Rami's Readings #99 - DeepSeek 🔥 & Reading Widely
The latest on AI, LLMs, DeepSeek Engineering Fundamentals, Phi 4, Tyler Cowen, Reading Widely, Matcha, and more.
Welcome to Rami’s Readings #99 - a weekly digest of interesting articles, papers, videos, and X threads from my various sources across the Internet. Expect a list of reads covering AI, technology, business, culture, fashion, travel, and more. Learn about what I do at ramisayar.com/about.
🤖 AI Reads
DeepSeek AI Released DeepGEMM, smallpond, and 3FS: Engineering Fundamentals 🔥
Notes: DeepSeek is showing how engineering fundamentals remains critical to solving problems at scale. DeepGEMM is a masterclass in hardware-aware programming - DeepSeek wanted to extract as much speed up as possible from Hopper-class GPUs, so they eschewed CUDA and built their own matrix multiplication library (2.7x faster). It’s already making its way into cutlass. They built their own distributed file system (3FS). Smallpond builds on it.
SiriuS: Self-improving Multi-agent Systems via Bootstrapped Reasoning
Notes: Excellent paper from Stanford on improving multi-agent systems with reasoning. The paper goes counter to other recent papers by showing that specialization does help improve reasonining.
Llama Stack
Notes: Open source building blocks for AI applications.
The Next Generation of the Phi Family
Notes: New Phi-4-multimodal and Phi-4-mini models.
Transforming Science with LLMs: A Survey on AI-assisted Scientific Discovery, Experimentation, Content Generation, and Evaluation
Notes: Great survey from multiple universities on how AI models are transforming science.
💼 Business Reads
Tyler Cowen, the Man Who Wants to Know Everything
Notes: I frequently reference Tyler Cowen in this newsletter and share his opinion pieces from Bloomberg. I started reading his work ten years ago, and while I don’t agree with all his viewpoints, I admire his breadth of reading and sharing. His “assorted links” posts on Marginal Revolution are a source of inspiration for this newsletter.
Many of my subscribers ask how I read so fast. Well, I’m not as fast as Tyler, but speed reading is a skill I developed in my senior year of high school. Back then, I loved spending most of my day reading Visual C++ 6 books or browsing through the library. As you read more and read widely, you start to notice that most published works are repetitive within themselves and rarely introduce unique ideas. Honestly, it seems every New York Times nonfiction best seller is a rewrite of another book published >10 years ago. Today, I skip reading most new books and wait a year to see if they’re still being talked about.
Once you recognize writing patterns, common ideas or derivations more quickly, you naturally increase your reading speed and expand the range of material you consume. For technical papers, I learned a scanning technique to quickly judge how much time to invest. Should I host office hours on this topic?
A former colleague once joked about working with me, saying, “He’s like a meta-search engine.” (Sorry for repeating this Shawn) That’s because, for most technical topics, I could vaguely recall reading something about it and quickly provide the key search terms needed to find the right document. In many ways, this newsletter is an extension of my high-volume reading and sharing habit.
How do you approach learning in your daily life?
The Global Matcha Boom Is Driving a Shortage in Japan
Notes: A Matcha or coffee shortage would be disastrous for my household.
Map of New Unicorns
Notes: This diagram is from Ilya Strebulaev at Stanford GSB.
🛡️CyberSecurity Reads
How to Gain Code Execution on Millions of People and Hundreds of Popular Apps
Notes: This isn’t surprising if you’re familiar with how most web development is done. With the rapid adoption of AI-powered coding models for web development, I expect an explosion of these kinds of vulnerabilities. Unfortunately, software supply chain security remains an unsolved problem—even for big tech.
Signing off from Redmond, WA.