DeepSeek-R1's Epic Pull Request  ## Metadata - Author: [[Changelog Master Feed]] - Full Title: DeepSeek-R1's Epic Pull Request - Category: #podcasts - URL: https://share.snipd.com/episode/0442f5d9-cced-4909-8f89-70a353eebc36 ## Document Notes ## Highlights - **00:52 DeepSeek-R1's Impressive Pull Request** An AI model wrote 99% of a pull request that significantly improved the speed of WASM code, showcasing the potential of LLMs in low-level programming. > Surprisingly, 99% of the code in this PR is written by DeepSeekR1. The only thing I do is to develop tests and write prompts with some trials and errors. > > — **Xuan-Son Nguyen** **02:09 Tailwind CSS 4.0 Launch** Adam Wathan announces the release of Tailwind CSS version 4.0, a major update focusing on performance and flexibility, with a reimagined configuration and customization experience. > Tailwind CSS version four is an all new version of the framework optimized for performance and flexibility. > > — **Adam Wathan** **05:30 AI-Induced Programmer Illiteracy?** A software developer confesses to relying heavily on AI for coding tasks, raising concerns about the potential for decreased problem-solving skills and long-term understanding among programmers. > We're not becoming 10x developers with AI. We're becoming 10x dependent on AI. There's a difference. > > — **Nemanye Goel** Transcript: Jerod Santo Speaking of big brain, Zan's son Nguyen opened a pull request to Georgi Gaganov's Llama.cpp repo that doubles the speed of WASM by optimizing SIMD instructions with the following PR comment. Quote, surprisingly, 99% of the code in this PR is written by DeepSeekR1. The only thing I do is to develop tests and write prompts with some trials and errors. Indeed, this PR aims to prove that LLMs are now capable of writing good low-level code to a point that it can optimize its own code. End quote. I can't judge whether this is good low-level code or not because I don't know what good low-level code looks like, but Georgi and Zonson sure are impressed. Zonson also shared the prompts they used to get the desired results. This of course resulted in a long x-thread where both humans and robots debate and meme whether or not it's over for folks like us or not quite yet. Tailwind CSS version four is official. Here's Adam Wavin. Quote, Tailwind CSS version four is an all new version of the framework optimized for performance and flexibility. With a reimagined configuration and customization experience and taking full advantage of the latest advancements the web platform has to offer. End quote. This looks like it was a massive undertaking. It has a new high-performance build engine, simplified installation, automatic content detection, reimagined, CSS-first config, and too much more to list here. The most influential papers in computer science history. Matthias Lima opens up the history books to create this admittedly subjective list of influential papers dating all the way back to 1936. Quote, these seven papers, sorted by date, stand out to me mostly because of their impact in today's world. For each paper, Matthias provides the big idea and why he thinks it still matters to this day. Here's the quick list. One, uncomputable numbers with an application to the N-Shidenzong problem by Alan Turing in 1936. Two, a mathematical theory of communication by Claude Shannon, 1948. Three, a relational model of data for large shared data banks. Edgar F. Codd, 1970. 4. The complexity of theorem proving procedures. Stephen A. Cook, 1971. 5. A protocol for packet network intercommunication. Vinton G. Cerf and Robert E. Kahn, 1974. Six, Information Management, a proposal by Tim Berners-Lee, 1989. And seven, The Anatomy of a Large-Scale Hypertextual Web Search Engine by Sergey Brin and Larry Page, 1998. He also provides a bonus list of five papers that almost made the list, finishing with this, quote, These days we are flooded with new stuff, fresh languages, mind-blowing AI breakthroughs, Quantum leaps, and the JavaScript framework of the week. It's all super exciting, but here's the thing. Foundations matter. Without them, we're just piling on new toys without fully understanding the ground we're building on. It's now time for sponsored news. Replay 2025 in London, March 3rd through 5th. Our friends at Temporal invite you to replay in London, March 3rd through 5th, to break free from the status quo. Replay 25 is an in-person conference focused on transitioning away from outdated monolithic systems and methodologies to embrace cutting edge technologies. Immerse yourself in two days of technical talks from back-end software engineering leaders at top organizations. Then enjoy connecting on day three at the after party. Live it up, connect, and continue conversations with food, drinks, and fun alongside your replayplay community. Early bird tickets are on sale now. Early bird pricing ends January 31st, which is right around the corner, so get your tickets soon if you plan to attend. Learn more and register at replay.temporal AI is creating a generation of illiterate programmers. Nemanye Goel has a confession to make. Quote, a couple of days ago, cursor went down during the chat GPT outage. I stared at my terminal facing those red error messages that I hate to see. An AWS error glared back at me. I didn't want to figure it out without AI's help. After 12 years of coding, I'd somehow become worse at my own craft. And this isn't hyperbole. This is the new reality for software developers. End quote. He doesn't think he's the only one who's become a human clipboard, a mere intermediary between his code and an LLM. Quote, we're not becoming 10x developers with AI. We're becoming 10x dependent on AI. There's a difference. Every time we let AI solve a problem we could have solved ourselves, we are trading long-term understanding for short-term productivity. ([Time 0:01:03](https://share.snipd.com/snip/426ed6a2-48f8-4299-83e6-21a05ddc87bc)) - **AI-powered Pull