| 1 | Oct 30, '25 | Crawling isn't innate (unlike walking). Every baby must *invent* crawling, from scratch, using extremely little data, and no reference to imitate. Which is why different babies end | 72.3× | 7.6M | 133k | 8.1k | 14k |
| 2 | Dec 20, '24 | Today OpenAI announced o3, its next-gen reasoning model. We've worked with OpenAI to test it on ARC-AGI, and we believe it represents a significant breakthrough in getting AI to ad | 21.3× | 2.2M | 8.5k | 1.5k | 3.1k |
| 3 | Aug 25, '25 | NVIDA chips are manufactured by TSMC, a Taiwanese company. They're created using EUV lithography machines manufactured by ASML, a Dutch company. These machines consist of >50% o | 20.8× | 2.2M | 33k | 2.1k | 5.7k |
| 4 | Dec 15, '23 | Unfortunately , too few people understand the distinction between memorization and understanding. It's not some lofty question like "does the system have an internal world model?", | 19.4× | 2M | 3.7k | 467 | 1.9k |
| 5 | May 16, '23 | It's been mask-off for EM since early 2022 (with many warning signs since 2017) -- so not exactly new, but still pretty sad. Imagine being able to do absolutely anythingv you want | 17.2× | 1.8M | 2.8k | 233 | 52 |
| 6 | Apr 14, '23 | Really striking how every AI hype tweet is explicitly trying to induce FOMO. "You're getting left behind. All your competitors are using this. Everyone else is making more money th | 15× | 1.6M | 4.8k | 585 | 224 |
| 7 | Jul 18, '25 | Today we're releasing a developer preview of our next-gen benchmark, ARC-AGI-3. The goal of this preview, leading up to the full version launch in early 2026, is to collaborate wit | 12.8× | 1.3M | 2.6k | 456 | 285 |
| 8 | Nov 28, '23 | Big news: we just released Keras 3.0! ▶ Run Keras on top of JAX, TensorFlow, and PyTorch ▶ Train faster with XLA compilation ▶ Unlock training runs with any number of devices & | 12.5× | 1.3M | 3k | 637 | 552 |
| 9 | Jun 24, '23 | Search interest over time, ChatGPT (blue) vs Minecraft (red). There's an obvious factor that might well underlie both trends (down for ChatGPT, up for Minecraft). Can you guess wha | 12.2× | 1.3M | 2.4k | 478 | 631 |
| 10 | Sep 9, '25 | A student who truly understands F=ma can solve more novel problems than a Transformer that has memorized every physics textbook ever written. | 11.1× | 1.2M | 3.7k | 294 | 783 |
| 11 | Mar 21, '18 | The problem with Facebook is not *just* the loss of your privacy and the fact that it can be used as a totalitarian panopticon. The more worrying issue, in my opinion, is its use o | 10.4× | 0 | 14k | 7.4k | 936 |
| 12 | Nov 4, '21 | View post on X | 10.4× | 0 | 28k | 3.8k | 345 |
| 13 | Jun 19, '23 | Here's an interesting anecdote on that topic. In 1979, a group of people who were "just asking questions" offered a $50,000 reward ($220,000 in 2023 dollars) to anyone who could pr | 10.3× | 1.1M | 2.6k | 429 | 294 |
| 14 | Sep 18, '25 | The 3rd edition of my book Deep Learning with Python is being printed right now, and will be in bookstores within 2 weeks. You can order it now from Amazon or from Manning. This ti | 9.7× | 1M | 6.4k | 1k | 3.8k |
| 15 | Jan 21, '23 | What people don't realize is that this is the default state of housing in a market economy, and it's great for consumers. Turning aging homes into an investment vehicle by artifici | 9.7× | 1M | 6.3k | 703 | 616 |
| 16 | Jul 11, '23 | We're launching Keras Core, a new library that brings the Keras API to JAX and PyTorch in addition to TensorFlow. It enables you to write cross-framework deep learning components a | 9.2× | 968k | 3.8k | 760 | 498 |
