My latest paper is available on arxiv: Low Rank Factorizations are Indirect Encodings for Deep Neuroevolution.

The general idea is that we can search for stronger neural networks in a gradient-free fashion by restricting search to networks of low-rank. We show that it works well for language modeling and reinforcement learning tasks. It’s essentially a crossover between the following papers:

• LoRA: Low-Rank Adaptation of Large Language Models
• Deep Neuroevolution: Genetic Algorithms Are a Competitive Alternative for Training Deep Neural Networks for Reinforcement Learning.

I’ll be presenting it virtually for the Neuroevolution@Work workshop at GECCO 2025.

If all goes well, I’ll leave Brandeis University with a PhD degree in computer science this summer. Here are my parting thoughts for our remaining students, divided into technical and non-technical sections.

I write this out of self-interest; students after me will represent our university in years to come, and it’s in my best interest for them to be as effective as possible.

NON-TECHNICAL

• Students often lack crucial skills. Technical students may lack communication skills, or social butterflies may lack depth. Some lack both. This doesn’t really matter, in my opinion, if they are willing to learn. Just like neural networks, they may be all over the place initially, but students who reflect on your feedback will (eventually) be worth more than those who never update their parameters. Don’t give up on those who listen and try; they often just need the right kind of mentorship.

I used to have a mental block against casually reading things that weren’t designed to be casually read, like old papers, textbooks, documentation, etc. If I was going to read for enjoyment, I’d read blogs or popular science/fiction books. Not sure why.

Over time, I’ve met people who read things I’d never think to read. I’ve gotten a lot of value from copying them, so I’m sharing my thoughts here.

Starting from REINFORCE, the original deep reinforcement learning algorithm, we will trace the evolution of policy gradient methods to the Group Relative Policy Optimization algorithm used to train Deepseek r1.

This post ignores the LLM side of things, less-related developments in RL, and most of the equations used for these algorithms, but captures the essence and intuition of the RL-timeline without wasting your time. This is all self-study, so feel free to send me any corrections/suggestions.¹

I just finished The Story of Civilization, Vol. 1 (1976) by Will Durant. It’s a fantastic read; Durant is eloquent, witty, and surprisingly relevant today. Below are some of my personal highlights of his writing, mixed with some quotes-of-quotes I found interesting.

Sections: Introduction | Communism | Government | Women | Morals | Language | Appearance | Sumeria | Egypt | Babylon | Assyria | Judea | Persia | India | China | Japan | Conclusions

Introduction

“the provincialism of our traditional histories, which began with Greece and summed up Asia in a line, has become no merely academic error, but a possibly fatal failure of perspective and intelligence”

I don’t think r1 will get us to artificial super intelligence, but whatever comes next probably will.

We are reaching a familiar bottleneck in AI. Previously, humans had to manually hardcode the patterns that AI could recognize. With deep learning, machines began to learn patterns on their own, without human assistance. With (relatively) expensive humans out of the loop, we threw machines at the world’s data until they began to talk, code, and paint. Many people believed this would be sufficient to reach artificial super intelligence–but it wasn’t.

I’m doing an internship right now, and thankfully, I read a few books on software design before starting.

I had to design a database schema, data submission page, submission approval page, and dynamic dashboard for the project I was assigned to. This is one of those projects that AI can obviously do 90% of to work for if designed appropriately–if the right abstractions are used, performance trade-offs are made, and the right tools are chosen. I wanted to make sure that, throughout the project, AI could always easily with any part of the codebase.

Wikipedia translates the Greek philosophical concept of Eudaimonia as “Human Flourishing”, particularly on the spiritual level. It refers to genuine happiness and fulfillment over hedonism or GDP.

The modern world feels meaningless at times. Instead of hunting or farming to put food on the table, many of us sit at a computer or stand behind a counter, far removed from the fruits of our labor—if our labor even bears fruit. Our screens bombard us with videos of people more beautiful and successful than we could ever hope to be while we stay inside, alone.

It used to be good enough just to copy others. Now, with AI in the hands of billions, there’s little value in copying.

For instance, take programming. Five years ago, building apps, websites, or games required a non-trivial amount of skill, and getting your first project off the ground was an accomplishment. Now, AI can generate most starter projects in hours, if not minutes. I think this decimates the reward, both internal and external, of actually completing the first few projects.

CUPERTINO, CALIFORNIA - Apple today announced the release of the groundbreaking iPad Pro Max, the world’s first tablet powered by the all-new quantum Q1 processor. Capable of 30 octillion floating point operations per second, the Q1 makes the iPad Pro Max the most powerful streaming device ever made, over 2x faster than the M4 iPad Pro. Available in both 11-inch and 13-inch sizes, Apple says that the iPad Pro Max is the perfect device for generating TV shows, movies, and games in real-time on the go using the power of AI, no network connection needed.