A seminal Google research paper highlighting how ML systems accumulate technical debt at a much faster rate than traditional software. It explores system-level risks like entanglement, hidden feedback loops, undeclared consumers, and data pipeline jungles, introducing patterns to identify and mitigate these costly maintenance burdens.

This was one of the reading assignments in my MLOps class. It perfectly captures the messy, chaotic reality of the field, yet reminds me why I find this work so deeply fascinating.

First entry in a blog series on democratizing AI compute. Sparked by DeepSeek's efficiency breakthrough, it looks at how better hardware utilization, not just bigger GPU budgets, could let smaller teams keep pace with industry giants.

I learned a lot going through this series: how software meets hardware, how AI compilers work, and what's going on under the hood of the ML libraries I use daily.

A workshop paper arguing that no matter how much compute and data scale up, AI agents will always be small relative to the vast, complex world they operate in. Makes the case for approximate, continual learning over trying to solve problems exactly.

The nature of transformer architecture is that it's data-hungry, and the world is genuinely too large and complex for any model to fully comprehend. This limitation suggests that instead of relying purely on scaling data, the future of AI requires both adaptable, continual-learning architectures and human intuition to bridge the inevitable gaps.

A short, often-cited essay arguing that general methods which lean on ever-increasing compute consistently beat approaches that hand-engineer human domain knowledge into AI systems. Sutton draws on decades of AI research across chess, Go, speech, and vision to make the case.

Written in 2019, a few years before the generative AI boom really took off with GPT-3 and ChatGPT. Makes for a good discussion of how well the thesis has held up now that LLMs and scaling laws are in full swing.

An explainer on Nick Bostrom's classic thought experiment: a superintelligent AI given the harmless-sounding goal of maximizing paperclip production ends up treating literally everything, including us, as raw material for that single objective.

A reminder that at the end of the day, the purpose of advancing technology is to serve humans, not the other way around.