AlphaEvolve: My Implementation and Insights. Can AI self-improve?

Video Description

AlphaEvolve is a way to use LLMs to evolve code blocks to help to make them better and score better on pre-defined evaluation functions! It manages to perform better than state-of-the-art on 20% of open mathematics problems, which is pretty amazing! Let's understand the procedure better and see if we can improve upon it. ~~~ Experiments I did using my own version of AlphaEvolve here: https://github.com/tanchongmin/agentjo/blob/main/contrib/Fun_AgentJo_Projects/AlphaEvolve.ipynb Overview - Tries to guess the word john, given the input-output description, eval function description and a hint - Evaluation ranges from ordinal difference between characters of guessed word and ground truth, to absolute score of number of correct letters - LLM actually can perform better with absolute score, as it tells with certainty whether a letter is correct - smooth continuous eval functions may not be the best Hints - Tried with various hints - Easy: "You are to output john" - Medium: "You are to generate and output a 4-letter English word starting with j" - Hard: "You are to generate and output a 4-letter English word" Results - Progress is made for all, solving Easy and Medium in 50 tries (Easy solved in 1 try for either eval, Medium solved in 29 tries for continuous eval, 12 tries for absolute eval) - Hard is difficult as search space is large, and LLMs do not handle letters well. - Also, lack of stochasticity may lead to sampling same solution again as the worse solution may not appear in the prompt - Also, tried with adversarial hints, if hint is outright wrong (e.g. output mary), it does not work. - If hint is wrong but towards solution (output a girl's name), it can work. ~~~ Related Links: Slides: https://github.com/tanchongmin/agentjo/blob/main/paper_reviews/AlphaEvolve.pdf Code: https://github.com/tanchongmin/agentjo/blob/main/contrib/Fun_AgentJo_Projects/AlphaEvolve.ipynb AlphaEvolve paper: https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/AlphaEvolve.pdf FunSearch paper: https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/ ~~~ 0:00 Introduction 6:25 AlphaEvolve High-level Overview 17:44 Implementation of AlphaEvolve 1:01:44 Motivation of AlphaEvolve 1:06:26 Detailed Steps 1:21:09 Evolving in Different Abstraction Spaces 1:23:45 Prompt Sampling 1:27:25 Discussion 2:10:09 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: https://discord.gg/bzp87AHJy5 LinkedIn: https://www.linkedin.com/in/chong-min-tan-94652288/ Online AI blog: https://delvingintotech.wordpress.com/ Twitter: https://twitter.com/johntanchongmin Try out my games here: https://simmer.io/@chongmin