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.

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.

I’ve been an author on a few papers before, but I recently published the first research project where I was responsible for most of the work and direction. It’s in the first 2024 issue of the journal Artificial Life, which you can find here. You can find a non-paywalled version here Below, I tell the chronology of the project and summarize our findings.

We explore the conditions under which trade can emerge between four deep reinforcement learning agents that pick up and put down resources in a 2D foraging environment. Agents are rewarded for having both resources once, but the resources are distributed far apart from each other. To maximize reward, agents need to split up the work - agent 1 goes to resource A, agent 2 goes to resource B, etc, and then they meet to exchange resources, since meeting halfway can get them the most of each resource in the shortest amount of time.

How great would it be if the solution to most errors you face were in the first place you looked? That’s what the Arch Wiki has been for me: a massive wealth of information and troubleshooting resources to help me navigate the various configuration and installation issues I’ve encountered. Some people claim Arch Linux is too difficult for new users, but for me it’s been the only distribution I’ve been able to get consistently working, and it’s all thanks to the detailed documentation and known workarounds.

Flash attention, a recent implementation of attention which makes less calls to high-bandwidth memory, uses a version of the softmax function which is numerically stable. In this post, I’ll briefly showcase how this is done and an example of an unstable softmax.

The softmax function is used in machine learning to convert a vector of real numbers to a vector of probabilities which sum to 1, and is defined as:

Kittyplot is a program designed to plot experiment data in the kitty terminal using the kitty graphics protocol, primarily for use on HPC clusters.

Plots are rendered using matplotlib, and users can zoom into different regions of the plots by setting x and y limits using their editor. I use prompt_toolkit to accept regexp input and I override the tab-completion to instead display a list of all metrics that are matched by the current regexp.

Matthew Carlson’s blog post “Fighting Distraction With Unit Tests” inspired me to share some extra benefits of writing test code I’ve discovered during my PhD program.

I’m working on a weird project that’s constantly changing as I try new things, and naturally, debugging and ensuring correctness was a nightmare. So I started writing tests, cursing myself for needing to write so much code I’ll likely throw away soon. But as it turns out, testing can be pretty helpful in a few other ways: