AI 建设者应关注的 60 个 GitHub 账号

✍ Meta Alchemist🕐 2026-05-03📦 40.8 KB 🟢 已读 𝕏 文章列表

文章精选了 60 位 AI 领域的关键 GitHub 开发者，涵盖了从基础架构到应用层的各个方向。作者不仅列出了名单，更提供了一套将研究成果转化为产品的行动指南，建议读者深入挖掘这些开源项目背后的“为什么”，并从中寻找未被开发的蓝海产品机会。

AI 开源 GitHub Agent LLM 学习指南工具推荐开发者

# 60 GitHub accounts every AI builder should follow to be great at AI **作者**: Meta Alchemist **日期**: 2026-05-03T07:07:53.000Z **来源**: [https://x.com/meta_alchemist/status/2050834716157501710](https://x.com/meta_alchemist/status/2050834716157501710) --- ![Image](/doc/1499/image/img_94bac68f49b4.jpg) A curated walkthrough of the people behind the tools you already use, plus a playbook for translating their research into products you can ship. Most builders who say they're "in AI" rarely open the GitHub of the people actually building the field. They follow launch threads, demo clips, and influencer commentary. They miss the source. This article is a curated walk-through of 60 of those accounts. Real people, not company logos, mapped across the seven layers of the AI stack you touch every day. For each account you'll get a short, honest read: what they built, why the work matters, and the part most "follow these accounts" lists skip, what you as a builder can actually take from their work and turn into your own. A repo to run, a paper to read, a gap in the ecosystem to fill, a tiny product nobody else is building yet. Save this. Open one profile per day for the next 60 days. By the end of the month, you'll have a sharper map of the AI world, a working short list of repos worth cloning, patterns of the AI world, and blue oceans to claim. Within two months, you'll be at a level that really understands the world of AI. That's the difference between watching the AI moment from the outside and actually being part of it. ## How to read this list to actually get value out of it Following 60 accounts and feeling smarter is a trap. The point is to use these profiles as doorways into ideas, patterns, and projects you can build on, whether you write code or not. Here's how to do that without wasting your time. - Don't just follow them, dig into what they made. The interesting part of any builder's work is rarely the technical detail. It's the why. Why did they build this? What problem were they trying to solve? What did they say no to? You'll find the answers in their pinned posts, READMEs, podcast interviews, and old threads. Understand the why. That's where the lesson is. - Look for the unfair advantage they had. Every person on this list shipped because of one specific edge: deep knowledge of a niche, perfect timing, ruthless focus, or a willingness to do the unsexy work everyone else avoided. Spot the edge. Then ask yourself which kind of edge you have, and what you should be building because of it. - Engage with their work, even if you don't write code. Pick one builder per week. Read their writing, watch one of their talks, try the tool they made if you can, leave a thoughtful reply on something they posted. You don't have to be an engineer to learn from them. You just have to engage deeply with their style and what they stood for, rather than skimming. This will give you a ton of new insight into the AI world. - Treat their work as your starter kit. Most things on this list are missing a layer that someone with taste, hustle, or domain knowledge can add. A nicer interface on top of a powerful tool. A version of the tool made for a specific industry. A simpler way to package what they built so a regular person can use it. That "easy layer on top of brilliant work" is one of the highest-leverage moves in the AI economy right now, and you don't need a computer science degree to do it. Just need to spend enough time in the terminal after you get the inspiration. - Translate research into real businesses. Every important paper or repo on this list is a product opportunity in disguise. When a new AI technique drops, the first few people to wrap a clean product around it for a specific audience usually own the category. The technical part gets easier every quarter. The taste, the audience, and the packaging are where the money is. - Follow them, then follow who they follow. The AI ecosystem is smaller than it looks. The people on this list mostly know each other, build on each other's work, and amplify each other's research. Follow these 40, then follow 5 to 10 of the people each of them follows. Your feed will then turn into a frontier scanner instead of a hype filter. Group 1: Foundational and classic ## 1. karpathy, Andrej Karpathy The closest thing the AI world has to a teacher you can actually learn from. Every project he drops, from nanoGPT to llm.c to micrograd, walks you through how transformers really work, line by readable line. Follow him to go from "I use AI tools" to "I understand the model". You'll be 10x more dangerous as a builder once you know the machinery underneath. What you can build from this: rebuild nanoGPT yourself in a weekend, then write your own tutorial in your own voice. Teaching what he teaches is one of the highest-conversion content angles in AI right now. ## 2. ggerganov, Georgi Gerganov One developer made it possible to run massive LLMs on a laptop. That's llama.cpp, and Georgi built it almost solo while everyone else was waiting for cloud GPUs. Follow him for proof of what one focused builder can ship. His code sits under Ollama, llamafile, and most of the local AI tools you already use. What you can build from this: a vertical product on top of llama.cpp for users who can't send data to the cloud (lawyers, doctors, accountants). The runtime is solved, the verticalization is wide open. ## 3. geohot, George Hotz Jailbroke the iPhone at 17, now building tinygrad to compete with PyTorch. He ships in public, streams the work live, and doesn't care who he upsets along the way. Follow him for raw building energy and contrarian takes on the AI infra stack. You won't always agree, you'll always learn something. What you can build from this: a clean tutorial on