The Abstraction and Reasoning Corpus (ARC)

https://www.youtube.com/watch?v=jSAT_RuJ_C

1. you need reasoning when there is an explicit knowledge gap
2. reasoning is knowledge acquisition
3. if you can broadly expand your horizons from a few examples, normalised by the amount of priors and experiences you have, then its intelligence
4. human intelligence is non computable?
5. the basic formalism is non computable
6. generalisation efficiency or knowledge acquisition efficiency
7. statistical or perceptual shortcuts
8. sparse problem
9. bongard problems
10. core knowledge and combinations of the same
    1. objectness priors
    2. goal-directness priors
11. domain specific language
    1. unary functions on grid
12. patterns
    1. symmetry
    2. movement
    3. denoising
13. hamming distance
14. dreamcoder by Kevin Ellis
15. functional type system
    1. valid syntax and valid typed syntax
    2. cut down the space of possible programs to be generated
16. in natural world simple things seem to generalise and work much better than complex things
17. repository of crystallized intelligence
18. fluid intelligence
19. type 1 and type 2 problems
    1. system 1 and system 2
20. brittleness of LLMs
21. representation space
22. will acquisition efficiency
23. reasoning is a search problem
24. perceptual problems with infinite possibilities
25. building a library of core knowledge and then classify the problem of objective and then do a search for the same
26. skill templates and generalising well
27. llm wide range shallow generalisation
28. symbolic representation narrow but deep
29. generate data to run more fundamental experiments
30. what is active inference?
31. universal Transformers
32. algorithmic learning
33. self attention
34. cross attention
35. discover the weakness of the model and then see how you can mitigate it using synthetic data?
36. inductive priors
37. discrete program search
38. flexible search
39. measure of intelligence
40. cognitive resolution
41. you need priors to learn
42. generalisation must be normalised by price and experience
43. cognition
44. the solutions in kaggle didn't generalise outside the instance of the problem
45. ingenious creativity in the moment

References

1. https://arxiv.org/abs/2402.03507#:~:text=While%20specific%20neural%20networks%20are,of%20abstract%20visual%20reasoning%20tasks