# Why PyTorch Won  ## Episode metadata - Episode title: Why PyTorch Won - Show: The New Stack Podcast - Owner / Host: The New Stack - Guests: [Luca Antiga](https://share.snipd.com/person/e93beb98-9b87-47e1-94f8-8ffd071622e2) - Episode publish date: 2025-10-24 - Episode AI description: Luca Antiga, CTO of Lightning AI and chair of the PyTorch Foundation Technical Advisory Council, shares insights from his journey in AI. He discusses PyTorch's evolution and its role in major AI advancements, emphasizing its flexibility in reinforcement learning and generative models. Antiga highlights the integration of projects like vLLM and DeepSpeed within the PyTorch ecosystem, all while maintaining a focus on user experience. He also touches on exciting trends like efficient learning models and the potential for future projects within the foundation. - Duration: 29:43 - Episode URL: [Open in Snipd](https://share.snipd.com/episode/f35418b4-ed09-4268-8e64-3d3681cd10d3) - Show URL: [Open in Snipd](https://share.snipd.com/show/366c5a71-cfd2-4429-8bfb-8a29b6518715) - Export date: 2026-02-17T23:36:37 ## Snips ### [From Biomedical Engineer To PyTorch Advocate](https://share.snipd.com/snip/49719592-3d89-44be-87f9-b7732c0ce7d9) 🎧 00:30 - 01:54 (01:24) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/e2301e77-b891-4338-8b57-31468df11df7" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Luca Antiga described his path from biomedical engineering to contributing to Torch7 and PyTorch. - He wrote a PyTorch book and later co-founded Lightning AI, which naturally joined the PyTorch Foundation. #### 💬 Quote > I graduated at Polytechnic of Milano in 1999... then I started my own company... we started contributing to Torch7. and then as soon as PyTorch was released, then I started contributing... > — Luca Antiga Luca Antiga recounting his career journey into PyTorch and Lightning AI #### 📚 Transcript **Luca Antiga:** TAC? I said that I planned for it, but it would be impossible. So, yeah, I'm a biomedical engineer. Okay. Yeah. I at Polytechnic of Milano in 1999. Can you believe how long it was? That's crazy. I **Unknown Speaker:** graduated high school around the same time. Yeah, **Luca Antiga:** so then I was in research for about 10 years in biomedical research, doing medical image analysis. Back in the day, we were doing a lot of like partial differential equations. I didn't do any machine learning at all. And then I started my own company, Aerobics, in Italy in 2009. And as part of that, we started doing deep learning. And we started contributing to Torch7. And then as soon as PyTorch was released, then I started contributing, getting in touch with Sumit, Adam, and all the others. Oh, very well. Nice. For some reason, I ended up in the paper. and then I wrote a book about PyTorch. The paper. Yeah. And then, yeah, this is, and then Lightning happened, and then, yeah, Lightning, of course, was born out of PyTorch Lightning. Lightning, the company, **Unknown Speaker:** Lightning. **Luca Antiga:** Lightning AI, the company. Oh, you're the **Unknown Speaker:** CTO, right? Is that correct? Yeah, **Luca Antiga:** I'm CTO right now. And it was very natural for PyTorch Lightning, the company behind PyTorch Lightning, to become part of the PyTorch Foundation. --- ### [Pythonic Design Fueled PyTorch's Rise](https://share.snipd.com/snip/4e0c65bc-2bf4-4948-a99b-cfca54b558d7) 🎧 02:30 - 04:08 (01:37) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/94cd028c-a863-48df-976d-3e48d381228b" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - PyTorch is a Python library that automates optimization and runs fast on GPUs for neural network training. - Its Pythonic, researcher-friendly design made debugging and iteration much faster than earlier frameworks. #### 💬 Quote > PyTorch was kind of revolutionary in that sense. and iteration speed of Python and, you know, the bias toward action that Python has, right? > — Luca Antiga Luca Antiga on PyTorch's Pythonic, researcher-friendly advantage #### 📚 Transcript **Luca Antiga:** so pytorch is a python library uh that makes it easy to run programs that need to optimize themselves automatically but they need to run fast on gpu they need to do a lot of computations to do so. And what I mean by optimize themselves is the way neural networks work is that we have a lot of operations like multiplying very large tensors or multidimensional arrays of numbers by other multidimensional arrays of numbers. So you need GPUs to do that. But also, you need to then compute an outcome, which is an error, and then try to minimize that error. And this is how models have trained. And so if you combine the ability to compute, in mathematical terms, the derivatives of the error with respect to all the parameters in the next work to understand how to change the parameters so to minimize the error, PyTorx comes with that built-in. But at the same time, compared to other frameworks back in the day, it makes it in a way that is very researcher-framely. So it's very Pythonic. Whereas back in the day, there were many frameworks that forced you, that were a bit like, yeah, they were in Python, but you needed to code in a kind of meta language that was behind the problem in yourself. And it makes debugging a lot harder and so on. And was kind of revolutionary in that sense. and iteration speed of Python and, you know, the bias toward action that Python has, right? --- ### [Researchers' Adoption Scaled PyTorch Industry-Wide](https://share.snipd.com/snip/a839fe88-d8f9-46e5-bc0c-f8191f126732) 🎧 05:24 - 06:38 (01:14) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/db42dff0-ab79-4638-a5a9-5b2c5ebb50cc" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - PyTorch remained central across multiple AI revolutions from small models to generative LLMs. - Researchers adopted PyTorch early, then moved into industry, turning it into an industry-scale platform. #### 💬 Quote > It really took forward because it won the researchers, right? And then the researchers eventually graduated to go, work somewhere, eventually, or to do more research. > — Luca Antiga Luca Antiga on how researcher adoption drove PyTorch's industry dominance #### 📚 Transcript **Luca Antiga:** Throughout these revolutions that came, you always see PyTorch there. And there are, of course, other like JAX and so on, they're very strong. But comparatively, PyTorch is a whole industry, powers that whole industry. And together with that came a great ecosystem. Before, it was an ecosystem that still is focused around computer vision. And then with language, you had this very important inference. Typically, before, what you did is use PyTorch to train and then something else to run inference. There was Torcer, but they were all things that evolved at a bit of a slower pace. Today, if you go and see the inference frameworks are most popular, BLLM, SGLang, and they run PyTorch, and they run PyTorch in production. And it's actually Python, because we talk to developers right now, and they say, what? PyTorch in production? Like, in Power? And it is. If you interact with any chatbot today, chances are that you're running Python at that time. It's not that it's exported anywhere else. It's still PyTorch running. --- ### [PyTorch Powers Production Inference](https://share.snipd.com/snip/d993cc87-1bb0-49cb-8d77-f18d81f01b13) 🎧 05:51 - 06:26 (00:34) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/8ca24ae6-65c9-46c1-972a-e5bba858ed0b" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Modern inference frameworks like vLLM and SGLang run PyTorch in production for chatbots and LLM serving. - Many production chatbots today run Python and PyTorch for inference, not just training. #### 💬 Quote > If you interact with any chatbot today, chances are that you're running Python at that time. It's not that it's exported anywhere else. It's still PyTorch running. > — Luca Antiga Luca Antiga on PyTorch's role in production inference for chatbots #### 📚 Transcript **Luca Antiga:** Typically, before, what you did is use PyTorch to train and then something else to run inference. There was Torcer, but they were all things that evolved at a bit of a slower pace. Today, if you go and see the inference frameworks are most popular, BLLM, SGLang, and they run PyTorch, and they run PyTorch in production. And it's actually Python, because we talk to developers right now, and they say, what? PyTorch in production? Like, in Power? And it is. --- ### [Optimize For Fast Iteration](https://share.snipd.com/snip/bd8f5d9a-519b-40fd-9e25-be68a2e4eadd) 🎧 06:50 - 07:36 (00:45) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/de881d62-5043-4ea7-8fa5-3328c8abdbbe" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Prioritize iteration speed and developer productivity over building a perfect system upfront. - Use tools that let you prototype and improve quickly because conditions change rapidly. #### 💬 Quote > The speed of incremental improvement is what really matters here, because things are so fast that you cannot just stop and invest your time into building the perfect thing. > — Luca Antiga Luca Antiga advising on why iteration speed matters in AI development #### 📚 Transcript **Luca Antiga:** Well, I think the reason why PyTorch won is the same reason why Python won in the AI space, which means a ton of people can work in it with very low cognitive always. They don't need to know many things in order to assemble a system. And the speed of incremental improvement is what really matters here, because things are so fast that you cannot just stop and invest your time into building the perfect thing. There's no such thing. Like two months later, everything will have to be. So what you do, you need to optimize for your speed of iteration and execution at that particular point in time with the context that you have at that particular point in time. And that's why PyTorch and PyCon are ultimately the winners. --- ### [Foundation Hosting Lowers Contribution Friction](https://share.snipd.com/snip/2f8a185d-9b95-482d-bae7-496fca92891c) 🎧 07:52 - 08:33 (00:40) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/14350005-8b15-4346-8514-99a006d65fcd" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Moving projects like vLLM into the PyTorch Foundation makes them vendor-neutral and easier for companies to contribute to. - This neutrality reduces worries about licensing or future closed-source moves and increases corporate contributions. #### 💬 Quote > Whoever had VLLM gave the, say, copyright and auto-active property and assets to the Linux Foundation to be managed. > — Luca Antiga Luca Antiga explaining vLLM's transfer to a neutral foundation #### 📚 Transcript **Luca Antiga:** Yeah, the foundation initiative is just a few months, three, four months. I don't want to, I **Unknown Speaker:** don't know. **Luca Antiga:** Yeah, just months. And VLLM was, I think, the second project to come in as a hosted project. That means that whoever had VLLM gave the, say, copyright and auto-active property and assets to the Linux