tinygrad for a specific use case. The framework is great, the educational layer is thin, and that gap is yours to take. ## 4. soumith, Soumith Chintala Co-creator of PyTorch, the framework almost every serious AI lab on earth runs on. Quiet, technical, respected by everyone who matters in the field. Follow him to see where deep learning infrastructure is heading from someone actually shaping it. His takes on hardware and open ecosystems land before the trend does. What you can build from this: watch what he signal-boosts. The frameworks and hardware he points at six months early are the ones to skill-stack on now. ## 5. fchollet, François Chollet Built Keras, which made deep learning accessible to millions. Now he runs the ARC challenge, the cleanest benchmark for whether a model is actually reasoning or just memorizing. Follow him for sharp, principled thinking about what AI is and isn't. He'll keep your head clear when the hype cycle gets loud. What you can build from this: tooling around ARC. Visualizers, dataset explorers, leaderboards, agents that attempt the puzzles. The benchmark is the new frontier and the toolchain around it is barely built. ## 6. ylecun, Yann LeCun Turing Award winner (2018), recently left Meta after years as Chief AI Scientist to start his own venture, and the guy who keeps loudly disagreeing with the "LLMs are everything" crowd. Convolutional networks exist because of him. Follow him for a counterweight to the daily AI hype cycle. He's been right about a lot of things long before they were consensus. What you can build from this: read what he's skeptical about. Some of the biggest opportunities in AI sit exactly in the lanes the consensus is overhyping and a few sober voices are warning about. ## 7. lucidrains, Phil Wang Reimplements almost every important AI paper in clean PyTorch, usually within days of it dropping. One person, hundreds of repos, real quality across all of them. Follow him if you'd rather understand papers through code than math. Fastest way to grasp new architectures without doing a PhD. What you can build from this: turn his bare-bones implementations into productized notebooks, courses, or developer kits. He ships the science, the educational and product layer is wide open. ## 8. rwightman, Ross Wightman Built timm, the library every computer vision team quietly depends on. It powers a huge slice of production vision work today. Follow him if vision models are anywhere near your stack. He shares benchmarks, training tricks, and model comparisons you won't find elsewhere. What you can build from this: a vertical CV product (inventory counting, document parsing, defect detection) using timm as the backbone. The model zoo is solved, deployment-ready vertical apps are not. ## 9. srush, Sasha Rush Longtime Cornell professor, now at Cursor Research. The Annotated Transformer and MiniTorch are both his, and they're some of the cleanest learning material in ML. Follow him if you learn better through worked examples than dense papers. He's the bridge between research and practitioners who actually ship. What you can build from this: follow his pattern. Pick one ML topic, build the cleanest "annotated" version of it on the public web, and you'll have a portfolio piece that opens doors for years. ## 10. rasbt, Sebastian Raschka Author of "Build a Large Language Model from Scratch", one of the cleanest learning paths in the field. Hands-on, clear, zero fluff. Follow him to actually understand LLMs instead of just calling APIs. Every post is a small free lesson in modern ML. What you can build from this: his book is a content treasure chest. Each chapter is a YouTube video, a thread, a workshop, or a paid cohort waiting to be made for an audience that learns better with a different teacher. Group 2: Model training and efficiency ## 11. tridao, Tri Dao Co-author of FlashAttention and Mamba, two of the most important papers of the last few years. Any fast inference setup you're using probably has his code in it. Follow him if you care about how the next generation of models will run. He's at the frontier of making them faster and cheaper. What you can build from this: any product whose unit economics depend on cheaper inference (voice agents, video pipelines, real-time personalization). The papers he ships compound directly into your margins. ## 12. albertfgu, Albert Gu Co-author of Mamba and the leading researcher on state space models, the architecture genuinely challenging transformers. His work could shape what comes after attention. Follow him to be early on the next architecture wave. He's been building toward it for years. What you can build from this: early Mamba-based products. The ecosystem is at the same moment transformers were in 2018. Whoever ships the first great Mamba-native vertical app looks prescient in two years. ## 13. TimDettmers, Tim Dettmers Built bitsandbytes and led QLoRA, which are the reason you can finetune big models on a single consumer GPU. He gave the open source world cheap finetuning, full stop. Follow him for accessible writing on quantization, hardware, and training economics. Every post saves you weeks of trial and error. What you can build from this: a productized "finetune your model on your own data, on your own hardware, in one click" wrapper. The infra exists, the consumer-grade UX layer barely does. ## 14. danielhanchen, Daniel Han Co-founder of Unsloth (with his brother Michael), which makes finetuning LLMs roughly 2x faster on consumer hardware. He keeps catching subtle bugs in major open models that the labs themselves missed. Follow him if you finetune anything at all. Every thread is a free optimization for your training loop. What you can build from this: a finetuning-as-a-service business for a vertical (legal AI, medical AI, sales AI) sitting on top of Unsloth. The cost reduction is the moat. ## 15. winglian, Wing Lian Maintainer of Axolotl, the default open source framework for finetuning LLMs. If a Hugging Face model was