{"slug": "pytorch", "name": "PyTorch", "category": "dev-libs", "category_name": "Libraries and frameworks", "description": "Open-source Python deep learning framework providing tensors, automatic differentiation, and GPU acceleration.", "recommenders": 5, "commenters": 15, "warners": 1, "score": 5.359, "rising": -1.64, "first": "2025-11-07", "latest": "2026-09-21", "why": "Commenters use it daily for ML work and like its level of abstraction, and it is seen as the default framework for sharing and reproducing ML results.", "why_is_quote": false, "has_verdict": true, "sentiment": {"recommend": 6, "neutral": 7, "compare": 3, "warn": 1}, "spark": [2, 2, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0], "listed": true, "loved": ["Daily workhorse", "Ecosystem default"], "caveat": "Python lock-in", "short_quote": "I use pytorch daily, it has made it possible for me to play with ideas that I would have only dreamed of.", "icon": true, "domain": null, "github_repo": null, "official_url": "https://pytorch.org", "pricing": "open source", "aliases": ["PyTorch"], "verdict": [{"text": "Commenters use it daily for ML work", "cites": [{"id": 782973664, "item_id": 48935854, "url": "https://news.ycombinator.com/item?id=48935854", "by": "andy_ppp", "date": "2026-07-16", "stance": "recommend", "quote": "have built many projects with PyTorch and use it daily for all my work", "excerpt": "I love AI and have built many projects with PyTorch and use it daily for all my work and am building a startup in my spare time here: https://veloa.com/ Being concerned about how AI will concentrate power doesn\u2019t mean I don\u2019t think it\u2019s extremely useful.", "context": "building a startup in spare time", "use_case": "building AI projects", "switched_from": "", "story": "Generative AI Is an Engineering Disaster"}, {"id": 733562064, "item_id": 45847629, "url": "https://news.ycombinator.com/item?id=45847629", "by": "jmward01", "date": "2025-11-07", "stance": "recommend", "quote": "I use pytorch daily, it has made it possible for me to play with ideas that I would have only dreamed of.", "excerpt": "All I can say is 'thanks!'. It does take a team, and a community, but individuals matter. I use pytorch daily, it has made it possible for me to play with ideas that I would have only dreamed of. It is a big part of my life so, thanks for your contributions and best of luck on the next thing!", "context": "", "use_case": "ML experimentation", "switched_from": "", "story": "Leaving Meta and PyTorch"}, {"id": 737775232, "item_id": 46110952, "url": "https://news.ycombinator.com/item?id=46110952", "by": "Grosvenor", "date": "2025-12-01", "stance": "neutral", "quote": "I use the normal stacks you'd expect. Python, Pytorch, Spark/ray, Jupyter/Merimo, AWS, Postgres, Mathematica", "excerpt": ".-~~\\ / \\ _ ~x .-~_)_ ~x\".-~ ~-. _ ( / \\ _ || T o o Y || ==:l l < ! I;== \\\\ \\ .__/ / // \\\\ ,r\"-,___.-'r.// }^ \\.( ) _.'//. / }~Xi--~ // \\ Y Y I\\ \\ \" Y | | |o\\ \\ | | l_l Y T | l \"o l_j ! \\ / ___,.---^. o .^---.._____ \"~~~ \" ~ ~~~\" SEEKING WORK - Data scientist, consulting & fractional leadership, US/remote worldwide, email in profile. All I want for christmas is some gnarly problems to chew on, otherwise it's coal for", "context": "data scientist with 20+ years experience", "use_case": "data science consulting", "switched_from": "", "story": "Ask HN: Who wants to be hired? (December 2025)"}]}, {"text": " and like its level of abstraction,", "cites": [{"id": 733532016, "item_id": 45845751, "url": "https://news.ycombinator.com/item?id=45845751", "by": "stared", "date": "2025-11-07", "stance": "recommend", "quote": "the more it developed, the better it was, with (at least to my mind) the perfect level of abstraction.", "excerpt": "In 2018, I co-wrote a blog post with the inflammatory title \u201cDon\u2019t use TensorFlow, try PyTorch instead\u201d (https://news.ycombinator.com/item?id=17415321). As it gained traction here, it was changed to \u201cKeras vs PyTorch\u201d (some edgy things that work for a private blog are not good for a corporate one). Yet the initial title stuck, and you can see it resonated well with the crowd. TensorFlow (while a huge step on top of T", "context": "deep learning practitioner, co-wrote 2018 PyTorch vs TensorFlow post", "use_case": "deep learning", "switched_from": "Keras", "story": "Leaving Meta and PyTorch"}, {"id": 740585008, "item_id": 46286563, "url": "https://news.ycombinator.com/item?id=46286563", "by": "stared", "date": "2025-12-16", "stance": "recommend", "quote": "When