Request** - Xuan-Son Nguyen used DeepSeek-R1 to optimize Llama.cpp's WASM speed. - The pull request doubled the speed, showcasing AI's low-level coding capabilities. Transcript: Jerod Santo Speaking of big brain, Zan's son Nguyen opened a pull request to Georgi Gaganov's Llama.cpp repo that doubles the speed of WASM by optimizing SIMD instructions with the following PR comment. Quote, surprisingly, 99% of the code in this PR is written by DeepSeekR1. The only thing I do is to develop tests and write prompts with some trials and errors. Indeed, this PR aims to prove that LLMs are now capable of writing good low-level code to a point that it can optimize its own code. End quote. I can't judge whether this is good low-level code or not because I don't know what good low-level code looks like, but Georgi and Zonson sure are impressed. Zonson also shared the prompts they used to get the desired results. This of course resulted in a long x-thread where both humans and robots debate and meme whether or not it's over for folks like us or not quite yet. ([Time 0:01:03](https://share.snipd.com/snip/11325f16-b15f-4369-89de-6eae02d17a68)) - **Tailwind CSS v4 Release** - Tailwind CSS v4 prioritizes performance and flexibility. - It leverages the latest web advancements with a redesigned configuration. Transcript: Jerod Santo Tailwind CSS version four is official. Here's Adam Wavin. Quote, Tailwind CSS version four is an all new version of the framework optimized for performance and flexibility. With a reimagined configuration and customization experience and taking full advantage of the latest advancements the web platform has to offer. End quote. This looks like it was a massive undertaking. It has a new high-performance build engine, simplified installation, automatic content detection, reimagined, CSS-first config, and too much more to list here. ([Time 0:02:09](https://share.snipd.com/snip/54d28988-3941-4d13-a87c-43b77456f2fd)) - **Influential CS Papers** - Matheus Lima curated a list of influential computer science papers. - He emphasizes the importance of foundational knowledge amidst new technologies. Transcript: Jerod Santo The most influential papers in computer science history. Matthias Lima opens up the history books to create this admittedly subjective list of influential papers dating all the way back to 1936. Quote, these seven papers, sorted by date, stand out to me mostly because of their impact in today's world. For each paper, Matthias provides the big idea and why he thinks it still matters to this day. Here's the quick list. One, uncomputable numbers with an application to the N-Shidenzong problem by Alan Turing in 1936. Two, a mathematical theory of communication by Claude Shannon, 1948. Three, a relational model of data for large shared data banks. Edgar F. Codd, 1970. 4. The complexity of theorem proving procedures. Stephen A. Cook, 1971. 5. A protocol for packet network intercommunication. Vinton G. Cerf and Robert E. Kahn, 1974. Six, Information Management, a proposal by Tim Berners-Lee, 1989. And seven, The Anatomy of a Large-Scale Hypertextual Web Search Engine by Sergey Brin and Larry Page, 1998. He also provides a bonus list of five papers that almost made the list, finishing with this, quote, These days we are flooded with new stuff, fresh languages, mind-blowing AI breakthroughs, Quantum leaps, and the JavaScript framework of the week. It's all super exciting, but here's the thing. Foundations matter. Without them, we're just piling on new toys without fully understanding the ground we're building on. ([Time 0:02:44](https://share.snipd.com/snip/29866a75-2019-4ecd-a898-74bc6486bf10)) - **AI Dependency Concerns** - Namanyay Goel argues that AI is creating dependence, not skill improvement. - Over-reliance on AI tools may hinder long-term understanding for developers. Transcript: Jerod Santo Nemanye Goel has a confession to make. Quote, a couple of days ago, cursor went down during the chat GPT outage. I stared at my terminal facing those red error messages that I hate to see. An AWS error glared back at me. I didn't want to figure it out without AI's help. After 12 years of coding, I'd somehow become worse at my own craft. And this isn't hyperbole. This is the new reality for software developers. End quote. He doesn't think he's the only one who's become a human clipboard, a mere intermediary between his code and an LLM. Quote, we're not becoming 10x developers with AI. We're becoming 10x dependent on AI. There's a difference. Every time we let AI solve a problem we could have solved ourselves, we are trading long-term understanding for short-term productivity. We are optimizing for today's commit at the cost of tomorrow's ability. End quote. Does this sentiment resonate with you? If so, see also a linked recent paper on metacognitive laziness. ([Time 0:05:30](https://share.snipd.com/snip/2e354f06-6556-445b-a25c-a71fdd625357)) - **WFH Lighting Tips** - Russell Baylis suggests optimizing home lighting for reduced eye strain. - He recommends diffused lighting, proper brightness, and utilizing natural light. Transcript: Jerod Santo How to improve work from home lighting to reduce eye strain. Russell Bayless is not an ergonomist or optometrist. He's just a worker from Homer who is susceptible to eye strain, eye pain, and dizziness. In the linked post, Russell shares what he's learned about optimizing home lighting to reduce eye strain. Here's the quick list. One, an even diffused lighting environment is the best for the eyes. Two, when it comes to light brightness, too much is just as problematic as too little. Three, use natural light wherever possible. Four, quality of artificial light matters. Five, the best lighting for camera is not necessarily the best lighting for ergonomics. And six, even the perfect lighting environment will fatigue you. Take breaks and take care of yourself. ([Time 0:06:30](https://share.snipd.com/snip/62e8a14e-fe66-4102-bb05-030ad0e9e08b))