finetuned in the last two years, Axolotl was probably involved. Follow him for no-nonsense takes on training pipelines. He's solving the boring infra problems that make finetuning actually work. What you can build from this: turnkey Axolotl recipes for specific use cases (function-calling, structured output, domain assistants). Recipe libraries are content gold and lead magnets at the same time. ## 16. teknium1, Teknium Core researcher at Nous Research, the most respected open weights crew in the game. Hermes, OpenHermes, and a long list of finetunes that punch way above their weight. Follow him if you care about open models that aren't gatekept by big labs. He sits at the center of the open AI ecosystem. What you can build from this: products serving users whose use cases need open weights for cost or sovereignty reasons. His finetunes are production-ready substitutes for many GPT-4-class workloads. ## 17. mlabonne, Maxime Labonne Built one of the best free LLM courses online and pioneered model merging as a serious technique. His repos are some of the cleanest learning material in the space. Follow him for a real path from "I read about LLMs" to "I'm shipping finetuned models". He shows every step of the work. What you can build from this: merge models for niche domains and ship them as paid downloads or API endpoints. Model merging takes hours, distribution and packaging is where the value is. ## 18. ehartford, Eric Hartford The mind behind Dolphin, Cognitive Computations, and a long line of uncensored finetunes. Also writes some of the clearest essays on alignment, openness, and what these models really are. Follow him for an unfiltered view of where open models are heading. No safety theater, no corporate framing. What you can build from this: products serving users whose legitimate use cases get refused by closed models. Creative writing tools, research assistants, niche professional tools. The market is big and underbuilt. Group 3: Frontier labs (OpenAI, Anthropic, Cohere) ## 19. gdb, Greg Brockman OpenAI co-founder and President. He's been at the absolute center of every major model release since GPT-3, and writes some of the most candid posts about what it's like inside a frontier lab. Follow him for unfiltered signal from the most-watched company in AI. When OpenAI ships, his feed is usually the first place you see what's actually going on. What you can build from this: pay attention to what excites him about a release. The features he highlights tend to be the ones that turn into category-defining products by the time mainstream coverage catches up. ## 20. nottombrown, Tom Brown First author on the GPT-3 paper, Anthropic co-founder, now leading API engineering. He literally signed his name on the document that started this entire era. Follow him for takes from one of the very few people technically in the room for every step of the modern LLM. Posts rarely, but every post is high signal. What you can build from this: when he comments on what's working in production, treat it as priority signal. He sees the inside of one of the most-used LLM APIs in the world. ## 21. bcherny, Boris Cherny Lead engineer behind Claude Code, the coding agent half the developers reading this are using right now. Built from the inside of Anthropic, fast, and continues to push the platform forward. Follow him if you ship code with AI agents. He posts the design choices, edge cases, and patterns the rest of the agentic ecosystem will copy in 6 months. What you can build from this: tools, plugins, and workflow extensions on top of Claude Code. The ecosystem around it is still small, and he's effectively documenting the platform in public. ## 22. catherio, Catherine Olsson Mechanistic interpretability researcher at Anthropic. Her work is part of the small group actually trying to look inside the black box and understand what these models are doing. Follow her to think clearly about what an LLM is and isn't. The interpretability field is small, and she's at the center of it. What you can build from this: turn her group's research into accessible explainers, demos, or visualizers. Mechanistic interp is one of the most under-explained important areas in AI, and the audience for clear teaching is huge. ## 23. sholtodouglas, Sholto Douglas Anthropic researcher working on scaling and reinforcement learning, ex-DeepMind. One of the most thoughtful voices on what actually drives model capability. Follow him for substantive thinking on training, scaling laws, and where the field is heading. His Dwarkesh interviews are some of the best public material on modern AI training. What you can build from this: turn his interviews into structured notes, threads, or written essays. Translating long-form research conversations into shareable artifacts is a high-leverage content lane that still has room. ## 24. aidangomez, Aidan Gomez Cohere CEO and one of the original co-authors of "Attention Is All You Need", the paper that started the transformer era. He was 20 when he signed it. Follow him for a builder-CEO's view of the LLM market, especially around enterprise and non-English models. He thinks about distribution and economics differently from the US lab consensus. What you can build from this: enterprise AI products that serve markets US labs underweight (multi-language, regulated, sovereignty-sensitive). Cohere has built a real moat there and the wider ecosystem barely catches up. Group 4: Agentic and coding agents ## 25. yoheinakajima, Yohei Nakajima Built BabyAGI, the first viral autonomous agent that made everyone realize agents were a real category. He keeps shipping experimental agents that the rest of the ecosystem copies later. Follow him if you build agents. What he prototypes today usually becomes a standard pattern a month from now. What you can build from this: productize his prototypes. He ships the loop, you wrap it in a clean UI for a real user (real estate agents, recruiters, accountants) and charge for the result. ## 26. Torantulino, Toran