it comes to matrix notation, I prefer PyTorch over NumPy", "excerpt": "On the contrary, I think that well-designed general-purpose languages beat domain-specific languages. Even in the example given, in NumPy you can use np.array, but to make a fair comparison, use np.matrix. import numpy as np X = np.matrix([1, 2, 3]) Y = np.matrix([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) Z = Y * X.T W = np.hstack([Z, Z]) That way, we can extend our languages. If np.matrix is \"too many keystrokes\", it can be", "context": "", "use_case": "tensor math", "switched_from": "", "story": "In Defense of Matlab Code"}]}, {"text": " and it is seen as the default framework for sharing and reproducing ML results.", "cites": [{"id": 761259952, "item_id": 47578747, "url": "https://news.ycombinator.com/item?id=47578747", "by": "kadushka", "date": "2026-03-30", "stance": "recommend", "quote": "use good old Pytorch", "excerpt": "1. If you have good results on sufficiently large models (check latest papers re: which benchmarks are still relevant), post them on Github, along will detailed instructions how to reproduce. 2. Post the link to the GH repo in \"Show HN\" section. 3. If results are solid, write up a paper and upload it to arxiv. Next step would be try to publish in an ML conference. p.s. To increase your chances of anyone actually clic", "context": "", "use_case": "ML research code", "switched_from": "", "story": "Ask HN: Release Path for 'Transformers Alternatives'?"}, {"id": 742124898, "item_id": 46382806, "url": "https://news.ycombinator.com/item?id=46382806", "by": "RomanPushkin", "date": "2025-12-24", "stance": "recommend", "quote": "Python has won for me: fastapi, pytorch, langchain, streamlit, and so on and on.", "excerpt": "I'm happy to see v4.0, but 2025 was the year I switched from Ruby to Python after gradually drifting back to it more and more. The tipping point was when I had Claude Code automatically convert one of my Ruby projects to 100% Python - and after that, I just had no Ruby left. I spent over a decade enjoying Ruby and even wrote a book about it. At this point, though, Python has won for me: fastapi, pytorch, langchain, s", "context": "Ruby developer for over a decade, wrote a book on Ruby", "use_case": "", "switched_from": "", "story": "Ruby 4.0.0 Released"}]}, {"text": " The main caveat raised is that a hard PyTorch requirement shuts out developers working in other languages, such as C++.", "cites": [{"id": 796538369, "item_id": 49783648, "url": "https://news.ycombinator.com/item?id=49783648", "by": "cookiengineer", "date": "2026-09-21", "stance": "warn", "quote": "because of the hard pytorch requirements", "excerpt": "OMG! Just yesterday I published my reworked GoNEAT library that implements HyperNEAT combined with phased search and backpropagation [1]. But it's kind of impossible to enter for me because of the hard pytorch requirements :( would love to see the project as a gym, so that you can run your own ANN design algorithm. I get that most data science students still use python, but the evolutionary world is kinda in C++ and ", "context": "", "use_case": "", "switched_from": "", "story": "Show HN: A competition for small neural networks that play strategy games"}]}], "praise": [{"label": "Daily workhorse", "summary": "Commenters use it every day and credit it with making ambitious ideas feasible.", "quotes": [{"id": 733562064, "item_id": 45847629, "url": "https://news.ycombinator.com/item?id=45847629", "by": "jmward01", "date": "2025-11-07", "stance": "recommend", "quote": "I use pytorch daily, it has made it possible for me to play with ideas that I would have only dreamed of.", "excerpt": "All I can say is 'thanks!'. It does take a team, and a community, but individuals matter. I use pytorch daily, it has made it possible for me to play with ideas that I would have only dreamed of. It is a big part of my life so, thanks for your contributions and best of luck on the next thing!", "context": "", "use_case": "ML experimentation", "switched_from": "", "story": "Leaving Meta and PyTorch"}, {"id": 782973664, "item_id": 48935854, "url": "https://news.ycombinator.com/item?id=48935854", "by": "andy_ppp", "date": "2026-07-16", "stance": "recommend", "quote": "have built many projects with PyTorch and use it daily for all my work", "excerpt": "I love AI and have built many projects with PyTorch and use it daily for all my work and am building a startup in my spare time here: https://veloa.com/ Being concerned about how AI will concentrate power doesn\u2019t mean I don\u2019t think it\u2019s extremely useful.", "context": "building a startup in spare time", "use_case": "building AI projects", "switched_from": "", "story": "Generative AI Is an Engineering Disaster"}, {"id": 737775232, "item_id": 46110952, "url": "https://news.ycombinator.com/item?id=46110952", "by": "Grosvenor", "date": "2025-12-01", "stance": "neutral", "quote": "I use the normal stacks you'd expect. Python, Pytorch, Spark/ray, Jupyter/Merimo, AWS, Postgres, Mathematica", "excerpt": ".