Image by ChatGPT
More than a year after the release of DeepSeek R1, another Chinese AI company, Moonshot AI, has once again sent shockwaves through the AI community with the launch of Kimi K3. The company has even drawn accusations from advisers to the White House, alleging that it distilled advanced AI models developed by Anthropic and relied on restricted NVIDIA hardware located outside China to train its latest model.
Whether Kimi K3 has, to any extent, appropriated American AI technology will undoubtedly remain the focus of much public attention. Yet what interests me far more is the architecture behind the model itself. Rather than relying on a single monolithic neural network, Kimi K3 employs a Mixture-of-Experts (MoE) architecture consisting of 896 specialised “experts.” During inference, only 16 of these experts are activated for each token, dramatically reducing computational requirements and inference costs.
The significance of MoE extends well beyond engineering efficiency. On the one hand, it highlights how computational optimisation has become a decisive factor in today’s AI arms race. On the other hand, it illustrates an important principle of intelligence itself: within a large and complex system, it is neither necessary nor desirable to activate every subsystem for every task. Different situations require different forms of expertise. For problems that depend upon specialised knowledge, the most appropriate experts should take the lead, while unnecessary participation by others may actually reduce the quality of the outcome.
This insight carries profound implications not only for artificial intelligence, but also for human society and political institutions.
Modern parliamentary democracies are likewise composed of hundreds of legislators. Yet under the logic of party politics and dynamics, every representative is expected to vote on virtually every issue, regardless of whether they possess relevant expertise. Kimi K3 invites us to imagine a different possibility. Suppose we had a “multi-agent parliament” composed of 896 specialised representatives. Rather than operating solely according to the principle of majority rule, deliberation would become issue-specific. Different members would concentrate on the domains in which they possess genuine competence, engaging in repeated discussion, negotiation and analysis without requiring everyone to participate equally in every decision.
Of course, this does not mean that MoE can immediately replace parliamentary democracy.
The objective of today’s MoE is to combine the opinions of relevant experts in order to generate the most accurate prediction. Parliamentary democracy, however, pursues a fundamentally different goal. It is not merely a search for a homogeneous optimal consensus, but an institutional framework for discovering diverse—and often conflicting—perspectives. Environmental protection may conflict with economic growth. The interests of vulnerable communities may diverge from those of the majority. Future generations may hold priorities very different from our own. While democratic institutions seek to reconcile such differences wherever possible, disagreement itself is not a sign of failure. On the contrary, it is an indispensable feature of democracy.
Looking ahead, if we one day seek to constitutionalise AI, we may not require substantially larger foundation models or ever more specialised intelligent agents — those technological ingredients are already rapidly falling into place. What matters far more is whether each AI agent is capable of independent reasoning; whether it can challenge the assumptions of other agents, critically evaluate competing proposals, and balance perspectives drawn from different domains of knowledge. Most importantly, disagreements between AI agents must be made fully visible to citizens. Transparency about competing viewpoints is essential if AI-assisted governance is to remain accountable and democratically legitimate.
From this perspective, the significance of Kimi K3 lies not only in its Mixture-of-Experts architecture, but equally in its decision to release the model as open weights.
Today, the development of frontier AI remains concentrated in the hands of a small number of tech giants. Open-weight models allow universities, research institutes and civil society organisations to participate more actively in adapting, improving and deploying these models for diverse local contexts. AI development can therefore become substantially more decentralised and locally embedded.
This seemingly modest decision to release model weights may carry profound political consequences. It shifts part of the initiative away from organisations possessing virtually unlimited financial and computational resources, creating opportunities for a far broader range of stakeholders to develop alternative AI applications and governance models. In doing so, it encourages a more open and pluralistic AI ecosystem. Like today’s highway networks or the Internet, AI itself may gradually become a form of shared public infrastructure, upon which different institutions, communities and societies build very different paths of technological development.
The true significance of Kimi K3 therefore lies in the combination of two innovations: the pluralistic potential of its MoE architecture and the democratising potential of its open-weight strategy. AI no longer needs to be imagined as a single superintelligence controlled by a small number of actors. Instead, it can evolve into a diverse ecosystem of specialised intelligences capable of flourishing in many different institutional settings — more importantly, this technological architecture offers a concrete foundation upon which transparency, accountability and governance may eventually be built. In this sense, it provides part of the technical groundwork for AI Constitutionalism as a genuinely utopian political project.
One final clarification is essential: nothing in this argument should be interpreted as an endorsement of Moonshot AI, DeepSeek or any other particular company, nor does it imply that they will necessarily advance transparency or accountability in AI governance. Rather, these developments demonstrate a possibility. If companies in Mainland China can pursue such technical directions, there is no inherent reason why liberal democracies cannot do the same.
Once one country demonstrates a viable model, others will rapidly adapt and improve upon it, thus shaping the trajectory of a shared future for all humankind. Over the next ten to twenty years, the constitutional choices societies make about AI may prove decisive and foundational.
The race is therefore not merely technological; it is institutional. I have been repeating such a statement since 2018 in my various books (in Chinese). It is once again be highlighted in my recently published Demotopia: an Exploration of Democracy and Artificial Intelligence (2026).