Bruce Richards Created AutoGPT, which kicked off the entire autonomous agents wave in early 2023. It hit 100k+ stars in weeks, the kind of timing most builders only dream about. Follow him to study what it looks like to launch one of the defining repos of an era. He still ships and still has strong opinions on agent design. What you can build from this: AutoGPT-style agents narrowed to one specific outcome (qualifying leads, monitoring competitors, summarizing inboxes). The general agent went viral, the specialized agents are where the revenue is. ## 27. KillianLucas, Killian Lucas Built Open Interpreter, the tool that lets natural language actually run code on your machine. One of the first agents that genuinely felt like the future of computer use. Follow him for sharp thinking on local agents and computer control. The repo has 50k+ stars for a reason, and he keeps pushing it forward. What you can build from this: "do this on my computer" verticals. Bookkeepers, legal staff, ops teams all need natural language wrappers around tasks they already do in clunky desktop tools. ## 28. paul-gauthier, Paul Gauthier Built aider, the AI pair programmer that beat most commercial tools to the punch. It runs in your terminal, integrates with git, and stays open source. Follow him if you actually want to ship code with AI, not just demo it. His benchmarks and design choices are some of the smartest in the entire agent space. What you can build from this: aider workflows, prompt presets, and project templates packaged as a paid product or course for engineers who want pair programming dialed in by default. ## 29. AntonOsika, Anton Osika Built gpt-engineer, one of the original "describe an app, watch it appear" tools. He's now co-founder of Lovable (with Fabian Hedin), one of the hottest AI coding products on the planet. Follow him to see the full arc from viral open source repo to scaled product. He's living the journey most vibe coders are just starting. What you can build from this: services and templates that ride the Lovable wave. Component libraries, prompt packs, post-generation polish services. When the pickaxe goes mainstream, every accessory wins. ## 30. hwchase17, Harrison Chase Created LangChain, which became the default LLM application framework basically overnight. Whatever you think of the framework, he saw the category before anyone else and built the rails. Follow him if you build anything multi-step with LLMs. He's at the center of the agent tooling decisions the whole industry is making. What you can build from this: opinionated, clean alternatives to LangChain for a specific niche (voice agents, RAG over PDFs, sales workflow agents). The market accepted the framework, it didn't promise loyalty. ## 31. jerryjliu, Jerry Liu Co-founder and CEO of LlamaIndex, which owned RAG before most people knew RAG was a thing. He shares concrete, usable patterns for getting LLMs to actually answer with your data. Follow him if your work touches documents, search, or knowledge bases. Every other post is a pattern you can ship this week. What you can build from this: "talk to your X" products for specific industries (talk to your contracts, talk to your medical records, talk to your codebase). The patterns are open, the verticalization is wide open. ## 32. jxnl, Jason Liu Built Instructor, the library that made structured outputs from LLMs actually reliable. Beyond engineering, he writes the best content on AI consulting and pricing your work properly. Follow him for both engineering depth and business sense. Most builders only have one of those, he covers both. What you can build from this: a productized AI consultancy of your own. He has basically open-sourced the playbook for charging real money for AI work. Read it, adapt it, run it. ## 33. mckaywrigley, Mckay Wrigley Built Chatbot UI, the open source ChatGPT clone everyone has forked at some point. Now runs Takeoff School, a course teaching builders to actually ship with AI instead of just playing with it. Follow him for build-in-public energy and tutorials that don't waste your time. The kind of educator most "AI gurus" pretend to be. What you can build from this: your own niche fork of Chatbot UI as a paid template. Customer-support specific, sales-specific, internal-tools specific. The UI is solved, the niche packaging is not. ## 34. saoudrizwan, Saoud Rizwan Created Cline, the open source autonomous coding agent that exploded in 2024 and became one of the most-used VSCode AI extensions. Now founder and CEO of the Cline company. Follow him if you ship code with AI. The patterns Cline figures out tend to become standard in the coding agent space soon after. What you can build from this: extensions, MCP servers, and custom tools on top of Cline. The platform is open, the surrounding ecosystem is still small. ## 35. eyaltoledano, Eyal Toledano Built claude-task-master / Task Master AI, the tool that turns vague PRDs into structured, AI-executable task lists. Quietly became one of the most-used Claude Code companions. Follow him if you actually ship features with AI agents. He thinks deeply about how to break down work in a way that AI can finish without supervision. What you can build from this: opinionated Task Master configs for specific industries (mobile apps, web SaaS, ecommerce). The tool is general, the use-case-specific recipes are not. Group 5: Inference runtimes ## 36. jmorganca, Jeffrey Morgan Co-founder of Ollama (with Michael Chiang), which made running local LLMs as easy as installing an app. If you've run a model on your laptop in the last year, his work is probably under it. Follow him if local AI is anywhere on your roadmap. Every Ollama release pulls in new models the day they drop. What you can build from this: desktop products with Ollama bundled. Privacy-first AI assistants, offline research tools, on-device coaches. The packaging story is solved, the consumer apps barely exist. ## 37. jart, Justine Tunney One of the most legendary