-~~\\ / \\ _ ~x .-~_)_ ~x\".-~ ~-. _ ( / \\ _ || T o o Y || ==:l l < ! I;== \\\\ \\ .__/ / // \\\\ ,r\"-,___.-'r.// }^ \\.( ) _.'//. / }~Xi--~ // \\ Y Y I\\ \\ \" Y | | |o\\ \\ | | l_l Y T | l \"o l_j ! \\ / ___,.---^. o .^---.._____ \"~~~ \" ~ ~~~\" SEEKING WORK - Data scientist, consulting & fractional leadership, US/remote worldwide, email in profile. All I want for christmas is some gnarly problems to chew on, otherwise it's coal for", "context": "data scientist with 20+ years experience", "use_case": "data science consulting", "switched_from": "", "story": "Ask HN: Who wants to be hired? (December 2025)"}]}, {"label": "Ecosystem default", "summary": "Seen as the standard ML framework, recommended for sharing research and listed in job requirements.", "quotes": [{"id": 761259952, "item_id": 47578747, "url": "https://news.ycombinator.com/item?id=47578747", "by": "kadushka", "date": "2026-03-30", "stance": "recommend", "quote": "use good old Pytorch", "excerpt": "1. If you have good results on sufficiently large models (check latest papers re: which benchmarks are still relevant), post them on Github, along will detailed instructions how to reproduce. 2. Post the link to the GH repo in \"Show HN\" section. 3. If results are solid, write up a paper and upload it to arxiv. Next step would be try to publish in an ML conference. p.s. To increase your chances of anyone actually clic", "context": "", "use_case": "ML research code", "switched_from": "", "story": "Ask HN: Release Path for 'Transformers Alternatives'?"}, {"id": 742124898, "item_id": 46382806, "url": "https://news.ycombinator.com/item?id=46382806", "by": "RomanPushkin", "date": "2025-12-24", "stance": "recommend", "quote": "Python has won for me: fastapi, pytorch, langchain, streamlit, and so on and on.", "excerpt": "I'm happy to see v4.0, but 2025 was the year I switched from Ruby to Python after gradually drifting back to it more and more. The tipping point was when I had Claude Code automatically convert one of my Ruby projects to 100% Python - and after that, I just had no Ruby left. I spent over a decade enjoying Ruby and even wrote a book about it. At this point, though, Python has won for me: fastapi, pytorch, langchain, s", "context": "Ruby developer for over a decade, wrote a book on Ruby", "use_case": "", "switched_from": "", "story": "Ruby 4.0.0 Released"}, {"id": 739102129, "item_id": 46193883, "url": "https://news.ycombinator.com/item?id=46193883", "by": "konz", "date": "2025-12-08", "stance": "neutral", "quote": "a range of ML frameworks such as TensorFlow, PyTorch, or HuggingFace Transformers", "excerpt": "ML6 | Machine Learning Engineer, Data Engineer | Python, TensorFlow, PyTorch, GCP, AWS, Azure | Full-time | Amsterdam, Berlin, Munich, Eindhoven, Ghent (EU) On-site/hybrid We are a Machine Learning consulting company that builds end-to-end Machine Learning solutions. By applying the latest AI research, we keep our clients at the forefront of innovation. If you are interested check out: https://www.ml6.eu/knowledge-hu", "context": "hiring at ML6", "use_case": "machine learning projects", "switched_from": "", "story": "Ask HN: Who is hiring? (December 2025)"}]}, {"label": "Reference implementation", "summary": "Used as the baseline when validating or reimplementing models and losses.", "quotes": [{"id": 747083764, "item_id": 46692735, "url": "https://news.ycombinator.com/item?id=46692735", "by": "fotcorn", "date": "2026-01-20", "stance": "neutral", "quote": "validated it against the PyTorch version", "excerpt": "I used Claude Opus 4.5 inside Cursor to write RISC-V Vector/SIMD code. Specifically Depthwise Convolution and normal Convolution layers for a CNN. I started out by letting it write a naive C version without intrinsic, and validated it against the PyTorch version. Then I asked it (and two other models, Gemini 3.0 and GPT 5.1) to come up with some ideas on how to make it faster using SIMD vector instructions and write ", "context": "", "use_case": "reference implementation", "switched_from": "", "story": "Ask HN: Do you have any evidence that agentic coding works?"