| 17 | Dec 20, '24 | One very important thing to understand about the future: the economics of AI are about to change completely. We'll soon be in a world where you can turn test-time compute into comp | 9.2× | 964k | 4.1k | 392 | 1.5k |
| 18 | Jul 25, '23 | GitHub Copilot and LLM chatbots are causing a big drop in StackOverflow traffic. What's fascinating about it is that these systems work largely thanks to the user-generated content | 9.1× | 957k | 4.3k | 600 | 482 |
| 19 | Aug 21, '25 | LLM adoption among US workers is closing in on 50%. Meanwhile labor productivity growth is lower than in 2020. Many counter-arguments can be made here, e.g. "they don't know yet ho | 8.9× | 928k | 4.5k | 545 | 1.4k |
| 20 | Dec 20, '24 | TikTok just open-sourced their recommender system framework (Monolith) -- and it uses Keras This means that nearly all the major recommender systems in the industry are built on Ke | 8.8× | 921k | 10k | 831 | 6.1k |
| 21 | Aug 7, '25 | Grok 4 is still state-of-the-art on ARC-AGI-2 among frontier models. 15.9% for Grok 4 vs 9.9% for GPT-5. | 8.3× | 869k | 2.5k | 377 | 211 |
| 22 | Aug 6, '23 | YYYY-MM-DD is obviously the best date format: it's the one where alphanumerical order matches chronological order. | 8.2× | 854k | 8.3k | 507 | 189 |
| 23 | May 4 | I wrote Deep Learning with Python to be the definitive guide to how deep learning works and how to best make use of it. Tens of thousands of people got their career start via this | 8.0× | 835k | 4.1k | 640 | 3.8k |
| 24 | Nov 6, '23 | Turns out, fitting a curve to a dataset produces a model that only generalizes to that specific data distribution -- big if true. | 7.8× | 813k | 3.7k | 377 | 665 |
| 25 | Oct 3, '23 | My interpretation of prompt engineering is this: 1. A LLM is a repository of many (millions) of vector programs mined from human-generated data, learned implicitly as a by-product | 6.8× | 711k | 2.9k | 505 | 1.9k |
| 26 | Oct 21, '24 | People have been rewriting history and saying that "everyone has always believed that LLMs alone wouldn't be AGI and that extensive scaffolding around them would be necessary". No, | 6.6× | 695k | 2.7k | 285 | 544 |
| 27 | Aug 13, '25 | GenAI isn't just a technology; it's an informational pollutant—a pervasive cognitive smog that touches and corrupts every aspect of the Internet. It's not just a productivity tool; | 6.5× | 681k | 6.4k | 942 | 2.2k |
| 28 | Jun 2, '23 | To be clear, at this time and for the foreseeable future, there does not exist any AI model or technique that could represent an extinction risk for humanity. Not even in nascent f | 6.1× | 636k | 2.7k | 536 | 182 |
| 29 | Jan 1, '24 | We need a "where are they now" series about the developers who claimed in February-March 2023 that they had become "100x" more productive thanks to LLMs. Surely by now they must ha | 6.0× | 626k | 3.7k | 404 | 355 |
| 30 | Sep 29, '24 | We haven't needed programmers since the advent of automated programming in the 1970s. Once you have a high level language like COBOL and a compiler to automatically turn it into ac | 5.9× | 616k | 5.6k | 433 | 635 |
| 31 | Jun 11, '24 | I'm partnering with @mikeknoop to launch ARC Prize: a $1,000,000 competition to create an AI that can adapt to novelty and solve simple reasoning problems. Let's get back on track | 5.8× | 603k | 2.5k | 513 | 1k |
| 32 | Dec 21, '22 | Twitter is great because it gives you unfiltered visibility into the thoughts and decision-making process of some of the most successful people in the world. This has completely cu | 5.2× | 548k | 6k | 736 | 218 |
| 33 | Mar 12, '24 | If you could fully automate software engineering (my job), I think that would be great, since I could then move on to higher-leverage things. Making software is a means to an end, | 5.1× | 539k | 2.7k | 284 | 983 |
| 34 | Jul 10, '25 | In my personal experience with Duolingo, I found that it isn't an education app, it's a mobile game -- it's completely ineffective at teaching languages but it's quite entertaining | 4.9× | 510k | 5.9k | 220 | 903 |
| 35 | Sep 30, '23 | The main downside of working in AI is having to constantly fight off assassination attempts from random time travelers | 4.5× | 471k | 3.6k | 451 | 130 |
| 36 | Jan 1, '25 | I believe that if you understood intelligence, you could build it in software form on a $1M budget, including training. It would not need to be trained on the entire Internet, nor | 4.5× | 471k | 3.6k | 232 | 853 |
| 37 | Jan 24, '23 | Been working with Python for 13 years and I still occasionally end up with a hopelessly borked environment where I have to actually nuke and reinstall the Python interpreter. And y | 4.4× | 460k | 2.5k | 95 | 169 |
| 38 | Mar 2, '25 | Twitter used to be my favorite place on the Internet. I've derived enormous value from it in the past 16 years. Not true anymore. Most of the people I enjoyed reading have left. My | 4.4× | 457k | 6.8k | 394 | 563 |
| 39 | Sep 15, '25 | To solve any problem, you don't have to be super smart. You just have to 1) be able to break down problems into subproblems, 2) be slightly smarter than the hardest of the atomic s | 4.2× | 439k | 5.4k | 488 | 1.6k |
| 40 | Sep 4, '24 | Computing used to feel fast -- everything ran locally, software was mostly in C/C++ and was kept in check by the need to run on all kinds of old hardware. Now any one of my Chrome | 4.1× | 428k | 6.5k | 368 | 569 |
| 41 | Nov 14, '24 | Some personal news -- I'm leaving Google to go start a new company with a friend. More to be announced soon! I will stay deeply involved with the Keras project from the outside -- | 4.0× | 417k | 4.8k | 244 | 589 |
| 42 | Jan 5, '25 | I kinda miss when "AI Twitter" was folks doing AI research or app development, posting under their own identities. Slightly more intellectual depth than anime avatars whose main AI | 4.0× | 416k | 4.1k | 206 | 297 |
| 43 | Sep 28, '25 | The idea that we will automate work by building artificial versions of ourselves to do exactly the things we were previously doing, rather than redesigning our old workflows to mak | 4.0× | 414k | 5.6k | 547 | 889 |
| 44 | May 10, '24 | One thing that even relatively senior ML people often fail to grasp is that deep learning models are curves fitted to a data distribution. You cannot expect them to solve tasks out | 3.9× | 412k | 2.7k | 411 | 1.1k |
| 45 | Mar 21, '23 | Asking if it's still worth learning to code in the age of AI is like asking if it's still worth learning to write in the age of the printing press. It's even more valuable than bef | 3.9× | 406k | 3.2k | 533 | 311 |
| 46 | Oct 4, '24 | Not too many people realize this, but my actual job is just writing (and reviewing) a lot of code. I've shipped about 45k lines of code per year on average since 2015. Writing code | 3.8× | 399k | 4.2k | 144 | 578 |
| 47 | Dec 22, '24 | The most reliably predictable trend of the next 100 years: every year, humanity will use significantly more computing power than the previous year. Someone should start an ETF base | 3.8× | 395k | 2.6k | 191 | 1.2k |
| 48 | Oct 12, '24 | LLMs are trained on much more than the whole Internet -- they also consume handcrafted answers produced by armies of highly qualified data annotators (often domain experts). Today | 3.6× | 376k | 2.6k | 292 | 822 |