low-level engineers alive. Built llamafile and Cosmopolitan libc, the projects that let you package an entire LLM into a single executable that runs on basically any computer. Follow her to see real systems mastery applied to AI. Her writeups are gold for anyone who cares how things actually work under the hood. What you can build from this: llamafile-distributed AI tools. Single-binary downloadable products are a marketing dream and a friction killer, and most builders haven't realized this is a real distribution channel. ## 38. abetlen, Andrei Betlen Maintains llama-cpp-python, the Python bindings under a huge chunk of the local AI Python world. Quiet, but absolutely essential infrastructure. Follow him if you build Python tools on top of llama.cpp. His work is what lets the rest of us treat local inference like a normal library. What you can build from this: Python-first developer tools and notebooks targeting local inference. The ecosystem is hungry for opinionated, batteries-included tooling on top of these bindings. Group 6: Image, video, and audio AI ## 39. comfyanonymous, ComfyUI creator Built ComfyUI, the node-based interface that's quietly become the default way to do serious AI image and video generation. 111k stars on a project where the maintainer has stayed anonymous. Follow this account if you work with image, video, or any visual AI pipeline. Whatever's coming next in generative video, ComfyUI will be the surface most people use to build it. What you can build from this: vertical workflows on top of ComfyUI for specific creative roles (real estate stagers, ecommerce photographers, brand designers). The engine is open, the polished use-case kits are not. ## 40. AUTOMATIC1111, Stable Diffusion WebUI creator Built Stable Diffusion WebUI, the project that actually made AI image generation accessible to millions of regular users. 163k stars, with the maintainer staying anonymous. Follow this account for early signal on the open image and video generation stack. Their decisions shape what an entire generation of creators learn AI on. What you can build from this: tutorials, custom nodes, and packaged extensions for the WebUI ecosystem. The user base is huge, the polished educational layer is shockingly thin. ## 41. erogol, Eren Gölge Built Coqui TTS, one of the most important open source voice synthesis projects of the last few years. Now at Cantina.ai pushing voice AI further. Follow him if voice is anywhere in your stack. The voice agent wave is just starting, and people who built the foundations of open TTS will shape the next phase. What you can build from this: voice agent products for specific verticals (call centers, restaurants, healthcare scheduling). The voice quality is now production-ready, the focused vertical apps are wide open. Group 7: Tooling and infra ## 42. simonw, Simon Willison Co-creator of Django, builder of Datasette, and author of the most useful AI blog on the open web. His llm CLI is the cleanest way to talk to any model from your terminal. Follow him for daily, practical AI experiments written up clearly. If something useful happened in AI yesterday, Simon probably already wrote about it today. What you can build from this: small CLI tools and plugins on top of his llm framework. He has built a real plugin ecosystem and it is shockingly underpopulated for the demand. ## 43. jph00, Jeremy Howard Co-founder of fast.ai, one of the most generous teachers in machine learning history. Now running Answer.AI, a research lab building practical open source AI tools. Follow him if you learn better through doing than reading. Everything he ships is built so a normal developer can actually use it. What you can build from this: apply the fast.ai teaching pattern to your own niche. Hands-on, top-down, results-first. Almost every domain in AI is missing a fast.ai-style course for it, and that's a content business. ## 44. erikbern, Erik Bernhardsson Built Annoy at Spotify, then founded Modal, the serverless GPU platform half the AI startup world runs on. Also writes some of the sharpest engineering essays on the internet. Follow him if you care about ML infra economics or just want opinionated, clear engineering thinking. Every long post is worth a slow read. What you can build from this: products on Modal where unit economics demand serverless GPU. Image generation tools, batch transcription pipelines, embedding services. The infra removes 90% of the painful work. ## 45. charliermarsh, Charlie Marsh Founded Astral and built ruff and uv, the Rust-based Python tools that finally made Python developer experience feel modern. Astral was acquired by OpenAI in March 2026. Follow him if you ship Python anywhere in your AI stack. He's redefining what fast tooling means for a 30 year old language. What you can build from this: "rewrite a slow Python tool in Rust" is one of the most lucrative content and product playbooks in software right now. He showed the path. Pick a different target and walk it. Group 8: Open ecosystem (Hugging Face and Qwen) ## 46. thomwolf, Thomas Wolf Co-founder of Hugging Face (with Clem Delangue and Julien Chaumond), the company that basically created the open AI ecosystem we all rely on. Without his work, there's no easy path to any open model. Follow him for a builder's view of where open AI is heading. He was right about openness winning long before it was the consensus. What you can build from this: products that depend on the open ecosystem he helped seed. Anything where users want sovereignty, customizability, or pricing flexibility is downstream of the bet he made years ago. ## 47. patrickvonplaten, Patrick von Platen One of the core engineers behind Hugging Face's transformers and diffusers libraries, now Research Engineer at Mistral AI. If you've finetuned, run, or generated images with open models, his code ran somewhere in your stack. Follow him for technical depth on diffusion and transformer internals. He explains the things most tutorials skip over. What you can