}, {"id": 788114913, "item_id": 49257182, "url": "https://news.ycombinator.com/item?id=49257182", "by": "ikergarcia1996", "date": "2026-08-11", "stance": "compare", "quote": "The loss is mathematically equivalent to the standard KL-divergence loss available in PyTorch.", "excerpt": "Hello everyone! I\u2019ve been working on an efficient implementation of the KL-divergence loss that reduces its VRAM usage from quadratic to linear, using an approach very similar to Flash Attention: chunking and fusing the forward and backward passes. The loss is mathematically equivalent to the standard KL-divergence loss available in PyTorch. Until now, Knowledge Distillation required a huge amount of VRAM and was pra", "context": "", "use_case": "KL loss", "switched_from": "", "story": "Chunked KL loss, running Knowledge Distillation locally in less <6GB VRAM"}, {"id": 741753040, "item_id": 46359565, "url": "https://news.ycombinator.com/item?id=46359565", "by": "libraryofbabel", "date": "2025-12-22", "stance": "neutral", "quote": "implementing a transformer in PyTorch just to refresh my memory", "excerpt": "I read this article back when I was learning the basics of transformers; the visualizations were really helpful. Although in retrospect knowing how a transformer works wasn't very useful at all in my day job applying LLMs, except as a sort of deep background for reassurance that I had some idea of how the big black box producing the tokens was put together, and to give me the mathematical basis for things like contex", "context": "applies LLMs in day job", "use_case": "implementing a transformer", "switched_from": "", "story": "The Illustrated Transformer"}]}, {"label": "Clean abstraction", "summary": "Praised for the right level of abstraction compared to TensorFlow, and for matrix notation compared to NumPy.", "quotes": [{"id": 733532016, "item_id": 45845751, "url": "https://news.ycombinator.com/item?id=45845751", "by": "stared", "date": "2025-11-07", "stance": "recommend", "quote": "the more it developed, the better it was, with (at least to my mind) the perfect level of abstraction.", "excerpt": "In 2018, I co-wrote a blog post with the inflammatory title \u201cDon\u2019t use TensorFlow, try PyTorch instead\u201d (https://news.ycombinator.com/item?id=17415321). As it gained traction here, it was changed to \u201cKeras vs PyTorch\u201d (some edgy things that work for a private blog are not good for a corporate one). Yet the initial title stuck, and you can see it resonated well with the crowd. TensorFlow (while a huge step on top of T", "context": "deep learning practitioner, co-wrote 2018 PyTorch vs TensorFlow post", "use_case": "deep learning", "switched_from": "Keras", "story": "Leaving Meta and PyTorch"}, {"id": 740585008, "item_id": 46286563, "url": "https://news.ycombinator.com/item?id=46286563", "by": "stared", "date": "2025-12-16", "stance": "recommend", "quote": "When it comes to matrix notation, I prefer PyTorch over NumPy", "excerpt": "On the contrary, I think that well-designed general-purpose languages beat domain-specific languages. Even in the example given, in NumPy you can use np.array, but to make a fair comparison, use np.matrix. import numpy as np X = np.matrix([1, 2, 3]) Y = np.matrix([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) Z = Y * X.T W = np.hstack([Z, Z]) That way, we can extend our languages. If np.matrix is \"too many keystrokes\", it can be", "context": "", "use_case": "tensor math", "switched_from": "", "story": "In Defense of Matlab Code"}]}, {"label": "Performance", "summary": "A from-scratch trainer using the same cuBLAS backend ran 1.5-2.5x slower than PyTorch.", "quotes": [{"id": 779489922, "item_id": 48718120, "url": "https://news.ycombinator.com/item?id=48718120", "by": "vforno", "date": "2026-06-29", "stance": "compare", "quote": "Compared to PyTorch, it's a bit slower (probably 1.5-2.5x)", "excerpt": "Hi, in nanoeuler I use cuBLAS (NVIDIA's super-optimized library) for all matrix multiplications, with the tensor cores in TF32 mode. It's the same thing PyTorch uses underneath, so it's very fast. What I've optimized (and will improve even more) and written by hand are the kernels for the other parts (like FlashAttention, which gave a nice 3x speedup), while I've delegated the large matrices to cuBLAS. Training the 1", "context": "", "use_case": "ML training baseline", "switched_from": "", "story": "Show HN: NanoEuler \u2013 GPT-2 scale model in pure C/CUDA from scratch"}]}], "caveats": [{"label": "Python lock-in", "summary": "A hard PyTorch requirement can exclude projects built in native languages like C++ or Go.", "quotes": [{"id": 796538369, "item_id": 49783648, "url": "https://news.ycombinator.com/item?id=49783648", "by": "cookiengineer", "date": "2026-09-21", "stance": "warn", "quote": "because of the hard pytorch requirements", "excerpt": "OMG! Just yesterday I published my reworked GoNEAT library that implements HyperNEAT combined with phased search and backpropagation [1]. But it's kind of impossible to enter for me because of the hard pytorch requirements :( would love to see the project as a gym, so that you can run your own ANN design algorithm. I get that most data science students still use python, but the evolutionary world is kinda in C++ and ", "context": "", "use_case": "", "switched_from": "", "story": "Show HN: A competition for small neural networks that play strategy games"}]}, {"label": "Memory use", "summary": "One commenter built a lower-VRAM alternative to PyTorch's standard KL-divergence loss for long contexts.", "quotes": [{"id": 788114913, "item_id": 49257182, "url": "https://news.ycombinator.com/item?id=49257182", "by": "ikergarcia1996", "date": "2026-08-11", "stance": "compare", "quote": "The loss is mathematically equivalent to the standard KL-divergence loss available in PyTorch.", "excerpt": "Hello everyone! I\u2019ve been working on an efficient implementation of the KL-divergence loss that reduces its VRAM usage from quadratic to linear, using an approach very similar to Flash Attention: chunking and fusing the forward and backward passes. The loss is mathematically equivalent to the standard KL-divergence loss available in PyTorch. Until now, Knowledge Distillation required a huge amount of VRAM and was pra", "context": "", "use_case": "KL loss", "switched_from": "", "story": "Chunked KL loss, running Knowledge Distillation locally in less <6GB VRAM"}]}], "who": [{"by": "andy_ppp", "context": "building a startup in spare time", "url": "https://news.ycombinator.com/item?id=48935854", "date": "2026-07-16"}, {"by": "konz", "context": "ML6 recruiter", "url": "https://news.ycombinator.com/item?id=48780213", "date": "2026-07-03"}, {"by": "MPender08", "context": "independent researcher", "url": "https://news.ycombinator.com/item?id=47115828", "date": "2026-02-22"}, {"by": "RomanPushkin", "context": "Ruby developer for over a decade, wrote a book on Ruby", "url": "https://news.ycombinator.com/item?id=46382806", "date": "2025-12-24"}, {"by": "libraryofbabel", "context": "applies LLMs in day job", "url": "https://news.ycombinator.com/item?id=46359565", "date": "2025-12-22"}, {"by": "Grosvenor", "context": "data scientist with 20+ years experience", "url": "https://news.ycombinator.com/item?id=46110952", "date": "2025-12-01"}, {"by": "stared", "context": "deep learning practitioner, co-wrote 2018 PyTorch vs TensorFlow post", "url": "https://news.ycombinator.com/item?id=45845751", "date": "2025-11-07"}], "tags": ["for deep learning", "for ML research", "for Python developers", "for GPU training"], "use_cases": ["kl loss", "training pitch outcome model", "building ai projects", "ml engineering", "ml training baseline", "ml research code"], "compared": [{"slug": "gemini", "name": "Gemini", "co": 2, "switched_from": 0, "switched_to": 0}], "makers": [], "threads": [{"url": "https://news.ycombinator.com/item?id=45843948", "title": "Leaving Meta and PyTorch", "mentions": 2}, {"url": "https://news.ycombinator.com/item?id=49776523", "title": "Show HN: A competition for small neural networks that play strategy games", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49257181", "title": "Chunked KL loss, running Knowledge Distillation locally in less <6GB VRAM", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=49113999", "title": "Show HN: NegativeEV \u2013 I built a tool to help people see how bad their bets are", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48934046", "title": "Generative AI Is an Engineering Disaster", "mentions": 1}, {"url": "https://news.ycombinator.com/item?id=48747976", "title": "Ask HN: Who is hiring? (July 2026)", "mentions": 1}], "sentiment_by_month": {"2025-11": {"recommend": 2, "compare": 1}, "2025-12": {"neutral": 3, "recommend": 2}, "2026-01": {"neutral": 1}, "2026-02": {"neutral": 1}, "2026-03": {"recommend": 1}, "2026-04": {}, "2026-05": {}, "2026-06": {"compare": 1}, "2026-07": {"neutral": 2, "recommend": 