A curated collection of articles, videos, and posts I find interesting, including what I think are the most important historical articles in computer science and AI. Let me know if you have any other articles that would go well here.

May 30, 2026 Domain Expertise Has Always Been the Real Moat Aaron Brethorst I love it when people say obvious true things succinctly. We are missing this in all the back-and-forth AI hype swings going on right now.

May 22, 2026 The Coddling of the Tech Mind Nikunj Kothari The last sentence is the essence of what's wrong in many companies today when it comes to agency and empowerment and thus engagement and impact. Enjoy!

May 16, 2026 Detecting and preventing distillation attacks Anthropic I applaud Anthropic's approach to being careful about China and their intentions.

May 5, 2026 When everyone has AI and the company still learns nothing Robert Glaser Nobody has figured this out - but this is a great description of the problem and a start at a solution.

April 29, 2026 10 Tricks That Will Make You the Best Listener in the Room TIME I found these useful and am still working on a number of them...

April 23, 2026 The "AI Vulnerability Storm": Building a "Mythos-ready" Security Program Cloud Security Alliance Labs Good intro to the impact of Mythos (preview). Attackers will have (already have?) tools like this — Defenders better have equivalent or better tools — and more urgency...

April 17, 2026 Every engineer should work this way... The Pragmatic Engineer I normally don't recommend podcasts, but this one is different. Up until 47 minutes in, you get to hear Boris' and Anthropic's origin story — and then for the remaining 45 minutes of the podcast you get to learn how Anthropic operates. Pure gold.

April 15, 2026 Anthropic shifts enterprise billing to per-token pricing. The flat-fee era is over. Implicator.ai Anybody who didn't think this was coming wasn't doing the math. But that's not the real point — which is that the market is desperately trying to find a stable equilibrium, which is needed. This is good.

April 2, 2026 Embarrassing indeed Apple This paper shows how little we know about actually what happens inside LLMs. If we can change the behavior of an LLM by 10% by doing something so, as Apple says, embarrassingly simple, then it shows that we are in the early early days of this entire revolution that we are experiencing.

March 25, 2026 Thoughts on Slowing the Fuck Down Mario Zechner I don't necessarily agree with the extreme position here, but it does make some points. And... it's funny :-)

March 24, 2026 TurboQuant: Redefining AI Efficiency with Extreme Compression Google Research If inference gets 10x cheaper in memory, does the entire money-house-of-cards fall apart? Maybe we don't have to bankrupt the planet (i.e. the Matrix)

March 12, 2026 Months In: What Building with AI Actually Feels Like Joe Lynch Joe points out a lot of strengths and weaknesses in this article which happen to unsurprisingly (at least to me) align perfectly with Yann LeCun's concerns about the limitations of LLMs.

March 9, 2026 Applying Statistics to LLM Evaluations Cameron R. Wolfe / Deep (Learning) Focus I found this very clarifying at the basic level I needed.

March, 2026 Dylan Patel — The Single Biggest Bottleneck to Scaling AI Compute Dwarkesh Patel This house-of-cards of AI power consumption will fall. Bet against it. The silliness of this hype is unfathomable.

February 11, 2026 No Coding Before 10am Michael Bloch / Substack For any of you who've read "The Phoenix Project", you remember that Brett was the problem. Brett is no longer the problem. Deep meaningful thought about customers is now the limiting factor.

December, 2025 Why LinkedIn Is Replacing PMs with AI-Powered "Full-Stack Builders" Lenny Rachitsky / Lenny's Podcast My whole career, I've watched engineers pigeon-hole themselves into being water carriers for their PMs. This was always silly. Now it's unviable. EVERYBODY needs to have customer delight at the center of their cognitive decision matrix.

December, 2025 There Is Something Faster Than Light Veritasium Ok, this is completely off topic for my website. But enjoy it anyways!

October 11, 2025 The Circular AI Money Shell Game Ryan Levesque / The Digital Contrarian Ponzi schemes always end up failing. And the billionaires will act surprised as all the people they lied to lose their money. There is a rea$on that all these people are giggling behind their hands as they lie to the rest of us.

