The Inevitable Public Good
The release of Deepseek's R1 is a wake-up call
In late January, a Chinese-built large language model called DeepSeek-R1 thrust itself onto the global stage. The initial frenzy, much of it within AI circles, was warranted: R1 promised an open-source, affordable competitor to proprietary ‘reasoning’ systems, most famously, OpenAI’s o1. By many available metrics, R1 comes impressively close to matching o1’s ability to tackle advanced tasks in chemistry, mathematics, and coding. For anyone following the trajectory of generative AI, it’s a déjà vu moment: as soon as a groundbreaking proprietary model appears, an open-source counterpart arrives, sooner, and more capable, than expected.
This pattern is not new in technology, but it’s accelerating. The time gap between the release of cutting-edge proprietary AI models and the unveiling of open-source challengers is shrinking. More importantly, the barrier to replication for these models, while not trivial, is nowhere near insurmountable. After a new technique or architecture is introduced, it’s often a matter of reading papers, studying observable model behaviours, and applying the same or similar training heuristics. Soon enough, a capable, and sometimes equally groundbreaking, model emerges from a different lab.
Information, by its nature, defies containment. Models and methods, once revealed in the wild, are not easily locked away. The “secret sauce” behind each new AI innovation, be it a novel training technique or an ingenious architecture tweak, tends to become obvious1 in hindsight. As soon as something is out there, replicating or refining it is increasingly a question of resources and will.
DeepSeek-R1’s emergence is the latest exhibit in this argument. With R1, we’re seeing what appears to be a near-clone of o1’s step-by-step, chain-of-thought reasoning capability. Not only has it arrived faster than many expectd, but it also robustly underscores how porous AI research has become. A dedicated, well-funded research team in China, working on this as a side project, read the relevant publications, made some smart inferences, gleaned enough of the training methodology, bought the necessary hardware, and promptly built a rival. If we keep repeating this scenario, and there’s no reason to believe we won’t, the trajectory seems clear: any new, privately funded, secretive AI “miracle” will soon have a low-cost or open-source version.
I’ve written that AI can, and should, be a public good, a resource that is both non-rivalrous and non-excludable, like clean air or public radio: everyone could use it without depleting it, and no one could be effectively fenced out. And I’ve argued that we need strong policy intervention to ensure that we can reach this dream. But I’m starting to think I was misguided.
It’s becoming clear that open AI, that can serve as the basis for a public good, is not just a desirable outcome, it may be inevitable. Once you have enough people with the know-how, enough countries and organizations with the means, enough open discussions, and enough desperate graduate students looking for a new challenge, the genie can’t be shoved back into the bottle. AI will soon be embedded everywhere, from edge devices to massive cloud platforms, unstoppable by trade secrets or regulatory sign-offs. As soon as a new model makes a splash, the reverse-engineering begins.
Far from needing regulation to enforce AI as a public good, it may, in fact, be the case that no regulation could possibly prevent it. This is great news for those who believe proliferation will ultimately empower humanity, but it has some sobering implications for those who crave a top-down approach to AI safety.
There is no shortage of governments, government agencies, IGOs, and others searching for ways to control or govern AI. But a heavy-handed regulatory approach may be doomed to frustration if the moment a new cutting-edge (and carefully governed and regulated) system is revealed, an uncontrollable open alternative quickly follows. This doesn’t mean that safety, fairness, and equity should be cast aside; on the contrary, it means we need to think very hard about how we can effectively promote AI safety in a world of unavoidable proliferation.
Welcome to the age of the unstoppable diffusion of AI models.
My earlier stance — that we ought to create policy to ensure AI becomes a global public good — still has moral merit, but I got the details wrong: we probably don’t need top-down legislation to force AI into the public domain. Indeed, it may be that sheer technical inevitability will carve that path on its own. When the knowledge, talent, and resources for building powerful models are spread across continents and industries, absolute control is an unattainable dream.
Given the rapid progress from closed to open that we see in DeepSeek-R1, it’s time to adjust our thinking. AI, in its many forms, is almost certainly fated to be a public good, whether we like it or not.
This phenomenon also aligns with computer scientist Rich Sutton’s “bitter lesson”: that broad, simple methods, scaled up with enough data and compute, almost always outperform meticulously engineered solutions.