1}, "2026-08": {"compare": 1}, "2026-09": {"warn": 1}, "2026-10": {}}, "quotes": [{"id": 796538369, "item_id": 49783648, "url": "https://news.ycombinator.com/item?id=49783648", "by": "cookiengineer", "date": "2026-09-21", "stance": "warn", "quote": "because of the hard pytorch requirements", "excerpt": "OMG! Just yesterday I published my reworked GoNEAT library that implements HyperNEAT combined with phased search and backpropagation [1]. But it's kind of impossible to enter for me because of the hard pytorch requirements :( would love to see the project as a gym, so that you can run your own ANN design algorithm. I get that most data science students still use python, but the evolutionary world is kinda in C++ and ", "context": "", "use_case": "", "switched_from": "", "story": "Show HN: A competition for small neural networks that play strategy games"}, {"id": 788114913, "item_id": 49257182, "url": "https://news.ycombinator.com/item?id=49257182", "by": "ikergarcia1996", "date": "2026-08-11", "stance": "compare", "quote": "The loss is mathematically equivalent to the standard KL-divergence loss available in PyTorch.", "excerpt": "Hello everyone! I\u2019ve been working on an efficient implementation of the KL-divergence loss that reduces its VRAM usage from quadratic to linear, using an approach very similar to Flash Attention: chunking and fusing the forward and backward passes. The loss is mathematically equivalent to the standard KL-divergence loss available in PyTorch. Until now, Knowledge Distillation required a huge amount of VRAM and was pra", "context": "", "use_case": "KL loss", "switched_from": "", "story": "Chunked KL loss, running Knowledge Distillation locally in less <6GB VRAM"}, {"id": 785824306, "item_id": 49114019, "url": "https://news.ycombinator.com/item?id=49114019", "by": "qkwrv", "date": "2026-07-30", "stance": "neutral", "quote": "it builds a PyTorch model targeting the exact pitch result", "excerpt": "I know I say it in the title, but I want to be very clear up front: the predominant output of this tool is \"don't place this bet.\" Betting apps want you to bet. If they sense you are tired of losing and might leave, they'll entice you with deposit bonuses, boosted odds and protected picks. There are even people, you see them self-promoting all over X, who sell access to \"good\" bets. Obviously, they don't refund you i", "context": "", "use_case": "training pitch outcome model", "switched_from": "", "story": "Show HN: NegativeEV \u2013 I built a tool to help people see how bad their bets are"}, {"id": 782973664, "item_id": 48935854, "url": "https://news.ycombinator.com/item?id=48935854", "by": "andy_ppp", "date": "2026-07-16", "stance": "recommend", "quote": "have built many projects with PyTorch and use it daily for all my work", "excerpt": "I love AI and have built many projects with PyTorch and use it daily for all my work and am building a startup in my spare time here: https://veloa.com/ Being concerned about how AI will concentrate power doesn\u2019t mean I don\u2019t think it\u2019s extremely useful.", "context": "building a startup in spare time", "use_case": "building AI projects", "switched_from": "", "story": "Generative AI Is an Engineering Disaster"}, {"id": 780483409, "item_id": 48780213, "url": "https://news.ycombinator.com/item?id=48780213", "by": "konz", "date": "2026-07-03", "stance": "neutral", "quote": "a range of ML frameworks such as TensorFlow, PyTorch, or HuggingFace Transformers", "excerpt": "Can you change this to the following: ML6 | Senior AI Engineer | Python, TensorFlow, PyTorch, GCP, AWS, Azure | Full-time | Amsterdam, Berlin, Munich, Eindhoven, Ghent (EU) On-site/hybrid We are an AI engineering company that builds end-to-end AI solutions. By applying the latest AI research, we keep our clients at the forefront of innovation. If you are interested check out: https://www.ml6.eu/en/customers/cases/ Wo", "context": "ML6 recruiter", "use_case": "ML engineering", "switched_from": "", "story": "Ask HN: Who is hiring? (July 2026)"}, {"id": 779489922, "item_id": 48718120, "url": "https://news.ycombinator.com/item?id=48718120", "by": "vforno", "date": "2026-06-29", "stance": "compare", "quote": "Compared to PyTorch, it's a bit slower (probably 1.5-2.5x)", "excerpt": "Hi, in nanoeuler I use cuBLAS (NVIDIA's super-optimized library) for all matrix multiplications, with the tensor cores in TF32 mode. It's the same thing PyTorch uses underneath, so it's very fast. What I've optimized (and will improve even more) and