Emad Mostaque’s bet: open small models, not one giant brain

Sarah Guo and Elad Gil opened No Priors in February 2023 with a guest who had spent barely a year turning a London startup into the gravitational center of open image models. Emad Mostaque’s Stability AI was already known for Stable Diffusion, the system that takes a natural-language prompt and returns a picture, but he described a wider bet: language, voice, video, and biology, all released as open tools. The hosts wanted to know how a company that young had changed the landscape, and whether small open models could actually outperform the closed giants.

Mostaque’s favorite adoption statistic came a few minutes in. On GitHub Stars, he said, Stable Diffusion had overtaken Ethereum and Bitcoin *combined* — projects that took ten years to get there — in three or four months. Widen the comparison past crypto to all of open-source software, AI or otherwise, and he still called it the fastest-growing project ever. The reason, in his telling, was not hype. It was that a two-gigabyte file had become a new kind of translation layer, and the inference cost of turning text into an image was far lower than anyone would have guessed five years earlier.

Mostaque’s first career was computer science, then hedge funds, where he says he became one of the largest investors in video games and artificial intelligence. The real turn was personal. About twelve years before the taping, his son was diagnosed with autism and doctors told him there was no cure or treatment. He assembled a team, ran an AI literature review of existing therapies, and did biomolecular pathway analysis of neurotransmitters looking for drug-repurposing options. Working with physicians, they found a few levers that might be causing the condition. His son later went to mainstream school.

He went back to the fund, won some awards, and then tried to aim the same toolkit at the world. The first project was non-AI education tablets for refugees through Imagine Worldwide, his co-founder’s charity. When COVID arrived in 2020, he saw another multi-system condition that existing institutions — which extrapolated the future from the past — could not keep up with. He stood up an AI initiative with the World Bank, UNESCO, and others to understand what caused COVID and to publish the findings. The project hit what he called an “institutional wall.” Two lessons stuck: the new models were far beyond prior technology, and they were locked inside a handful of labs.

That is the moral origin of Stability. “We’re now going to very narrow superhuman performance,” he told the hosts, “and everyone should have access to that.” A personal quirk made image models feel urgent. Mostaque has aphantasia: he cannot visualize anything. Ask him to imagine a beach and he gets the word and a feeling, not a picture. He had assumed “imagine” was a metaphor until OpenAI released CLIP and generative models could suddenly be steered by text. Recovering from COVID himself, he built a small model for his daughter. She sold the output as an NFT for $3,500 and donated the money to India COVID relief. He started writing personal checks: salaries for EleutherAI developers, compute for model creators, funding for Disco Diffusion notebooks, even early grants to Midjourney.

About fifteen months before the interview he tried to turn those scattered communities into a common good. His first idea was a DAO, or a DAO of DAOs. After a week he decided that would not work. Commercial open source became the vehicle instead — a way to build aligned technology across images and everything beyond them, and to offer researchers a third path besides a couple of academic GPUs, a job at Big Tech, or a brutal startup. EleutherAI, Stability, and the other communities he incubated were meant to be that option.

Classical open source, Mostaque argued, is a misunderstanding of this technology. You do not invite a hundred people to train a foundation model. It is usually five to ten people, a supercomputer, and a data team. Transformers already accounted for about 80 percent of the research, and scaling laws were pulling talent, data, and GPUs into private labs. The output, though, behaves like infrastructure: a small file, a deterministic transform, cheap inference. That is why he thinks the right analogy is Red Hat and public goods, not a chaotic bazaar of weekend contributors.

His forecast for a closed-only world was grim and specific. Big companies would keep the stack private for four or five years. Then someone would defect, open-source the weights, and collapse the market by giving away everyone else’s complement — the Gmail play, applied to models. Governments would still need a Japan model and an India model, and commercial labs selling B2B would not build them. Stability’s job, as he framed it, was to coordinate those national efforts, sit like Switzerland in the middle, and tap infrastructure budgets. Self-driving had drawn about $100 billion; web3, hundreds of billions; 5G, trillions. AI, on that scale, was a rounding error, and he wanted it spent as equitably as possible.