| 49 | May 3, '21 | Within 10-20 years, nearly every branch of science will be, for all intents and purposes, a branch of computer science. Computational physics, comp chemistry, comp biology, comp me | 3.5× | 0 | 6.5k | 1.1k | 954 |
| 50 | Apr 12, '25 | Google has never been good at marketing its products, but boy is it good at technology. Very bullish on Gemini, the Gemma line of open models, and the JAX/Keras/TPU stack | 3.3× | 342k | 3.1k | 161 | 431 |
| 51 | Sep 17, '25 | Software engineers shouldn't fear being replaced by AI. They should fear being asked to maintain the sprawling mess of AI-generated legacy code their employer's systems will soon r | 3.2× | 333k | 8.1k | 909 | 740 |
| 52 | Oct 25, '21 | Announcement: my book Deep Learning with Python (2nd edition) has been released. 500 pages of code examples, theory, context, practical tips... If you want to really understand how | 2.9× | 0 | 5.2k | 846 | 1.4k |
| 53 | Sep 7, '24 | In science and engineering, I feel like having a sense of aesthetics is dramatically more useful than being marginally more intelligent | 2.7× | 279k | 4k | 324 | 660 |
| 54 | Oct 30, '25 | @guy_dar1 There is no AI approach that can do what these babies do, not even close. The sample efficiency gap (at equal space dimensionality) is 3-4 OOM. So if the bar is "baby lev | 2.7× | 279k | 9k | 58 | 93 |
| 55 | Nov 6, '19 | I've just released a fairly lengthy paper on defining & measuring intelligence, as well as a new AI evaluation dataset, the "Abstraction and Reasoning Corpus". I've been workin | 2.6× | 0 | 4.5k | 1.3k | 748 |
| 56 | Nov 3, '25 | Programming isn't like playing an instrument or speaking a language: whether you started coding at 5 or 20 makes no difference to your skill at 30. Programming is about clarity of | 2.5× | 264k | 5.9k | 423 | 1.3k |
| 57 | Nov 5, '23 | You never really learn git, you just get used to it. | 2.4× | 256k | 2.8k | 193 | 96 |
| 58 | Jan 27, '21 | People who say $GME is overvalued don't understand the value proposition. It's no longer a stock, it's a store of value. It's the new digital gold. And for now it has only 3% of th | 2.4× | 0 | 6.1k | 607 | 138 |
| 59 | Jul 19, '24 | Imagine writing a bit of code at your dull software engineer job job and messing up so bad the president has to be briefed about it | 2.4× | 253k | 2.8k | 205 | 142 |
| 60 | Apr 25, '18 | 🌉 SF bay area: area: 18,040 km2 population: 7.8M density: 425.7/km2 1BR rent: $3.3k 🗼 Tokyo bay area: area: 14,034 km2 (22% less) population: 37.8M (385% more (4.85x)) density: 2 | 2.4× | 0 | 4.9k | 1.4k | 98 |
| 61 | Oct 30, '25 | The bottleneck for deep skill isn't usually intelligence, but boredom tolerance. Learning has an activation energy: below a certain skill threshold, practice is tedious, but above | 2.3× | 236k | 6.5k | 823 | 2.3k |
| 62 | Feb 1, '22 | Software is this weird space where you can spend basically nothing and create a billion dollars of value, or spend a billion dollars and create basically no value. | 2.2× | 0 | 5.7k | 782 | 202 |
| 63 | Aug 25, '25 | The most interesting fact about this globalized chain is the number of irreplaceable single points of failure. There's only one company that can make these chips at scale. It runs | 2.2× | 228k | 11k | 457 | 514 |
| 64 | Mar 11, '19 | Are you a deep learning researcher? Wondering if all this TensorFlow 2.0 stuff you heard about is relevant to you? This thread is a crash course on everything you need to know to u | 2.0× | 0 | 3.1k | 1.1k | 752 |