build from this: diffusers-based vertical image and video products (real estate staging, ecommerce photo generation, brand-style content engines). The library is open, the verticals are largely empty. ## 48. osanseviero, Omar Sanseviero Spent years as one of Hugging Face's most visible engineers, now at Google DeepMind. He's the rare bridge between cutting-edge research and the open source crowd. Follow him for one feed that filters the entire AI ecosystem for you. He's plugged into both sides at the highest level. What you can build from this: watch what he highlights from each release wave. The models, datasets, and tools he features early are the ones that get adopted broadly. ## 49. lvwerra, Leandro von Werra Hugging Face engineer behind TRL (Transformer Reinforcement Learning), one of the most important open source libraries for RLHF and DPO. 18k+ stars, and effectively the canonical implementation of nearly every alignment technique. Follow him if you finetune or align models. TRL is what most published alignment work reduces to in code. What you can build from this: turnkey TRL recipes for specific alignment goals (style consistency, factuality, refusal calibration). Recipes are content gold and lead magnets. ## 50. younesbelkada, Younes Belkada Spent years deep in Hugging Face transformers and PEFT, now Senior AI Engineer at TII (Falcon team) in Abu Dhabi. One of the most prolific contributors to open finetuning infrastructure. Follow him for hands-on takes on training, quantization, and adapter methods. His commits are essentially a curriculum. What you can build from this: clean walkthroughs of his PRs and PEFT techniques. He builds the plumbing, the educational layer is thin. ## 51. JustinLin610, Junyang Lin Lead on Alibaba's Qwen team, the most successful open weights project coming out of China. Qwen models routinely top open leaderboards. Follow him for one of the best windows into Chinese open source AI, a scene most Western builders sleep on. Their release cadence is brutal. What you can build from this: products and tutorials that take Qwen seriously as a Western dev. The performance is there, the bridge content for English-speaking builders barely exists. Group 9: Practitioners, writers, and dev rel ## 52. chiphuyen, Chip Huyen Wrote "Designing Machine Learning Systems" and "AI Engineering", two of the cleanest books on actually shipping ML in production. She thinks in systems, not just models. Follow her if you want to build AI products that survive contact with reality. Hype-free, practitioner-focused, every post is worth saving. What you can build from this: productized "AI engineering audits" for small companies. Her frameworks are the rubric, your job is to apply them to a specific business and charge for the report. ## 53. hamelsmu, Hamel Husain Spent years at GitHub on ML tooling, now one of the most respected voices on LLM evals and finetuning. His writing is what you read after you've moved past "wow, ChatGPT works". Follow him if you ship LLM features that need to actually be reliable. Evals, finetuning, debugging, he covers the unsexy parts that decide if you win. What you can build from this: evals-as-a-service for AI startups. Almost no early-stage AI product has real evals. The ones that survive will. Be the team that sets them up. ## 54. eugeneyan, Eugene Yan Applied scientist who turns hard ML systems lessons into clear, free essays. His writeups on RAG, evals, and recsys are bookmarked across the entire industry. Follow him for practitioner depth without academic gatekeeping. He distills years of production ML into posts you can read in fifteen minutes. What you can build from this: turn his essays into starter repos. He explains the pattern, you ship the runnable reference implementation. That pairing is one of the most valuable content artifacts on the open web. ## 55. swyxio, Shawn Wang (swyx) Coined the term "AI Engineer" and runs Latent Space, the most respected podcast and newsletter for serious AI builders. Also building smol.ai. Follow him for the cleanest weekly map of what's happening in AI. He sits at the intersection of every important conversation in the space. What you can build from this: bring his frameworks (AI Engineer, Latent Space taxonomies) into your own niche or geography. He's done the heavy mapping work, the localized application of it is wide open. ## 56. transitive-bullshit, Travis Fischer Built the agentic library (18k+ stars), one of the cleanest ways to turn APIs into AI-callable tools. Long history of high-quality open source across the React-Notion ecosystem and AI tooling. Follow him for sharp engineering taste applied to agentic tooling. His repos consistently look like products, not weekend projects. What you can build from this: vertical agents on top of his agentic library. The plumbing is there, the productized agents for specific workflows aren't. ## 57. danielmiessler, Daniel Miessler Built Fabric, an open source AI augmentation framework with 41k+ stars. Long history of clear writing on AI, security, and human-AI workflows. Follow him to think about AI as augmentation, not automation. Fabric is one of the cleanest examples of designing AI tools around real personal workflows. What you can build from this: Fabric "patterns" for your specific role or industry. The framework is general, the role-specific pattern libraries are still mostly empty. ## 58. logankilpatrick, Logan Kilpatrick Now leading Google AI Studio at DeepMind after years as OpenAI's most public dev rel face. One of the most consistent voices on what builders actually need from foundation model APIs. Follow him for builder-side signal from the inside of the API platforms you use every day. He shapes what gets prioritized. What you can build from this: tools and content that bridge gaps between API platforms (cross-provider abstractions, model comparison tooling, structured output libraries). He's openly explaining what's missing every week. ## 59. sqs, Quinn Slack Co-founder