written by hand are the kernels for the other parts (like FlashAttention, which gave a nice 3x speedup), while I've delegated the large matrices to cuBLAS. Training the 1", "context": "", "use_case": "ML training baseline", "switched_from": "", "story": "Show HN: NanoEuler \u2013 GPT-2 scale model in pure C/CUDA from scratch"}, {"id": 761259952, "item_id": 47578747, "url": "https://news.ycombinator.com/item?id=47578747", "by": "kadushka", "date": "2026-03-30", "stance": "recommend", "quote": "use good old Pytorch", "excerpt": "1. If you have good results on sufficiently large models (check latest papers re: which benchmarks are still relevant), post them on Github, along will detailed instructions how to reproduce. 2. Post the link to the GH repo in \"Show HN\" section. 3. If results are solid, write up a paper and upload it to arxiv. Next step would be try to publish in an ML conference. p.s. To increase your chances of anyone actually clic", "context": "", "use_case": "ML research code", "switched_from": "", "story": "Ask HN: Release Path for 'Transformers Alternatives'?"}, {"id": 753853249, "item_id": 47115828, "url": "https://news.ycombinator.com/item?id=47115828", "by": "MPender08", "date": "2026-02-22", "stance": "neutral", "quote": "built a theoretical hardware framework and PyTorch simulation", "excerpt": "Hey HN, I\u2019m an independent researcher. I've been exploring the thermodynamic limits of AI and built a theoretical hardware framework and PyTorch simulation to address it. The Problem: Current digital AI is hitting the \"Landauer Wall\": the massive energy cost of erasing bits and brute-forcing hierarchical data in flat, static Euclidean matrices. The Project: This is the Manifold Chip, a Dynamically Gated Analog Crossb", "context": "independent researcher", "use_case": "simulation", "switched_from": "", "story": "Show HN: The Manifold Chip \u2013 Bypassing the Landauer Wall in Analog Silicon"}, {"id": 747083764, "item_id": 46692735, "url": "https://news.ycombinator.com/item?id=46692735", "by": "fotcorn", "date": "2026-01-20", "stance": "neutral", "quote": "validated it against the PyTorch version", "excerpt": "I used Claude Opus 4.5 inside Cursor to write RISC-V Vector/SIMD code. Specifically Depthwise Convolution and normal Convolution layers for a CNN. I started out by letting it write a naive C version without intrinsic, and validated it against the PyTorch version. Then I asked it (and two other models, Gemini 3.0 and GPT 5.1) to come up with some ideas on how to make it faster using SIMD vector instructions and write ", "context": "", "use_case": "reference implementation", "switched_from": "", "story": "Ask HN: Do you have any evidence that agentic coding works?"}, {"id": 742124898, "item_id": 46382806, "url": "https://news.ycombinator.com/item?id=46382806", "by": "RomanPushkin", "date": "2025-12-24", "stance": "recommend", "quote": "Python has won for me: fastapi, pytorch, langchain, streamlit, and so on and on.", "excerpt": "I'm happy to see v4.0, but 2025 was the year I switched from Ruby to Python after gradually drifting back to it more and more. The tipping point was when I had Claude Code automatically convert one of my Ruby projects to 100% Python - and after that, I just had no Ruby left. I spent over a decade enjoying Ruby and even wrote a book about it. At this point, though, Python has won for me: fastapi, pytorch, langchain, s", "context": "Ruby developer for over a decade, wrote a book on Ruby", "use_case": "", "switched_from": "", "story": "Ruby 4.0.0 Released"}, {"id": 741753040, "item_id": 46359565, "url": "https://news.ycombinator.com/item?id=46359565", "by": "libraryofbabel", "date": "2025-12-22", "stance": "neutral", "quote": "implementing a transformer in PyTorch just to refresh my memory", "excerpt": "I read this article back when I was learning the basics of transformers; the visualizations were really helpful. Although in retrospect knowing how a transformer works wasn't very useful at all in my day job applying LLMs, except as a sort of deep background for reassurance that I had some idea of how the big black box producing the tokens was put together, and to give me the mathematical basis for things like contex", "context": "applies LLMs in day job", "use_case": "implementing a transformer", "switched_from": "", "story": "The Illustrated Transformer"}, {"id": 740585008, "item_id": 46286563, "url": "https://news.ycombinator.com/item?id=46286563", "by": "stared", "date": "2025-12-16", "stance": "recommend", "quote": "When