| 65 | Jul 16, '20 | "TensorFlow, Keras and Deep Learning (without a PhD)" by @martin_gorner is a clear, approachable, fun introduction to neural networks. Highly recommended! | 2.0× | 0 | 3.4k | 801 | 1k |
| 66 | Sep 13, '25 | The most important skill for a researcher is not technical ability. It's taste. The ability to identify interesting and tractable problems, and recognize important ideas when they | 2.0× | 211k | 3.7k | 500 | 981 |
| 67 | Oct 5, '19 | I just wrote an extensive TensorFlow 2.0 + Keras overview, targeted at deep learning researchers: Hope you will find it useful! Let me know if you have any feedback. | 2.0× | 0 | 3.5k | 961 | 733 |
| 68 | Aug 27, '25 | Saying that deep learning is "just a bunch of matrix multiplications" is about as informative as saying that computers are "just a bunch of transistors" or that a library is "just | 2.0× | 208k | 2.9k | 227 | 602 |
| 69 | Sep 24, '22 | Human intelligence is a poor metaphor for what "AI" is doing. AI displays essentially none of the properties of human cognition, and in reverse, most of the useful properties of mo | 1.9× | 0 | 4k | 700 | 490 |
| 70 | Jul 5, '24 | The main benefit of having extensive programming experience isn't the ability to write software (which is a bit of a commodity -- you could hire someone to do it), it's how it chan | 1.9× | 196k | 2.4k | 273 | 428 |
| 71 | May 29, '25 | The secret to life satisfaction is agency maxxing | 1.8× | 193k | 2.5k | 229 | 652 |
| 72 | Dec 8, '22 | Python is usually an elegant language. But `if __name__ == "__main__"` is one exception. | 1.8× | 0 | 5.4k | 269 | 143 |
| 73 | Nov 24, '24 | The real problem with EU cookie banners is that they're still around after several years of achieving nothing but being an annoyance. Mistakes happen. The key is to identify them a | 1.8× | 192k | 2.5k | 205 | 91 |
| 74 | Jul 19, '25 | Intelligence isn't a collection of skills. It's the efficiency with which you acquire and deploy new skills. It's an efficiency ratio. And that's why benchmark scores can be very m | 1.8× | 190k | 2.6k | 287 | 575 |
| 75 | Jul 23, '22 | I've been asked about this a lot, so let me provide a quick FAQ. Q: What's the nature of the issue? A: Anyone who has bought my book from Amazon in the past few month hasn't bought | 1.8× | 0 | 2.7k | 1k | 312 |
| 76 | Jul 21, '24 | An underappreciated aspect of productivity: momentum. The more stuff you do the easier it becomes to do stuff (literally anything). And inversely. | 1.8× | 190k | 2.9k | 301 | 672 |
| 77 | Apr 28, '23 | The most powerful debugging tool: walking away from your computer. | 1.8× | 187k | 2.5k | 331 | 69 |
| 78 | Jun 5, '19 | You don't need to know everything. You don't really need a formal background in this or that -- though it helps. You don't even need a PhD. You do, however, need to be constantly l | 1.7× | 0 | 3.3k | 1k | 127 |
| 79 | Oct 28, '25 | To really understand a concept, you have to "invent" it yourself in some capacity. Understanding doesn't come from passive content consumption. It is always self-built. It is an ac | 1.6× | 165k | 3.8k | 532 | 1.1k |
| 80 | Jan 7, '20 | Our field isn't quite "artificial intelligence" -- it's "cognitive automation": the encoding and operationalization of human-generated abstractions / behaviors / skills. The "intel | 1.6× | 0 | 2.9k | 786 | 326 |
| 81 | Apr 6, '19 | Learning: 1) Make something. It sucks. 2) Realize it sucks. 3) Actually care about it. 4) Find ways to improve. 5) Try again; repeat forever. A lot of people get stuck at either 2) | 1.5× | 0 | 3.1k | 902 | 157 |