and CEO of Sourcegraph, the company behind Cody and one of the most ambitious bets in code intelligence. Strong builder energy across his GitHub history. Follow him if you build dev tools or care about code understanding at scale. His thinking on enterprise AI coding is some of the sharpest available. What you can build from this: enterprise-grade integrations and extensions on top of Sourcegraph or Cody. Big companies will pay for code intelligence tooling, and most of the ecosystem hasn't realized it yet. ## 60. steipete, Peter Steinberger Built PSPDFKit (now Nutrient), then went all-in on AI dev workflows and became one of the most influential voices on personal agents thanks to OpenClaw. His tool reviews and workflow posts are bookmarked across the industry. Follow him for the most practical, no-hype takes on actually using AI to ship software. He's a canary for what's about to hit mainstream in AI dev. What you can build from this: replicate his workflow posts in your own voice and tooling. He has effectively built a content category, and there's room for a dozen more practitioners working in it. > The bigger pattern across these 60 > Almost everyone on this list shipped something tiny first, then iterated for years. nanoGPT was one Python file. llama.cpp started as a weekend port. AutoGPT was a viral hackweek project. ComfyUI was one developer staying anonymous. They didn't predict the wave, they just kept building useful, focused things and the wave came to them. Take that personally. ## This week's move Pick three accounts from this list. One whose work you already use every day, one whose work you've heard of but never run, and one whose name was new to you when you read it here. For each, do three things in the next 7 days. Read their pinned repo's README. Run their code locally. Ship one tiny thing on top of it (a tweet thread, a fork, a wrapper, a tutorial, a screenshot of you using it). That's nine concrete actions in seven days, costing you basically nothing, that compound into a real map of the AI ecosystem and a real reputation as someone who actually engages with it instead of consuming it. Save this article. Come back to it. The list will outlive the trends. ## 相关链接 - [Meta Alchemist](https://x.com/meta_alchemist) - [@meta_alchemist](https://x.com/meta_alchemist) - [1.1K](https://x.com/meta_alchemist/status/2050834716157501710/analytics) - [karpathy](https://github.com/karpathy) - [ggerganov](https://github.com/ggerganov) - [geohot](https://github.com/geohot) - [soumith](https://github.com/soumith) - [fchollet](https://github.com/fchollet) - [ylecun](https://github.com/ylecun) - [lucidrains](https://github.com/lucidrains) - [rwightman](https://github.com/rwightman) - [srush](https://github.com/srush) - [rasbt](https://github.com/rasbt) - [tridao](https://github.com/tridao) - [albertfgu](https://github.com/albertfgu) - [TimDettmers](https://github.com/TimDettmers) - [danielhanchen](https://github.com/danielhanchen) - [winglian](https://github.com/winglian) - [teknium1](https://github.com/teknium1) - [mlabonne](https://github.com/mlabonne) - [ehartford](https://github.com/ehartford) - [gdb](https://github.com/gdb) - [nottombrown](https://github.com/nottombrown) - [bcherny](https://github.com/bcherny) - [catherio](https://github.com/catherio) - [sholtodouglas](https://github.com/sholtodouglas) - [aidangomez](https://github.com/aidangomez) - [yoheinakajima](https://github.com/yoheinakajima) - [Torantulino](https://github.com/Torantulino) - [KillianLucas](https://github.com/KillianLucas) - [paul-gauthier](https://github.com/paul-gauthier) - [AntonOsika](https://github.com/AntonOsika) - [hwchase17](https://github.com/hwchase17) - [jerryjliu](https://github.com/jerryjliu) - [jxnl](https://github.com/jxnl) - [mckaywrigley](https://github.com/mckaywrigley) - [saoudrizwan](https://github.com/saoudrizwan) - [eyaltoledano](https://github.com/eyaltoledano) - [jmorganca](https://github.com/jmorganca) - [jart](https://github.com/jart) - [abetlen](https://github.com/abetlen) - [comfyanonymous](https://github.com/comfyanonymous) - [AUTOMATIC1111](https://github.com/AUTOMATIC1111) - [erogol](https://github.com/erogol) - [Cantina.ai](https://cantina.ai/) - [simonw](https://github.com/simonw) - [jph00](https://github.com/jph00) - [fast.ai](https://fast.ai/) - [Answer.AI](https://answer.ai/) - [fast.ai](https://fast.ai/) - [erikbern](https://github.com/erikbern) - [charliermarsh](https://github.com/charliermarsh) - [thomwolf](https://github.com/thomwolf) - [patrickvonplaten](https://github.com/patrickvonplaten) - [osanseviero](https://github.com/osanseviero) - [lvwerra](https://github.com/lvwerra) - [younesbelkada](https://github.com/younesbelkada) - [JustinLin610](https://github.com/JustinLin610) - [chiphuyen](https://github.com/chiphuyen) - [hamelsmu](https://github.com/hamelsmu) - [eugeneyan](https://github.com/eugeneyan) - [swyxio](https://github.com/swyxio) - [smol.ai](https://smol.ai/) - [transitive-bullshit](https://github.com/transitive-bullshit) - [danielmiessler](https://github.com/danielmiessler) - [logankilpatrick](https://github.com/logankilpatrick) - [sqs](https://github.com/sqs) - [steipete](https://github.com/steipete) - [3:07 PM · May 3, 2026](https://x.com/meta_alchemist/status/2050834716157501710) - [1,182 Views](https://x.com/meta_alchemist/status/2050834716157501710/analytics) - [View quotes](https://x.com/meta_alchemist/status/2050834716157501710/quotes) --- *导出时间: 2026/5/3 16:12:49* --- ## 中文翻译 # 每个 AI 构建者都应关注的 60 个 GitHub 账号，助你成就卓越 AI **作者**: Meta Alchemist **日期**: 2026-05-03T07:07:53.000Z **来源**: [https://x.com/meta_alchemist/status/2050834716157501710](https://x.com/meta_alchemist/status/2050834716157501710) --- ![Image](/doc/1499/image/img_94bac68f49b4.jpg) 一份精心策划的指南，带你深入了解那些你正在使用的工具背后的人们，并提供一份实战手册，教你如何将他们的研究成果转化为你可以交付的产品。大多数自称“投身 AI”的构建者，很少会去真正构建这个领域的人的 GitHub 主页。他们只关注发布推文、演示片段和网红评论。他们错过了源头。这篇文章精选了 60 个此类账号的深度指南。是真实的人，而不是公司 Logo，映射于你每天接触的 AI 技术栈的七个层级。