August 27, 2025 Apple Intelligence Foundation Language Models: Tech Report 2025 Apple / arXiv Apple's tech report on their on-device 3B model and server-side Parallel-Track Mixture-of-Experts transformer powering Apple Intelligence — KV-cache sharing, multimodal, tool use, and Private Cloud Compute.

January 20, 2025 DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning DeepSeek-AI / arXiv DeepSeek matched OpenAI o1 on reasoning benchmarks for roughly one-tenth the training cost using pure RL with verifiable rewards. The cost moat assumption — that frontier AI required US hyperscaler budgets — ended with this paper. If you work in AI infrastructure, this is the paper that changed the economics of your industry.

November 19, 2024 2024 Accelerate State of DevOps DORA / Google Cloud Even in the AI world, this is no less relevant than it ever was.

September 30, 2024 Queryable Encryption Technical Paper MongoDB Queryable Encryption, invented by geniuses Tarik and Seny and implemented by very clever Kenn White and others, is an underappreciated technology.

February, 2024 Is the "Modern Data Stack" Still a Useful Idea? The Analytics Engineering Roundup Tristan talks about how his initially groundbreaking idea is now changing. Sadly, despite the brilliance of Tristan, the success of dbt Labs, and a decade of work, data work still remains a hack at most companies.

January, 2024 3Blue1Brown's Deep Learning Series — Seven Chapters That Actually Explain Neural Networks 3Blue1Brown Grant Sanderson's seven-chapter visual walkthrough of neural networks, from the basics through transformers and how LLMs might store facts. The single best resource I've found for understanding what's actually happening inside modern AI.

March 22, 2023 Sparks of Artificial General Intelligence: Early Experiments with GPT-4 Bubeck et al. / arXiv (Microsoft Research) This paper arrived before GPT-4 was publicly available and argued — in 100+ pages of examples — that what we were seeing was qualitatively different from prior AI. Not everyone agreed with the framing, but watching a model solve novel problems in mathematics, law, and vision with human-level competence made it hard to argue the old benchmarks were still the right frame. The Sparks debate set the tone for everything that followed in 2023.

February, 2023 dbt Labs Names Data Industry Veteran Mark Porter as Chief Technology Officer dbt Labs Blog I was delighted to join dbt Labs, working for Tristan Handy and working with such amazing people as Sarah Riley, Ryan Segar, Brandon Sweeney, and (most of all) Meg Pittman. So very much work accomplished over two years to basically rework every layer of the software, operational, and human stack in Engineering at dbt Labs. An amazing experience I will always be thankful for.

June, 2022 A Path Towards Autonomous Machine Intelligence Yann LeCun / OpenReview My non-PhD intuition says he's right. Totally right. LLMs in their current (2026) form are a global max for conversations with humans, but a local max for actual useful intelligence and insight. I'm betting on the energy and self-training model.

June, 2022 We Went To MongoDB World 2022 So You Didn't Have To! See What You Missed! Linode When I went to MongoDB, I focused on getting the tech-focused team to be customer focused. 2022 was the first public release of a lot of what I accomplished. It was a start. Jim Scharf (MongoDB CTO) is focusing on what needs to be focused on - mission-critical performance, efficiency, availability, and operability. Will it be enough for MongoDB to succeed when the vast majority of agents suggest Supabase rather than MongoDB Atlas? (which is a shame)

June, 2022 MongoDB's New Tool to Migrate Data from Relational Systems The New Stack Strict relational is great. MongoDB documents are great. I hit a huge amount of organizational resistance to building and releasing this tool - I still can't figure out why. This tool was built to help people have choices.

March 4, 2022 Training Language Models to Follow Instructions with Human Feedback Ouyang et al. / OpenAI / arXiv This is the paper that turned GPT-3 into something you could actually talk to. RLHF — reinforcement learning from human feedback — is the technique that made language models useful rather than just impressive. Every chat AI you've used since 2022 runs on some version of this idea. Simple concept, massive practical consequence.

January 28, 2022 Chain-of-Thought Prompting Elicits Reasoning in Large Language Models Wei et al. / Google Brain / arXiv The discovery that simply asking a model to "think step by step" dramatically improved performance on reasoning tasks was both obvious in hindsight and genuinely surprising at the time. This paper formalized something that prompt engineers had started to notice empirically. It also opened up the question that still isn't settled: are the models reasoning, or very cleverly pattern-matching to reasoning-shaped outputs?