| 82 | Oct 30, '22 | Software companies aren't made of code. They're made of processes that produce and maintain code. And the foremost component of these processes are people. The code is just a by-pr | 1.5× | 0 | 3.4k | 534 | 333 |
| 83 | Mar 21, '18 | If Facebook gets to decide, over the span of many years, which news you will see (real or fake), whose political status updates you’ll see, and who will see yours, then Facebook is | 1.5× | 0 | 2.7k | 1.1k | 10 |
| 84 | Aug 26, '22 | Deep learning takes data points and turns them into a query-able structure that enables retrieval and interpolation between the points. You could think of it as a continuous genera | 1.5× | 0 | 2.6k | 415 | 715 |
| 85 | Jan 12, '18 | "Neural networks" are a sad misnomer. They're neither neural nor even networks. They're chains of differentiable, parameterized geometric functions, trained with gradient descent ( | 1.4× | 0 | 2.7k | 984 | 13 |
| 86 | Mar 21, '18 | We’re looking at a powerful entity that builds fine-grained psychological profiles of over two billion humans, that runs large-scale behavior manipulation experiments, and that aim | 1.3× | 0 | 2.4k | 885 | 19 |
| 87 | May 1, '21 | My wife and I are thrilled to announce the birth of our son Sylvain, earlier this week. Sylvain and his mom are both doing great. Feeling so blessed :) | 1.3× | 0 | 4k | 45 | 11 |
| 88 | Sep 7, '22 | I'll tell you a secret about having more impact. You've got to focus on fewer things, ideally just one -- your most ambitious idea. And you've got to commit and finish it. | 1.3× | 0 | 2.6k | 372 | 461 |
| 89 | Mar 21, '18 | If you work in AI, please don't help them. Don't play their game. Don't participate in their research ecosystem. Please show some conscience | 1.2× | 0 | 2.7k | 548 | 21 |
| 90 | Oct 28, '25 | You rarely solve hard problems in a flash of insight. It's more typically a slow, careful process of exploring a branching tree of possibilities. You must pause, backtrack, and wei | 1.2× | 127k | 2.8k | 307 | 1k |
| 91 | Nov 23, '22 | Many people in tech have so little exposure to philosophy (or the humanities in general) that when they get exposed to old ideas like Plato's cavern or the simulation hypothesis, t | 1.2× | 0 | 2.7k | 307 | 217 |
| 92 | May 7, '24 | My wife and I are overjoyed to announce the birth of our daughter Diane, a few days ago. Welcome to the world, little Diane! We're feeling so blessed to have you and your big broth | 1.2× | 126k | 3.2k | 26 | 33 |
| 93 | Apr 25, '21 | I'd guess I'd summarize it as, being a "deep learning expert" in 2021 is like being a "medicine expert" in 1800. You know a lot less than you think, and most of what you think you | 1.1× | 0 | 2.8k | 398 | 124 |
| 94 | Nov 29, '20 | Deep learning isn't a science, but rather an ever-changing set of empirically-derived engineering best practices, woven together by over-claiming, unreliable narratives. | 1.1× | 0 | 2.6k | 353 | 154 |
| 95 | Feb 24, '20 | "No code" is just the beginning. I'm looking forward to no software, no computer, just living in the woods | 1.1× | 0 | 2.7k | 384 | 41 |
| 96 | Jun 14, '22 | Mass deployment of self-driving cars was less than 5 years away in 2016. Then in 2021 it was 2-3 years away. In 2024 it will be 1 year away. Zeno's AI milestone. | 1.1× | 0 | 3.1k | 158 | 38 |
| 97 | May 16, '21 | Wait, what | 1.0× | 0 | 2.8k | 308 | 39 |
| 98 | May 21, '21 | I would rather press [▲] 50 times than actually type a command in the terminal | 0.9× | 0 | 2.5k | 208 | 28 |
@karpathy
@levelsio
@paulg
@naval
@marclou
@illyism