How does an open lab compete with trillion-dollar compute budgets? Mostaque’s answer was national supercomputers. He said Stability had more compute available than Microsoft or Google because it could tap machines private firms cannot touch, was helping multiple countries build exascale systems, and had just received a seven-million-hour grant on Summit, one of the fastest supercomputers in the United States. Academia cooperated more readily with open projects than with commercial agreements. Fully loaded, he claimed, their compute cost was close to zero, while closed labs had to raise tens of billions. Sarah later noted that the original Stable Diffusion training run was only 150,000 A100 hours — almost a rounding error in the frontier-model arms race.

The most schematic claim in the conversation was a four-stage map of deep learning, each with a different winner. Stage one is pretraining the large base model; Mostaque thought those foundation weights would all be open, with Stability as coordinator. Stage two is instruction and reinforcement learning — Flan-PaLM, InstructGPT — which needs careful annotation and is a private-company strength. Stage three is domain fine-tuning on specialist data. Stage four is the real workflow, with a human in the loop.

Google’s Med-PaLM paper, then only a few weeks old, was his worked example. PaLM, at 540 billion parameters, scored about 50 on medical questions. Flan-PaLM, the instructed version, reached 70. Med-PaLM, trained on medical information, hit 92 percent — “human level,” in his words. The last step was putting that system into clinical practice with doctors watching. Private firms, he said, should live in the instruct-to-human-in-the-loop band. The base models should be international and national infrastructure, available to everyone.

He added a composition thesis that he would repeat on AGI and on democracy: combining models beats training one giant multilingual system. Meta’s Cicero, which used eight language models together to beat humans at Diplomacy, was the citation. “Multiple models interacting with each other is the way, full stop,” he said later. The next decade, in this picture, is not one lab owning the stack. It is open bases, private adapters, and swarms of specialists that call each other.

Mostaque called Stability the only independent multimodal AI company in the world, set against research labs such as Meta’s FAIR and DeepMind. The point was not to clone every closed model. Stable Diffusion had been optimized for a 24GB VRAM GPU; a distilled version, he said, would run in a couple of seconds on an iPhone, with Intel helping on the neural engine. The emphasis was models that are customizable and editable, not merely bigger. He thought that was a better kind of scale than stacking layers.

DeepMind’s Chinchilla paper was exhibit A: a 67-billion-parameter model matching GPT-3 at 175 billion, which he read as a data-quality result more than a “train more tokens” result. On the language side, EleutherAI’s GPT-Neo, GPT-J, and GPT-NeoX had been downloaded 20 million times, topping out around 20 billion parameters. CarperAI’s trlX instruction framework was already public. They were training models up toward 100 billion parameters, which he considered Chinchilla-optimal enough to stand in for an open ChatGPT or Claude, and then to spawn sector models. He was also funding RWKV, an attention-free architecture, with a 14-billion-parameter version on the way.

Language, he conceded, is semantically denser than images and more unforgiving on accuracy, so it still wants more compute. Google’s Muse paper had shown a Transformer can replace a VAE, which meant image generation did not even need diffusion. The optimization cycle for language had barely started. OpenAI’s 1.3-billion-parameter InstructGPT beat 175-billion-parameter GPT-3; Google’s 3-billion-parameter Flan-T5 beat GPT-3 on some tasks. Stable Diffusion itself went from 5.6 seconds per image on an A100 on day one to 0.9 seconds, with 25 frames a second in sight — a 100x speedup from shipping the weights and letting people care. He expected language to follow the same curve.

That is also his AGI story. He said he does not care about AGI “except for it not killing us.” Stability’s mission is intelligence augmentation: “build the foundation to activate humanity’s potential.” Closed labs, in his caricature, stack more layers, swallow the internet, then press the result back toward human preference — “a massive elder god of weirdness.” The alternative is the one the Stable Diffusion community already demonstrated: hundreds of fine-tunes, merged, spontaneously learning perfect faces and fingers. Millions of models that reflect human diversity, then brought together, seemed to him more likely to stay aligned. His last-minute prediction closed the loop: small models will “outperform large models massively,” and ChatGPT-level systems will run on phones within five years.

On the business side, Mostaque named two edges: media, and private, regulated data. GPT-3 or PaLM weights will almost never land on-prem; an open model will. Private data, he said, is worth more than public data, which is why Amazon SageMaker was the go-to-market for nations and the largest corporates who will one day need to train their own models. That is a craft, not a science. Losses explode. “There’s only like 100 people who can train models in the world.” Stability’s open weights become the benchmark; the paid work is sitting inside a customer’s data center, training a model they control.