January 27, 2022 Is Relational The New COBOL? What The History Of Technology Tells Us About Change MongoDB Blog The blog-post companion to my AWS re:Invent 2021 talk — tracing the trends in data, computing, and programming from the 1970s onward to explain why relational is no longer the natural default for new apps.

July 15, 2021 Highly Accurate Protein Structure Prediction with AlphaFold Jumper et al. / DeepMind / Nature Fifty years of the protein folding problem, solved. Biology and chemistry have been stuck on this since 1972 and DeepMind cracked it with deep learning. This is the clearest example I know of AI producing a scientific result that humans could not have produced by any other means in any reasonable timeframe. Every drug discovery pipeline in the world changed after this paper.

July, 2021 MongoDB 5.0: Worth the Wait Database Trends and Applications This was my first release as MongoDB CTO. It was a start...

July, 2021 The Difference Between R and D MongoDB Blog Michael Cahill's piece on what's next at MongoDB.

May, 2021 One of Dwight's first slide presentations on MongoDB, 2009 myNoSQL Dwight and Eliot changed the world.

February, 2021 Your Company Is Too Risk-Averse Harvard Business Review When I wrote this, I wondered if this was MongoDB or not. To this day, I don't know. But the fact that MongoDB, which is a fabulous product, is still at less than 10% market share, tells me we needed to accept more risk and take more chances.

February, 2021 Cracking the Code of Sustained Collaboration Harvard Business Review "Teach People to Listen, Not Talk" - Communicating is really hard. I'm working on this too. Stop deluding yourself that it's the tech that makes producing great products hard. It's the people.

February, 2021 How Andy Jassy Was Trained by Jeff Bezos CNBC One way to build leaders — how Amazon does it

February, 2021 Things David Perell Learned from Jeff Bezos David Perell Key things to learn about how Bezos runs Amazon

February, 2021 Andy Jassy's 8 Steps to Reinvent a Business SiliconANGLE How companies can reinvent themselves

February, 2021 Connect Then Lead Harvard Business Review This is a top re-read for me. Once a quarter. People above you in the org chart expect competence first and then warmth second (though some toxic cultures don't care at all). People below you expect warmth first, that you care about them, and then competence second. This tightrope is crazy hard.

January, 2021 Secure Multiparty Computation Communications of the ACM Tarik and Seny (if you know who I mean, good for you) are absolutely brilliant. I learned so much from them, and I think I only learned maybe 15% of what they know. Queryable Encryption (which they invented and we hold co-patents on) easily wins as the most important innovation that never took over the world that I've been part of.

January, 2021 Stepping Up for a Truly Open Source Elasticsearch AWS Open Source Blog Quite an interesting way to wake up the open source community. Watching Amazon navigate their increased commitment to open source over the last 5 years has been a pleasure.

January, 2021 Elastic License v2 Elastic Blog Watching Shay navigate his "Not OK" drama with Amazon was...stressful. While I think it ended well, including the current Elastic License being a work of art, the market went through needless stress.

November, 2020 The Feedback Fallacy Harvard Business Review "The first problem with feedback is that humans are unreliable raters of other humans." So sadly true. "Focusing People on their shortcomings doesn't enable learning; it impairs it" - Taytay got this right in her song 'Mean'. Your job as a leader is to explain what excellence looks like. And give agency to your employees to aspire to that, to achieve it; or to move on.

October, 2020 High-Performing Teams Need Psychological Safety Harvard Business Review Can we feel safe in a high performing team?

October, 2020 Hanlon's Razor Farnam Street Maybe they aren't all out to get you?

September, 2020 The Authenticity Paradox Harvard Business Review It truly is a paradox. When people ask you to be authentic, they are really asking you to be exactly what they want. When you show up as yourself, with all your strengths and flaws equally exposed, you get skewered by most c-levels, especially founders, who have more blindness in this area than the rest of us. So... figure out your authentic self. Figure out what the company needs. And slowly, while keeping your job, make them the same thing.