it comes to matrix notation, I prefer PyTorch over NumPy", "excerpt": "On the contrary, I think that well-designed general-purpose languages beat domain-specific languages. Even in the example given, in NumPy you can use np.array, but to make a fair comparison, use np.matrix. import numpy as np X = np.matrix([1, 2, 3]) Y = np.matrix([[1, 2, 3], [4, 5, 6], [7, 8, 9]]) Z = Y * X.T W = np.hstack([Z, Z]) That way, we can extend our languages. If np.matrix is \"too many keystrokes\", it can be", "context": "", "use_case": "tensor math", "switched_from": "", "story": "In Defense of Matlab Code"}, {"id": 739102129, "item_id": 46193883, "url": "https://news.ycombinator.com/item?id=46193883", "by": "konz", "date": "2025-12-08", "stance": "neutral", "quote": "a range of ML frameworks such as TensorFlow, PyTorch, or HuggingFace Transformers", "excerpt": "ML6 | Machine Learning Engineer, Data Engineer | Python, TensorFlow, PyTorch, GCP, AWS, Azure | Full-time | Amsterdam, Berlin, Munich, Eindhoven, Ghent (EU) On-site/hybrid We are a Machine Learning consulting company that builds end-to-end Machine Learning solutions. By applying the latest AI research, we keep our clients at the forefront of innovation. If you are interested check out: https://www.ml6.eu/knowledge-hu", "context": "hiring at ML6", "use_case": "machine learning projects", "switched_from": "", "story": "Ask HN: Who is hiring? (December 2025)"}, {"id": 737775232, "item_id": 46110952, "url": "https://news.ycombinator.com/item?id=46110952", "by": "Grosvenor", "date": "2025-12-01", "stance": "neutral", "quote": "I use the normal stacks you'd expect. Python, Pytorch, Spark/ray, Jupyter/Merimo, AWS, Postgres, Mathematica", "excerpt": ".-~~\\ / \\ _ ~x .-~_)_ ~x\".-~ ~-. _ ( / \\ _ || T o o Y || ==:l l < ! I;== \\\\ \\ .__/ / // \\\\ ,r\"-,___.-'r.// }^ \\.( ) _.'//. / }~Xi--~ // \\ Y Y I\\ \\ \" Y | | |o\\ \\ | | l_l Y T | l \"o l_j ! \\ / ___,.---^. o .^---.._____ \"~~~ \" ~ ~~~\" SEEKING WORK - Data scientist, consulting & fractional leadership, US/remote worldwide, email in profile. All I want for christmas is some gnarly problems to chew on, otherwise it's coal for", "context": "data scientist with 20+ years experience", "use_case": "data science consulting", "switched_from": "", "story": "Ask HN: Who wants to be hired? (December 2025)"}, {"id": 733643665, "item_id": 45852729, "url": "https://news.ycombinator.com/item?id=45852729", "by": "sega_sai", "date": "2025-11-07", "stance": "compare", "quote": "how jax compares to this autotuning machinery in pytorch", "excerpt": "I switched from pytorch to jax just before triton appeared. Does anyone know how jax compares to this autotuning machinery in pytorch ? I know jax does jit, but i don't have a good intuition if jit is better than this type of autotuning.", "context": "", "use_case": "ML", "switched_from": "", "story": "Helion: A high-level DSL for performant and portable ML kernels"}, {"id": 733562064, "item_id": 45847629, "url": "https://news.ycombinator.com/item?id=45847629", "by": "jmward01", "date": "2025-11-07", "stance": "recommend", "quote": "I use pytorch daily, it has made it possible for me to play with ideas that I would have only dreamed of.", "excerpt": "All I can say is 'thanks!'. It does take a team, and a community, but individuals matter. I use pytorch daily, it has made it possible for me to play with ideas that I would have only dreamed of. It is a big part of my life so, thanks for your contributions and best of luck on the next thing!", "context": "", "use_case": "ML experimentation", "switched_from": "", "story": "Leaving Meta and PyTorch"}, {"id": 733532016, "item_id": 45845751, "url": "https://news.ycombinator.com/item?id=45845751", "by": "stared", "date": "2025-11-07", "stance": "recommend", "quote": "the more it developed, the better it was, with (at least to my mind) the perfect level of abstraction.", "excerpt": "In 2018, I co-wrote a blog post with the inflammatory title \u201cDon\u2019t use TensorFlow, try PyTorch instead\u201d (https://news.ycombinator.com/item?id=17415321). As it gained traction here, it was changed to \u201cKeras vs PyTorch\u201d (some edgy things that work for a private blog are not good for a corporate one). Yet the initial title stuck, and you can see it resonated well with the crowd. TensorFlow (while a huge step on top of T", "context": "deep learning practitioner, co-wrote 2018 PyTorch vs TensorFlow post", "use_case": "deep learning", "switched_from": "Keras", "story": "Leaving Meta and PyTorch"}]}