Computational biology was the breakthrough he was most willing to over-index on. Stability was a key driver of OpenFold, with a paper claiming much faster ablations than AlphaFold, plus DNA diffusion for sequence outcomes and BioLM for chemical reactions. He promised an open Med-StableGPT as a counterpart to Med-PaLM, and argued that misaligned incentives had stalled federated-learning efforts such as MELLODDY. Elad raised a fifty-year-old objection: Stanford’s MYCIN expert system already beat physicians at infectious-disease calls in the 1970s, and none of it reached the clinic. Mostaque blamed infrastructure barriers, and said open complex systems plus human-in-the-loop assistants were the first real way around them.

The leapfrog case was education and health in places with no legacy stack. Stability was deploying four million tablets to every child in Malawi, with a hope of reaching hundreds of millions of children the following year, working with governments and multilaterals to build care from the bottom up. The ratios he cited were one teacher for 400 children and one doctor for ten thousand. “I am certain that system will outperform anything in the West within five years.” In refugee camps, the most basic AI was already teaching literacy and numeracy in 13 months on one hour a day. Elad compared it to East Asia leapfrogging to mobile; Mostaque agreed, and said Asian governments were preparing to put tens of billions into generative AI while British headmasters called him over Christmas asking for a ChatGPT homework policy. His advice was to stop assigning essays. Western schools that made students handwriting papers without computers were, in his view, going backwards.

Media was the other obsession, and it predated Stability. His first job was organizing the British Independent Film Awards and reviewing for the Raindance Film Festival; he later became one of the world’s largest video-game investors. Visual communication had always been hard. If the acceleration continued, his mother sending memes would be the floor, not the ceiling, and art therapy — “the most effective form of therapy” — could be given to everyone. The metaverse, as he understood it, would not be one world everyone visits. It would be everyone sharing their own worlds. Descript already let you edit a podcast by editing the transcript; NeRFs let you reshoot a scene; movie studios were saving millions of dollars just by putting Stable Diffusion in production. The exclusive Eros deal was training video models on hundreds, soon thousands, of Bollywood films, with audio models and A. R. Rahman in the conversation.

Mostaque reached for Shannon when the conversation turned to politics. Information is valuable insofar as it changes state; political information, increasingly manipulated by stories, had widened the divide. National and public models could become a filter layer: translate a complex issue so a five-year-old, a ten-year-old, or a particular voter can actually understand it. Elad cited the novel *Lady of Mazes*, in which AGIs synthesize group opinion into virtual representatives. Mostaque connected it back to Cicero. People already trust Congress less than cockroaches, he joked, citing a poll, and they will hand more judgment to machines. Those coordinators have to be robust, because society is slowly ceding personal authority to whoever runs them.

He does not believe technology is neutral. EleutherAI had already split: capabilities work stayed with Stability, alignment research went to London-based Conjecture. Large models, he insisted, should be treated as fiction engines, not fact engines. Compressing terabytes into a few gigabytes cannot be lossless; what you lose is factualness, what you keep is principle-based analysis. Anything above a compute threshold should be registered like dual-use biology, with an international team to test systems the way Anthropic had begun to with constitutional models. One of the optimal ways to stop AGI, he noted, is to build your own first — a bind that makes an arms race likely, which is why he wanted multilateral action immediately rather than later.

Advertising was the near-term hazard. These models write convincing prose, and his sister-in-law’s company Sonantic had already rebuilt Val Kilmer’s voice for a documentary. Combined with engagement-maximizing recommenders at Meta and Google, synthetic ads could be co-opted across the political spectrum. He wanted mandatory identification of AI in advertising, opt-out and attribution for training data — Stability was building those mechanisms, including its own OpenCLIP and a CLIP-G checkpoint that week — and he thought the EU’s CE-mark plan, which makes model creators responsible for outputs, pointed the wrong way. A $200,000 deepfake-detector prize was withdrawn because a detector and a generator just bounce off each other. Attribution metadata, with Adobe and the Content Authenticity Initiative, seemed more durable than an arms race over detection. Slaughterbots and defense uses, he admitted, need an international compact he did not know how to write.