August, 2020 The Essential Clayton Christensen Articles Harvard Business Review Clay Christensen was a gift and I continue to be sad he is no longer with us

July, 2020 The Principle of Charity Effectiviology A piece on what it means to believe in best intentions

July, 2020 The Seven Transformations of Leadership Harvard Business Review I personally find my Diplomat very challenged

July, 2020 Steve Yegge's Medium Page Medium One of the funniest blogs on the Internet

June 19, 2020 Denoising Diffusion Probabilistic Models Ho, Jain & Abbeel / UC Berkeley / arXiv The paper that made diffusion models work well enough to replace GANs for high-quality image generation. Every image generation system you've used — Stable Diffusion, DALL·E, Midjourney — traces its architecture back here. The insight is surprisingly physical: model the reverse of a noise-adding process and you can generate images from pure static.

May 28, 2020 Language Models are Few-Shot Learners Brown et al. / OpenAI / arXiv GPT-3 was the moment scale became a strategy. 175 billion parameters and the model could do tasks it was never explicitly trained for — just from a few examples in the prompt. The "few-shot learning" capability felt like magic the first time you saw it. This paper ended the debate about whether big language models were a dead end and started the race we're still running.

January 23, 2020 Scaling Laws for Neural Language Models Kaplan et al. / OpenAI / arXiv This paper gave OpenAI a roadmap. It showed that model performance follows clean power laws with respect to compute, data, and parameters — and that bigger models are more sample-efficient. The implication was blunt: if you had the money, you knew in advance how good your model would get. That's not a research insight, that's a business plan. It's why the labs started spending billions.

June 12, 2017 Attention is all you need Arxiv The paper that started a very important discussion

January 28, 2016 Mastering the Game of Go with Deep Neural Networks and Tree Search Silver et al. / DeepMind / Nature Go was supposed to be the last human-holdout board game — 10^170 possible positions, too complex for brute force, requiring "intuition" that computers supposedly couldn't replicate. AlphaGo beat Lee Sedol 4-1 in 2016 and it wasn't close. Combining deep learning with Monte Carlo tree search, DeepMind demonstrated that AI could master domains previously thought to require human-like reasoning. The follow-on AlphaZero, which learned from self-play alone, made this result even more extraordinary.

January 16, 2013 Efficient Estimation of Word Representations in Vector Space Mikolov et al. / Google / arXiv King minus Man plus Woman equals Queen. That equation, which works in Word2Vec's embedding space, is one of the most striking results in the history of NLP. The idea that meaning could be encoded as arithmetic on vectors unlocked everything that followed — document similarity, recommendation systems, and eventually the embedding layers at the core of every transformer. Tomas Mikolov's two-page paper changed how we think about what language models are learning.

December 3, 2012 ImageNet Classification with Deep Convolutional Neural Networks Krizhevsky, Sutskever & Hinton / NeurIPS 2012 AlexNet won the 2012 ImageNet competition by such a large margin — 10.8 percentage points — that the field collectively pivoted overnight. Before this, most computer vision researchers were skeptical that deep neural networks were worth pursuing. After this, everyone was building GPUs and stacking layers. This is the paper that started the deep learning revolution. Hinton, LeCun, and Bengio won the Turing Award in 2018 for the work that culminated here.

August 19, 2006 A Tutorial on Energy-Based Learning LeCun, Chopra, Hadsell, Ranzato, Huang / Predicting Structured Data (MIT Press) The conceptual root of everything LeCun later branded as "world models." Twenty years on, the EBM framing in this tutorial is still what JEPA and H-JEPA are built on.

February, 2003 A Neural Probabilistic Language Model Bengio, Ducharme, Vincent, Jauvin / JMLR The paper that introduced word embeddings — learned, continuous vector representations of words — as a byproduct of training a neural language model. This was the original "language models learn representations" result. Everything from Word2Vec to GPT-4's token embeddings traces its conceptual lineage here. Bengio was doing in 2003 what the entire field caught up to a decade later.

November, 1998 Gradient-Based Learning Applied to Document Recognition LeCun, Bottou, Bengio & Haffner / Proceedings of the IEEE LeNet-5 was reading handwritten zip codes on US mail at 10 frames per second in 1998. The architecture — convolution, pooling, convolution, fully connected — is the template that AlexNet and every CNN since has used. The tragedy is that it took another 14 years for the hardware and data to catch up to what Yann LeCun already knew would work. He was right in 1998. The world just wasn't ready yet.

