Democratization of generative AI
Community resource and epistemological infrastructure
2024-06-15 — 2025-07-20
Wherein communities are shown to retrofit consumer GPUs and public datasets—exemplified by Stable Diffusion and LAION’s image corpus—to train and deploy capable generative models outside corporate walls.
A surprising phenomenon is unfolding in the world of artificial intelligence: despite the formidable barriers of technical know-how, expensive infrastructure, etc., communities are figuring out how to DIY large generative models. Since I grew up on cypherpunk DIY myself, I can’t help being captivated.
1 A Folk History of Open-Source Generative AI
The drive to democratize generative AI is a response to a field initially dominated by a few large, well-funded technology companies. Early, powerful models were often kept proprietary, accessible only through paid APIs. But hackers wanna hack.
That said, you need a few more things than a shell account and a copy of Phrack to build a large neural network. Let’s unpack the key ingredients that users needed to wrest from corporate control to make it happen.
The Big Four: Compute, Data, Algorithms, implementations To cook up one of these models, you need a mountain of GPUs (compute), a big dataset, and the secret sauce of the training algorithm. Also, you need implementations of the algorithms that work with the other components, because training big models is ahrd. For a long time, only the tech giants had all of these. The open-source story is about communities figuring out how to source, share, or substitute for each of these.
Image Generation: When models like DALL‑E first appeared, they lived in a corporate walled garden. The breakthrough for the rest of us was the 2022 release of Stable Diffusion. It was (is) a potent, open-source model that let anyone with a decent gaming PC generate high-quality images. This was possible because a non-profit called LAION had done the hard work of creating a massive, open dataset of images scraped from the web, feeding the model. The genie was out of the bottle.
The Language Model Pirates: In the world of text, the story is similar. When GPT‑3 was kept under lock and key, a scrappy crew of volunteer researchers on a Discord server formed EleutherAI in 2020. Their explicit goal was to replicate the model and give it away. They bootstrapped their own massive dataset, “The Pile,” and released a series of influential open models, proving that a distributed collective could take on the giants. The plot thickened when corporate players like Meta started open-sourcing their own powerful models like LLaMA, throwing a high-performance engine over the wall for the community to tinker with.
The Jugaad Ethos and Hacking on a Shoestring: The whole scene is underpinned by what some call the Jugaad philosophy: a kind of frugal, resourceful, get-it-done innovation that thrives when you don’t have a billion-dollar data centre. It’s about making do and hacking together what you need from what you have. This spirit is what’s globalizing AI, breaking it free from geographically concentrated power. The poster child for Jugaad might be Masakhane, a grassroots, pan‑African collective building NLP models for African languages that are completely ignored by Big Tech. Their name literally means “we build together,” and they operate as a distributed network of researchers, proving you don’t need a centralised HQ to do foundational work. You also see it in applied AI, like with Kenyan startup Tenakata, which uses AI on basic mobile phone data to create credit scores for smallholder farmers who have no formal banking history. It’s not about building a massive new LLM; it’s about a clever, resource‑light hack to solve a real, pressing problem. This is innovation that happens out of necessity, not excess.
And Then There Are the Fans: You can’t talk about DIY without mentioning the fans. [Fandoms](./fandoms.qmd( are tried-and-tested distributed innovation communities, so of course they’ve been at the forefront. Two prime examples show how this goes.
First, there’s the My Little Pony fandom. When the show Friendship is Magic ended in 2019, a segment of the fanbase decided they just weren’t done with it. This led to the grassroots Pony Preservation Project, which originated on 4chan’s /mlp/ board with a clear mission: to meticulously collect, clean, and align voice data from every episode. This wasn’t just archiving; it was building a high-quality dataset specifically for AI training. That dataset became a key component for 15.ai, a landmark text-to-speech platform that launched in 2020 and let anyone generate shockingly accurate dialogue in the characters’ voices. That led to much fun and equally much drama. The result was an explosion of fan-made content, including fully AI-voiced animations, effectively continuing the show through sheer force of will and technical ingenuity.
Second, you have communities like AI HUB. While the MLP fans focused on beloved fictional characters, AI HUB became the go-to Discord community for those interested in the much more ethically murky world of cloning the voices of real people, particularly musicians. This community doesn’t just use voice cloning tools; it’s a hotbed for the sharing and development of the tools themselves—from pre-trained models of specific singers to new techniques for separating vocals from tracks to create clean training data. It serves as a decentralised R&D and distribution network, dramatically lowering the barrier to entry for a controversial but technically demanding application of AI. It’s a perfect example of how this technology spreads not through press releases, but through the irrepressible, and sometimes unsettling, drive of communities to build the things they want to see in the world.
Fan projects frequently test the edges of copyright law, and fans often have smaller budgets than copyright holders or large tech companies, so these projects often get mired in drama or keep a low profile.
Mandatory AI safety qualm: the secrecy that keeps AI voice-cloning fan communities viable despite copyright law probably also enables weapon-building zealot communities to operate despite other laws.