Foundation to be managed. So basically, VLLM was moved into a venture-neutral space. And if you think about that, it's very fascinating because it's not a donation per se. Of course, you need to donate. --- ### [DeepSpeed Pioneered Large-Scale Training Tricks](https://share.snipd.com/snip/6064c163-36b1-4633-8375-13e81251328c) 🎧 09:34 - 10:33 (00:58) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/2be0e1d2-58b9-4b33-8bff-04f192e527e1" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - DeepSpeed pioneered distributed training strategies like parameter sharding and influenced multi-company contributions. - It enabled running large models across GPUs and machines without major efficiency losses. #### 💬 Quote > DeepSpeed... pioneered some distributed strategies with sharding parameters and so on, making it easy to define a model the way you define it mathematically. > — Luca Antiga Luca Antiga on DeepSpeed's impact on distributed training #### 📚 Transcript **Unknown Speaker:** Yes. I think people may be less familiar with them than PyTorch and VLLM. **Luca Antiga:** Yeah, yeah, exactly. So DeepSpeed dates back a while, and it was born out of Microsoft, and now it's being contributed by multiple companies, even before it was brought into the PyTorch Foundation. across GPUs, across machines. And so it pioneered some distributed strategies with sharding parameters and so on, making it easy to define a model the way you define it mathematically, let's say. And then how do I run that across machines without paying penalties on efficiency? And then they moved in also other fronts, even inference optimizations and compiler optimizations for distributors. So it's a very, very, it's more niche for sure than something like VLL. --- ### [Ray As A Composable Orchestration Fabric](https://share.snipd.com/snip/fe19bf4b-6888-452a-8f3c-f43f069e5555) 🎧 11:01 - 12:19 (01:17) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/3a661e20-563e-4b7b-9648-a78b5d3a5a3a" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Ray provides composable primitives to orchestrate distributed and remote Python computations at scale. - Ray fits workloads beyond AI, like massively parallel data transforms and reinforcement learning orchestration. #### 💬 Quote > Ray is a technology that came out of the need to distribute and orchestrate computations happening in different places. > — Luca Antiga Luca Antiga on Ray's role for distributed computation and orchestration #### 📚 Transcript **Luca Antiga:** Ray is a technology that came out of the need to distribute and orchestrate computations happening in different places. So it's a way to say, okay, I'm going to have some computations happening remotely, and I'm going to move some data from here to there and do it in a very controllable and hackable way with very strong primitives. So it's a very composable system. **Unknown Speaker:** Right. And so a **Luca Antiga:** lot of companies that operate at massive scales are usually ready to orchestrate their workloads, especially in the world where it's not just one training workload that has the same processes running identically across the cluster. But especially in a world where like reinforcement learning, for example, where you have something training, something managing experience, where traces and then something else that needs to generate more rollouts of the current policy. And so you have an inference part of it, too. And so you don't have a cluster anymore. You have more than one cluster. But they need to act in consultation. Otherwise, your GPUs are idle. Right, --- ### [Make Projects Easy To Adopt](https://share.snipd.com/snip/d92a3f90-1ec2-4ab7-a391-3f63dd1bff85) 🎧 14:05 - 15:17 (01:11) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/4741f0f6-d7df-45be-b640-1afc3330087a" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - Make ecosystem projects easy to adopt with clear READMEs, quickstarts, and version support. - Ensure tests and explicit compatibility so users can land and solve problems without friction. #### 💬 Quote > My goal is to have them success. You go there. The readme is clear, you have a quick start, you have tests, you have support for all the versions of PyTorch... > — Luca Antiga Luca Antiga on improving user experience in the PyTorch ecosystem #### 📚 Transcript **Luca Antiga:** The ecosystem is a set of projects that are not donated, not in this transfer, but there is a certain bar to enter the ecosystem. collect projects that build upon the foundations or are somehow related and that enable users to achieve some goals. Like, for example, there are a lot of projects in science verticals, like geosciences, biology and so on, physics, or other projects as well, more technological. And with the ecosystem, we want to not only create a cambrian explosion of projects from a developer standpoint. What I'm mostly concerned about in the ecosystem is the users. What journey do the users experience when they approach an ecosystem that is kind of sanctioned by the PyTorch Foundation? Yeah. My goal is to have them success. You go there. The readme is clear, you have a quick start, you have tests, you have support for all the versions of PyTorch, or at least you're explicit about what versions, and you support the latest ones. --- ### [LLMs Make Reinforcement Learning Practical](https://share.snipd.com/snip/0b3bbceb-9f43-4492-b68b-69b80b27802e) 🎧 16:43 - 19:06 (02:23) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/aa7ca316-1e92-4381-bade-d3729f279ab7" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - LLMs