April, 1998 The Anatomy of a Large-Scale Hypertextual Web Search Engine Sergey Brin & Lawrence Page / Computer Networks PageRank started as a PhD thesis. Two Stanford students noticed that citation patterns in academic papers could rank web pages by importance, and from that one insight built a trillion-dollar company. What's wild is that Google doesn't really use PageRank anymore — but the business it started is still very much here.

June, 1995 A Critique of ANSI SQL Isolation Levels ACM SIGMOD 1995 It was a privilege to work with Hal, one of the authors of this paper...

March, 1995 Temporal Difference Learning and TD-Gammon Gerald Tesauro / Communications of the ACM The first program to beat world-class human players at a complex board game — a decade before Deep Blue and two before AlphaGo. Tesauro trained a neural network purely through self-play and temporal difference learning, with no expert knowledge beyond the rules. The AI community largely ignored TD-Gammon. That was a mistake. Sutton and Barto's reinforcement learning textbook is dedicated partly to results like this one.

June, 1994 Oracle Media Server: providing consumer based interactive access to multimedia data SIGMOD Record The paper that defined truly scalable media servers — well before its time.

October 9, 1986 Learning Representations by Back-Propagating Errors Rumelhart, Hinton & Williams / Nature This is the algorithm that trains every neural network in the world. The chain rule of calculus, applied layer by layer, telling each weight exactly how much it contributed to the error and in which direction to adjust. Backpropagation had been discovered before — Werbos had it in 1974, Parker in 1985 — but Rumelhart, Hinton, and Williams published it in Nature with compelling experiments that made the field take notice. You cannot overstate how foundational this is. Every gradient descent step in every AI system is this paper.

July, 1978 Time, Clocks, and the Ordering of Events in a Distributed System Leslie Lamport / Communications of the ACM I've spent years of my career thinking about distributed systems, and every time I go back to first principles I end up here. There's no global clock. Events don't have a natural order. The only ordering we have is the one we construct from message passing. Lamport figured all of this out in 1978 — before the internet existed. This is the paper I recommend most to engineers joining a distributed systems team.

February, 1978 A Method for Obtaining Digital Signatures and Public-Key Cryptosystems Rivest, Shamir & Adleman / Communications of the ACM Three people took Diffie-Hellman's idea and built something you could actually ship a product on. That product now secures every credit card transaction, every encrypted email, every HTTPS connection in the world. Rivest, Shamir, and Adleman got the Turing Award in 2002. Worth the wait.

November, 1976 New Directions in Cryptography Whitfield Diffie & Martin E. Hellman / IEEE Transactions on Information Theory Before this paper, two people who had never met could not establish a secret. That wasn't a software problem — it was a mathematical impossibility. Then Diffie and Hellman solved it. Every HTTPS connection, every Signal message, every SSH session rests on an idea in this paper.

May, 1974 A Protocol for Packet Network Intercommunication Vinton G. Cerf & Robert E. Kahn / IEEE Transactions on Communications The internet was designed. Not discovered, not evolved by accident — designed. Two people figured out how to make heterogeneous networks talk to each other. TCP/IP works so well we've stopped appreciating it was a choice. It could have gone a completely different way.

May, 1971 The Complexity of Theorem-Proving Procedures Stephen A. Cook / ACM STOC P vs. NP is still unsolved. We don't know if the hard problems are actually hard or just hard for us. Cook formalized this question in 1971 and it's been one of the most consequential open problems in mathematics ever since. There's a million-dollar prize for solving it. Nobody has claimed it. Sleep on that.

June, 1970 A Relational Model of Data for Large Shared Data Banks E. F. Codd / Communications of the ACM The paper that started it all. Every relational database we've ever touched traces back to these twelve pages.

October, 1969 An Axiomatic Basis for Computer Programming C. A. R. Hoare / Communications of the ACM You can prove that code is correct. Not test it — prove it, mathematically. Hoare showed us how, and we've been mostly ignoring it ever since. Every bug we've ever shipped was a choice. This paper described the alternative.