2 The Players
A motley crew of non-profits, academic-adjacent labs, grassroots collectives, and a few companies playing the open-source game.
2.1 Hugging Face
- What they do: If this scene has a distro outlet, it’s Hugging Face. They started as a chatbot company but pivoted to become the central repository for nearly everything in open-source AI. They host the models, the datasets, and have built the essential
transformerslibrary, a kind of universal adapter that lets you actually use all this stuff without pulling your hair out. - Their role: They are the essential infrastructure. Nearly every other organisation on this list, from EleutherAI to DeepSeek, hosts their models on Hugging Face. They are the Schelling point where everyone meets.
2.2 Allen Institute for AI (AI2)
- What they do: Bankrolled by the late Microsoft co-founder Paul Allen, AI2 is the “benevolent rich uncle” of open AI research. They are a non-profit that tackles the hard, long-term research that corporations won’t and universities can’t. They built foundational tools like AllenNLP for language research and the Semantic Scholar search engine.
- Their role: They bring serious resources with an open-source ethos. With their OLMo model, they went a step further than most, releasing not just the model but all of its training data and code. They’re not just giving away the cooked meal; they’re publishing the entire recipe.
2.3 EleutherAI
- What they do: The pirate radio station of punk AI is EleutherAI. Born on Discord, this grassroots collective of volunteer researchers did the seemingly impossible: they replicated the closed-off GPT models and gave them away. They scraped the web to create “The Pile” dataset when no one else would.
- Their role: They are the spiritual godfathers of the open LLM movement. They proved it could be done, without corporate funding or a centralised lab, inspiring and directly enabling much of what came after.
2.4 LAION
- What they do: LAION is a German non-profit of data hoarders for the people. Their mission is simple but monumental: create and release the massive datasets that AI models need to learn. Their LAION-5B dataset, with its 5.8 billion image-text pairs, is the most famous example.
- Their role: Models need food. LAION provides the public buffet. Without their work, the open-source text-to-image revolution, particularly Stable Diffusion, simply would not have happened. They are a critical upstream provider of the ecosystem’s most basic resource.
2.5 BigScience
- What they did: BigScience wasn’t a permanent group, but a year-long “Woodstock for AI researchers” (This punk metaphor is getting overwrought, sorry). Coordinated by Hugging Face, it brought over 1,000 researchers together in a global jam session to build BLOOM, a massive multilingual language model.
- Their role: They provided a proof of concept for a radically different way of doing research: massively open, global, and collaborative. It was a powerful counter-narrative to the idea that only secretive, siloed corporate labs could build state-of-the-art AI.
2.6 Deepseek
- What they do: Deepseek is a Chinese company that has been dropping bombshells into the open-source world. They release models for coding and chat that are so good they make the proprietary players sweat, and they often do it with licences that allow for commercial use.
- Their role: They are pushing the performance ceiling of what we expect from open-source models. They represent a fascinating trend of commercial companies using aggressive open-sourcing as a core strategy, changing the competitive landscape for everyone.
2.7 DAIR Institute
- What they do: Founded by Dr. Timnit Gebru after her high-profile exit from Google, DAIR is an independent research institute, claiming “We publish interdisciplinary work uncovering and mitigating the harms of current AI systems, and research, tools and frameworks for the technological future we should build instead.”
- Their role: They are the conscience of the scene. DAIR provides the critical research and ethical frameworks to encourage the community to avoid rebuilding the same exploitative systems they are trying to escape.
2.8 Agora
- What they do: If EleutherAI is the pirate radio, Agora is the zine. It’s a hyper-grassroots collective of engineers and creators organised on Discord, seemingly driven by individuals like Kye Gomez. Their focus is on the bleeding edge: multi-agent systems, real-time learning, and augmenting AI reasoning.
- Their role: Agora represents the DIY, “let’s try it and see” punk rock ethos. They are radically open, sharing projects before they even work, all in the service of accelerating learning and collaboration at the fringes.
I bet there are other projects like Agora; it has the distinction of being the one I ran into.
2.9 Oumi
- What they do: Oumi is building the open-source workshop. While others provide the finished models, Oumi is creating an all-in-one platform and Python library to streamline the entire process of building your own, from curating data to fine-tuning and deployment.
- Their role: They are trying to democratize the process of creation. If Hugging Face is where you get the finished models, Oumi wants to give you all the tools to build your own from scratch.
3 Democratizing training on consumer hardware
4 Incoming
Kimi K2: Open Agentic Intelligence is an open-source AI model for agency
Kimi K2 is our latest Mixture-of-Experts model with 32 billion activated parameters and 1 trillion total parameters. It achieves state-of-the-art performance in frontier knowledge, math, and coding among non-thinking models. But it goes further — meticulously optimized for agentic tasks, Kimi K2 does not just answer; it acts.
i sensed anxiety and frustration at NeurIPS’24 – Kyunghyun Cho
LMSYS Chatbot Arena: Live and Community-Driven LLM Evaluation | LMSYS Org
cf Stability, RWKV