对于每个账号，你都会看到简短、诚实的解读：他们构建了什么，为什么这项工作很重要，以及大多数“关注这些账号”的列表都会跳过的部分——即作为构建者的你，究竟能从他们的工作中汲取什么，并将其转化为自己的东西。一个可运行的代码库，一篇可阅读的论文，一个待填补的生态空白，或是一个尚无人涉足的小型产品。收藏这篇文章。在接下来的 60 天里，每天打开一个主页。到本月底，你将拥有一张更清晰的 AI 世界地图，一份值得克隆的优质代码库简短清单，以及 AI 世界的规律和待占领的蓝海。两个月内，你将达到真正理解 AI 世界的水平。这就是旁观 AI 浪潮与真正成为其中一部分的区别。 ## 如何阅读这份列表以真正从中获益关注 60 个账号并觉得自己变得更聪明了，这其实是一个陷阱。关键在于将这些主页作为通向你可在此基础上构建的想法、模式和项目的入口，无论你是否写代码。以下是如何在不浪费时间的情况下做到这一点的方法： - 不要只是关注他们，要深入挖掘他们制作了什么。任何构建者工作中有趣的部分很少在于技术细节。而在于“为什么”。他们为什么要构建这个？他们试图解决什么问题？他们拒绝了什么？你会在他们的置顶帖子、README 文件、播客采访和旧推文中找到答案。理解“为什么”。那是经验所在。 - 寻找他们拥有的不公平优势。这份名单上的每个人之所以能发布产品，都是因为拥有某种特定的优势：对细分领域的深度知识、完美的时机、无情的专注，或者是愿意做其他人都在回避的枯燥工作。发现这种优势。然后问问自己，你拥有哪种优势，以及基于此你应该构建什么。 - 与他们的工作互动，即使你不写代码。每周挑选一位构建者。阅读他们的文章，看他们的一次演讲，如果可以的话尝试使用他们制作的工具，并对他们发布的内容留下深思熟虑的回复。你不必是工程师也能从他们那里学习。你只需要深入接触他们的风格和他们所代表的东西，而不是略读。这将给你大量关于 AI 世界的新见解。 - 把他们的工作作为你的入门套件。这份名单上的大多数东西都缺少一个有品味、有干劲或有领域知识的人可以添加的层级。在强大工具之上构建一个更好的界面。针对特定行业的工具版本。用一种更简单的方式封装他们的构建成果，以便普通人也能使用。这种“在杰出工作之上的简易层”是当前 AI 经济中杠杆率最高的举措之一，你不需要计算机科学学位就能做到。只需要在获得灵感后在终端上花足够的时间。 - 将研究转化为真正的生意。这份名单上的每一篇重要论文或代码库，都是一个伪装的产品机会。当一项新的 AI 技术出现时，最先围绕它为特定受众打包出整洁产品的几个人，通常拥有该类别。技术部分每季度都在变得更容易。品味、受众和包装才是赚钱的地方。 - 关注他们，然后关注他们关注的人。AI 生态系统比看起来要小。这份名单上的人大多互相认识，在彼此的工作基础上构建，并互相放大彼此的研究。关注这 40 个人，然后关注他们每个人关注的 5 到 10 个人。你的信息流将变成一个前沿扫描仪，而不是一个炒作过滤器。第一组：基础与经典 ## 1. karpathy, Andrej Karpathy 他是 AI 世界中为数不多的、能让你真正学到东西的老师。他发布的每一个项目，从 nanoGPT 到 llm.c 再到 micrograd，都会逐行带你理解 Transformer 的实际工作原理。关注他，让你从“使用 AI 工具”进阶到“理解模型”。一旦你了解了底层的 machinery，作为构建者，你的战斗力将提升 10 倍。你可以从中构建什么：在一个周末内自己重建 nanoGPT，然后用你自己的声音写一份教程。教授他所教的内容是目前 AI 领域转化率最高的内容角度之一。 ## 2. ggerganov, Georgi Gerganov 一位开发者让在笔记本电脑上运行大型 LLM 成为可能。那就是 llama.cpp，当其他人都在等待云 GPU 时，Georgi 几乎单枪匹马地构建了它。关注他，这是证明一个专注的构建者能发布什么产品的证据。他的代码支撑着 Ollama、llamafile 以及你已经使用的大多数本地 AI 工具。你可以从中构建什么：基于 llama.cpp 构建一个面向无法将数据发送到云端（律师、医生、会计师）的用户的垂直产品。运行时已经解决，垂直化领域尚是一片广阔天地。 ## 3. geohot, George Hotz 17 岁越狱了 iPhone，现在正在构建 tinygrad 以与 PyTorch 竞争。他公开构建产品，直播工作过程，并不在乎这一路上得罪了谁。关注他，获取原始的构建能量和对 AI 基础设施栈的反主流观点。你未必总是同意他的观点，但总能学到东西。你可以从中构建什么：针对特定用例的 tinygrad 清晰教程。这个框架很棒，但教育层级很薄弱，这个差距就是你的机会。 ## 4. soumith, Soumith Chintala PyTorch 的联合创作者，地球上几乎所有正经 AI 实验室都在用的框架。低调、技术流，受业内所有关键人物尊敬。关注他，从真正塑造深度学习基础设施的人那里，看它的发展方向。他对硬件和开放生态系统的看法总是领先于趋势。你可以从中构建什么：关注他转发支持的内容。他在六个月前指出的框架和硬件，就是现在应该进行技能堆栈的方向。 ## 5. fchollet, François Chollet 构建了 Keras，让数百万人能够接触深度学习。现在他负责 ARC 挑战赛，这是衡量模型是真正在推理还是仅仅在死记硬背的最清晰基准。关注他，获取关于 AI 是什么和不是什么的犀利、有原则的思考。当炒作周期喧嚣尘上时，他会让你保持清醒。你可以从中构建什么：围绕 ARC 构建工具。可视化工具、数据集探索器、排行榜、尝试解谜的 Agent。基准测试是新的前沿，而其周围的工具链几乎还没建成。 ## 6. ylecun, Yann LeCun 图灵奖得主（2018 年），在担任首席 AI 科学家多年后最近离开了 Meta 创办了自己的风险投资公司，也是那个大声反对“LLM 就是一切”群体的人。卷积网络的存在归功于他。关注他，作为对日常 AI 炒作周期的对冲。在很多事情成为共识之前，他早就正确了。你可以从中构建什么：读读他在怀疑什么。AI 领域的一些最大机会，恰恰存在于共识过度炒作而少数清醒声音发出警告的领域。 ## 7. lucidrains, Phil Wang 用干净的 PyTorch 重新实现了几乎每一篇重要的 AI 论文，通常在论文发布后的几天内。一个人，数百个代码库，质量都很高。如果你宁愿通过代码而不是数学来理解论文，关注他。这是在不攻读博士学位的情况下掌握新架构的最快方法。你可以从中构建什么：将他那些骨架般的实现转化为产品化的笔记本、课程或开发者工具包。他负责科学，教育和产品层则是广阔的开放领域。 ## 8. rwightman, Ross Wightman 构建了 timm，每个计算机视觉团队都在默默依赖的库。它支撑着当今生产环境视觉工作的很大一部分。如果视觉模型与你的技术栈沾边，关注他。他分享了你在别处找不到的基准、训练技巧和模型比较。你可以从中构建什么：使用 timm 作为骨干构建垂直 CV 产品（库存计数、文档解析、缺陷检测）。模型库已经解决，可部署的垂直应用还没有。 ## 9. srush, Sasha Rush 长期的康奈尔大学教授，现在任职于 Cursor Research。“Annotated Transformer”和 MiniTorch 都是他的作品，它们是 ML 中最清晰的学习材料之一。如果你通过做过的示例比通过晦涩的论文学得更好，关注他。他是研究人员与真正交付产品的从业者之间的桥梁。你可以从中构建什么：遵循他的模式。选择一个 ML 主题，在公共网络上构建它最清晰的“注释”版本，你将拥有一份能为你打开多年机会之门的作品集。 ## 10. rasbt, Sebastian Raschka 《从零开始构建大语言模型》的作者，这是该领域最清晰的学习路径之一。动手、清晰、零废话。关注他，真正理解 LLM 而不仅仅是调用 API。每一篇帖子都是现代 ML 的一小节免费课程。你可以从中构建什么：他的书是一个内容宝库。每一章都是 YouTube 视频、推文、研讨会或付费 cohorts（ cohort-based course），等待着为那些需要不同老师风格的受众制作。第二组：模型训练与效率 ## 11. tridao, Tri Dao FlashAttention 和 Mamba 的合著者，这是过去几年最重要的两篇论文。你正在使用的任何快速推理设置可能都有他的代码。如果你关心下一代模型如何运行，关注他。他处于让模型更快、更便宜的前沿。你可以从中构建什么：任何单位经济学依赖于更便宜推理的产品（语音 Agent、视频管道、实时个性化）。他发表的论文直接复利进你的利润率。 ## 12. albertfgu, Albert Gu Mamba 的合著者，状态空间模型（真正挑战 Transformer 的架构）的领先研究员。他的工作可能会塑造“注意力”之后的时代。关注他，以便在下一波架构浪潮中抢占先机。他为此已经准备了数年。你可以从中构建什么：早期的基于 Mamba 的产品。生态系统处于 2018 年 Transformer 时的阶段。谁发布了第一个优秀的 Mamba 原生垂直应用，谁在两年后就会显得有先见之明。 ## 13. TimDettmers, Tim Dettmers 构建了 bitsandbytes 并主导了 QLoRA，这是你能在单个消费级 GPU 上微调大模型的原因。他给了开源世界廉价的微调能力，仅此而已。关注他，获取关于量化、硬件和训练经济学的通俗易读的文章。每一篇帖子都能为你省去数周的试错时间。你可以从中构建什么：一个产品化的“在你自己的数据上、在你自己的硬件上、一键微调你的模型”的封装。基础设施已经存在，消费级的 UX 层几乎没有。 ## 14. danielhanchen, Daniel Han Unsloth 的联合创始人（与他的兄弟 Michael 一起），使在消费级硬件上微调 LLM 的速度提高了大约 2 倍。他不断在主要开源模型中发现实验室本身错过的细微 Bug。如果你要微调任何东西，关注他。每一个推文都是对你训练循环的免费优化。你可以从中构建什么：建立在 Unsloth 之上的垂直领域（法律 AI、医疗 AI、销售 AI）的微微即服务业务。成本降低就是护城河。 ## 15. winglian, Wing Lian Axolotl 的维护者，这是用于微调 LLM 的默认开源框架。如果一个 Hugging Face 模型在过去两年被微调过，很可能就用了 Axolotl。关注他，获取关于训练流程的务实观点。他正在解决那些让微调真正起作用的枯燥基础设施问题。你可以从中构建什么：针对特定用例（函数调用、结构化输出、领域助手）的交钥匙 Axolotl 配方。配方库既是内容金矿，也是获客磁铁。 ## 16. teknium1, Teknium Nous Research 的核心研究员，这是业内最受尊敬的开源权重团队。Hermes、OpenHermes 以及一长串实力远超其体量的微调模型。如果你关心不受大实验室把关的开源模型，关注他。他处于开源 AI 生态系统的中心。你可以从中构建什么：服务于因成本或主权原因需要开源权重的用户的产品。他的微调模型是许多 GPT-4 级别工作负载的生产就绪替代品。 ## 17. mlabonne, Maxime Labonne 构建了最好的免费 LLM 在线课程之一，并将模型合并确立为一种严肃的技术。他的代码库是该领域最清晰的学习材料之一。关注他，找到一条从“我读过 LLM 相关内容”到“我正在发布微调模型”的真正路径。他展示了工作的每一个步骤。你可以从中构建什么：为细分领域合并模型，并将它们作为付费下载或 API 端点发布。模型合并需要几小时，分发和包装才是价值所在。 ## 18. ehartford, Eric Hartford Dolphin、Cognitive Computations 以及一系列非审查微调模型背后的灵魂人物。他还撰写了一些关于对齐、开放性以及这些模型本质的最清晰的文章。关注他，获得关于开源模型走向的未经过滤的观点。没有安全作秀，没有企业框架。你可以从中构建什么：服务于那些合法用例被封闭模型拒绝的用户的产品。创意写作工具、研究助手、利基专业工具。市场巨大且未被充分开发。第三组：前沿实验室（OpenAI, Anthropic, Cohere） ## 19. gdb, Greg Brockman OpenAI 联合创始人兼总裁。自 GPT-3 以来，他一直处于每一次重大模型发布的绝对中心，并撰写了一些关于前沿实验室内部情况的最坦率的文章。关注他，获取来自最受瞩目的 AI 公司的未经过滤的信号。当 OpenAI 发布产品时，他的信息流通常是你看到实际情况的第一个地方。你可以从中构建什么：关注他对某个发布中感到兴奋的部分。他强调的功能往往会在主流媒体追赶上时，转化为定义类别的产品。 ## 20. nottombrown, Tom Brown GPT-3 论文的第一作者，Anthropic 联合创始人，现在领导 API 工程。他确实是在开启整个时代的文档上签了名的人。关注他，听取极少数在每一步现代 LLM 发展中都在技术现场的人之一的观点。发帖很少，但每篇都是高信号。你可以从中构建什么：当他评论在生产环境中什么有效时，将其视为优先信号。他看到了世界上使用最广泛的 LLM API 之一的内部。 ## 21. bcherny, Boris Cherny Claude Code 背后的首席工程师，这是一半正在阅读本文的开发人员正在使用的编码 Agent。从 Anthropic 内部构建，快速，并持续推动平台发展。如果你使用 AI Agent 交付代码，关注他。他发布了设计选择、边缘情况和模式，其余的 Agent 生态系统将在 6 个月内模仿。你可以从中构建什么：基于 Claude Code 构建的工具、插件和工作流扩展。它周围的生态系统仍然很小，他实际上是在公开记录这个平台。 ## 22. catherio, Catherine Olsson Anthropic 的机制可解释性研究员。她的工作是试图窥视黑盒内部并理解这些模型在做什么的小组的一部分。关注她，以便清晰地思考 LLM 是什么和不是什么。可解释性领域很小，而她处于中心地位。你可以

🔗 相关文档