March, 1968 Letters to the Editor: Go To Statement Considered Harmful Edsger W. Dijkstra / Communications of the ACM Not even a paper — a letter. Dijkstra changed how an entire industry wrote code, not by building anything, but by arguing we should stop doing something. Some of the best ideas in engineering are exactly that. I think about this every time I review code.

April, 1965 An Algorithm for the Machine Calculation of Complex Fourier Series James W. Cooley & John W. Tukey / Mathematics of Computation This is the math that I did at JPL/NASA and Caltech. Every MP3, every JPEG, every wireless signal, every radar system runs on this algorithm. Going from O(n²) to O(n log n) sounds like a footnote. For n = 1,000,000, it's the difference between possible and not.

April, 1960 Recursive Functions of Symbolic Expressions and Their Computation by Machine, Part I John McCarthy / Communications of the ACM LISP gave us recursion, garbage collection, and treating code as data. Most developers in 2026 use all three every day and couldn't tell you where they came from. McCarthy just wanted to do AI research. He built a programming language as a side effect.

1959 A Note on Two Problems in Connexion with Graphs Edsger W. Dijkstra / Numerische Mathematik Three pages. The most-used algorithm in the world. Every time your GPS recalculates, every packet that routes across the internet, every game NPC that finds its way — it's Dijkstra. I've implemented this four times in my career and never once stopped to appreciate what I was holding.

October, 1950 Computing Machinery and Intelligence Alan M. Turing / Mind "Can machines think?" is the wrong question, and Turing knew it. That's why he reframed it as a game. In 2026, every AI company on Earth is trying to pass Turing Tests that weren't designed for the AI we've built. We keep answering the question he asked instead of the harder one he was pointing at.

July, 1948 The Transistor, A Semi-Conductor Triode John Bardeen & Walter H. Brattain / Physical Review Without this, the rest of the list doesn't exist. The transistor isn't an improvement on vacuum tubes — it's a different category of thing entirely. Every AI model, every database, every website runs on billions of these. Two pages in Physical Review. The ROI per page is unmatched in the history of science.

July, 1948 A Mathematical Theory of Communication Claude E. Shannon / Bell System Technical Journal Shannon did this twice. In the same year. He defined information itself — what it means to send a message, how much "information" is in a signal, what the theoretical limit of any channel is. We're still discovering new corners of what he built in 1948.

June 30, 1945 First Draft of a Report on the EDVAC John von Neumann / Moore School, University of Pennsylvania Read this. This is the architecture your laptop runs on today. Not inspired by — literally the same architecture. Input, memory, CPU, output. Eighty years and we haven't replaced it. That's either a tribute to von Neumann's genius or an indictment of our collective imagination. Probably both.

1943 A Logical Calculus of the Ideas Immanent in Nervous Activity Warren McCulloch & Walter Pitts / Bulletin of Mathematical Biophysics The paper that invented the neuron as a computational concept. McCulloch was a neuroscientist, Pitts was an 18-year-old runaway with no formal education who had memorized the Principia Mathematica. Together they showed that a network of binary threshold units could compute any logical function — the first theoretical proof that a brain-like system could perform computation. Every artificial neural network in the world traces its intellectual lineage to this 1943 paper.

1938 A Symbolic Analysis of Relay and Switching Circuits Claude E. Shannon / Transactions of the AIEE This is a master's thesis. Shannon was 22. He figured out that Boolean logic and electrical circuits were the same thing — and that insight unlocked every gate in every chip ever made. I've never felt worse about my own graduate work.

November, 1936 On Computable Numbers, with an Application to the Entscheidungsproblem Alan M. Turing / Proceedings of the London Mathematical Society The concept of computation was invented on paper, by a 23-year-old, with a pencil. Every compiler, every database, every LLM runs on the back of this one idea. The truly foundational papers don't just answer a question — they define the question. This is that paper.

1843 Sketch of the Analytical Engine Invented by Charles Babbage — with Notes by Ada Lovelace Ada Lovelace / Taylor's Scientific Memoirs Ada Lovelace wrote the first computer program for a machine that was never completed. There's something beautiful and tragic and deeply human about that. She figured out what the Analytical Engine could do before Babbage finished building it. That's not just a footnote — that's a proof that software is harder to build than hardware, and always has been.