serve as strong initial policies that make reinforcement learning effective on complex tasks. - Starting from pretrained models reduces reward sparsity and accelerates RL optimization. #### 💬 Quote > It works really well if your initial policy is kind of already good at solving that thing or at least gets close to the task. > — Luca Antiga Luca Antiga on why pretrained LLMs improve reinforcement learning outcomes #### 📚 Transcript **Luca Antiga:** And LLMs have been instrumental in that journey because RL has been there for a long time. Sure. But it works really well if your initial policy is kind of already good at solving that thing or at least gets close to the task. And then RL can bump it up and optimize it. **Unknown Speaker:** Explain that a little bit because I'm not sure everybody's familiar with how that works. absolutely. So reinforcement learning will **Luca Antiga:** encourage a model to do more of the things that work when confronted with an environment that, you know, you can go in with an action, the environment state changes and you get a reward for your actions. Because **Unknown Speaker:** that's how reinforcement learning with some of those early examples of models playing games or things like that, right? Which was done that way too. **Luca Antiga:** Yeah, it's exactly the same things and more like very often they're exactly the same algorithms that we're using the Atari games by DeepMind and so on. And what makes RL fragile and hard to pull off for complicated things is the fact that you're starting from scratch. So you're starting from a model that doesn't know how to do anything, which is not a problem for Atari kind of environments because the action space is small and the state is not very large. But then what happens when you have a very complicated to model environment where an agent needs to do actions that will bring a very sparse reward, you need to hit that reward sometimes. So if you're kind of spanning the space at random, your signal will be so sparse that it's really hard to understand when you're doing good and probably bad. Whereas if you start from something like an LLM that has some concept of how the world works, it means that more often than not, we'll get a reward that is not zero. And that will help the system get better at it. And there are a million details that will dictate whether the thing works or not. But I think that will kind of change in the era of LLM page regions. That is the thing. Being able to start from something that is already quite okay. --- ### [Miniaturize Efficiency, Not Just Model Size](https://share.snipd.com/snip/c8b35213-2bb1-49a3-8465-acad122b2d74) 🎧 21:56 - 24:13 (02:17) <iframe src="https://share.snipd.com/embed/obsidian-player/snip/fd7ace64-2190-4725-a61f-7bd06c919ad8" width="100%" height="100" style="border: none; border-radius: 12px;" sandbox="allow-scripts allow-same-origin allow-forms allow-popups allow-clipboard-write" ></iframe> - The future challenge is making models that ingest knowledge top-to-bottom efficiently and use far less energy. - Miniaturization means designing more efficient intelligence, not just shrinking parameter counts. #### 💬 Quote > I think about models that can... pick up new knowledge. When I read a differential geometry book... I can consume some power to do that kind of ingestion of knowledge. > — Luca Antiga Luca Antiga on the goal of making models learn deeply and efficiently #### 📚 Transcript **Luca Antiga:** don't know if smaller. I think about models that can... Because smaller depends on what the hardware is, right? right? So smaller of today was the super large a **Unknown Speaker:** few years ago. But **Luca Antiga:** I think, yeah, it's about the efficiency of, well, I think two things. One, how effectively a model can pick up new knowledge. When I read a differential geometry book, if that... Have one **Unknown Speaker:** dust on the weekend. **Luca Antiga:** No, all right. I'm picking a subject, it is hard, but then you need to read it once, read it twice, read it top to bottom. And this is not how models learn nowadays. This is not at all how they learn. so I would like something to be asked to get to the point where we can consume some power to do that kind of ingestion of knowledge. Whereas, and where after that, I can rely on that model to know that. Now, of course, we have retrieval of information and so on. But in a way, I think humankind have landed on something that is amazing, right? So it's training systems that go beyond our capabilities of explaining how to do things. We don't have to explain to a computer how to do things. But here it are. I call them kind of analogy machines. We are building analogy machines in very multidimensional space. But at the same time, the way we're doing that is like humankind may have landed on a flying machine that was like huge contraption of wheels and spinning things. But that's not the only design. The proof that we have is birds. **Unknown Speaker:** Yeah, yeah. **Luca Antiga:** the proof that we have that intelligence can be different is our brain right and how many watts we consume per thought is like it's a head it's in front of us and i don't think it's in managing mode that what we landed on with large language model and transformers and large scale stuff can be now miniaturized in a way that can learn top to bottom. And it won't consume so much memory, so much energy. --- Created with [Snipd](https://www.snipd.com) | Highlight & Take Notes from Podcasts