Alternative non-AGI goals
10:55 Example: my thumbnail generation process
11:54 Increasing automation through narrow AI
12:23 AGI-level AI is probably smarter than us
12:44 Copying biological neurons more closely
13:14 Running evolution at high speed
13:47 Conclusion
14:48 Outro: taking the AI's perspective