o3 (Part 2) - Tradeoffs of Heuristics, Tree Search, External Memory, In-built Bias

o3 (Part 2): Tradeoffs of Heuristics, Tree Search, External Memory, In-built Bias

o3 is indeed groundbreaking, and shows that we might be close to finding a general training procedure that can self-improve with fine-tuning.

Here's part 2 of my discussion session on how o3 works based on my own understanding of it (or more generally, architectures that bootstrap learning via fine-tuning on correct trajectories)!

While o3 is powerful, I do not think o3-type architecture is the only way ahead for learning.

I believe that fine-tuning on own trajectories is slow to learn, and having a procedure to learn with external memory is very important (and missing) right now!

I also believe that learning from arbitrary start and end states in a trajectory is important - for instance, in math, we do not want to just learn the goal / model answer, but perhaps also how to reach every intermediate step of the solution.

Moreover, we should consider imbuing some biases so that we can reduce samples needed for training - like filters in Convolutional Neural Networks to bias for neighbouring pixels, so we do not need too many translations of the original image to learn translational invariance/equivariance.

~~~

Slides: https://github.com/tanchongmin/agentj...

References:

Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters - https://arxiv.org/abs/2408.03314
Learning, Fast and Slow: https://arxiv.org/abs/2301.13758
LLMs as a System of Multiple Expert Agents: https://arxiv.org/pdf/2310.05146

~~~

0:00 Introduction and Recap
2:21 Impressive Benchmark Scores
8:50 Is it AGI?
9:50 Correct Trajectories
35:45 Tree Search vs Parallel Search
57:14 Path Ahead - Adaptive Benchmarks
1:07:33 Learning, Fast and Slow
1:24:33 Multiple Abstraction Spaces / Neurosymbolic Integration
1:31:25 Discussion
1:54:30 Conclusion

~~~

AI and ML enthusiast. Likes to think about the essences behind breakthroughs of AI and explain it in a simple and relatable way. Also, I am an avid game creator.

Discord:   / discord  
LinkedIn:   / chong-min-tan-94652288  
Online AI blog: https://delvingintotech.wordpress.com/
Twitter:   / johntanchongmin  
Try out my games here: https://simmer.io/@chongmin