The robot didn't get to train *at all* with tied fingers β€” it had to adapt on the fly. (Also, humans have a billion plus years of evolutionary practice to solve the cube with untied fingers; the robot only gets about 10,000 years of untied practice.)
π’Άπ“π“Œπ’Άπ“Žπ“ˆ πŸ„΅πŸ…πŸ„΄πŸ…‚πŸ„· @rhyolight
15 Oct 2019
I could learn it in a year with all my fingers. Then it would probably only take a couple days to adjust my mental model to solve it with two fingers tied. How many billion years of training did the Deep RL agent need again?

Oct 15, 2019 Β· 5:56 PM UTC

7
14
6
116
Replying to @gdb
Is it not, by that reasoning, more proper to say the robot had itself 1 billion + 10 thousand years of practice? It required that an entity with 1 billion years of practice and the ability to mass manufacture pencils and toasters, specify it in tens of millions of lines of code.
Replying to @gdb
How common is this prior assumption of β€œhumans have a billion years of evolutionary practice”. Rarely people question this.
Replying to @gdb
What algorithm are you using to train it? Q-learning, LSTD, policy gradient? Model is neural network? Do you use feature engineering?
Replying to @gdb
Do you think that some training in the real world will be important for complex tasks, or do you anticipate that these results will be able to scale up entirely in sim?
1
Replying to @gdb
Well, I can't solve the cube even with all my fingers so the robot beats this human for sure.
1
Replying to @gdb
Pretty sure there weren’t any animals with fingers β€” or any animals at all β€” a billion years ago :P
1
9
Replying to @gdb
I'm only being contrary because the the intelligence I am looking for is the one that learns how neurons learn in the brain, not how life learns to evolve. An empty brain takes very little time to learn something new. It evolved to be a container of knowledge.
2
1
8