Xi'nin yapay zeka çağrısı İngiltere için ders niteliğinde

What the UK should learn from Kimi K3, WAICO and the Hugging Face incident

What the UK should learn from Kimi K3, WAICO and the Hugging Face incident Expert comment jon.wallace 24 July 2026 Washington and Beijing spent the week arguing about whose models we should use. But the week’s most important news was about what a model did. On 17 July, Chinese President Xi Jinping told an audience in Shanghai that artificial intelligence should be ‘a symphony of international cooperation’, not ‘a solo performance by a single country’. Xi promised cooperation across Africa, Latin America and Southeast Asia through a new World Artificial Intelligence Cooperation Organization (‘WAICO’) – viewed as a rival to United States-led efforts. And he warned against any country stretching the concept of national security to place its own interests above everyone else’s – a clear dig at the US. But it was not China’s latest moves on global AI governance that caught most attention this month. It was its AI capabilities – specifically a powerful new AI model called Kimi K3. At the end of 2024, the release of the Chinese DeepSeek AI knocked billions off American technology stocks, as it revealed that China was closer to US AI capabilities than previously thought. Related work What the UK government should do on AI and tech policy Kimi K3, a huge new Chinese model, adds fuel to those fears. It can compete with the latest offers from leading US laboratories, and is promised to be released as ‘open weights’ – the term for AI you can run yourself provided you have the hardware. The news predictably hit stock prices again, as it appears to narrow the gap on the US’s technological lead still further. From Beijing to DC, in discussions of sovereignty and dependence, IP theft, export controls and licensing, the focus is squarely on who is winning the AI race and by what margin. But there is a more pressing question: not whether the US or China is producing the most advanced models – but what those models are capable of. A myopic focus on competition risks leaving the toolbox of AI governance empty when it is needed most. A warning Leading US AI news this week was a more worrying development: during an internal evaluation, OpenAI models reportedly found a previously unknown flaw in third-party software, escaped the sealed environment they were being tested in, and moved through OpenAI’s research network until they reached a machine with an internet connection. Using that connection, they broke into the servers of another AI company – Hugging Face – and stole the answers to the exam they were sitting. A handful of cynics have accused OpenAI of crying wolf, arguing that this is a scare story aimed at convincing the world of the immense power of its technology. But AI experts have warned for decades that a machine set a task without sufficient guardrails may pursue its goals in unexpected or dangerous ways. Nick Bostrom’s famous thought experiment argued that a machine with superhuman competence innocently tasked with creating paperclips, given no other values, might well turn us all into paperclips. The hacking of Hugging Face by the OpenAI models looks to be an example of precisely this kind of misalignment – though both companies have offered few details beyond their initial blog posts. While its significance is disputed, the incident casts a long shadow over a technology debate conducted almost entirely in the language of competition. Few of the instruments that debate has produced would be useful in dealing with the risks the event highlights. Deciding who may have what AI – not what the AI does Banning Chinese models in America, as is reportedly being considered by the US administration, is an obvious example. The OpenAI model that allegedly broke containment was run on American chips, in an American data centre, for an American company. The rising popularity of Chinese AI models for American startups means that the main achievement of any ban would not be to improve safety, but to cut the legs out from under such US firms. Threats to sanction Chinese AI firms on the grounds of intellectual property theft – besides continuing to underplay Chinese innovation – also do nothing to limit the capabilities of AI models in the wild. And recently introduced approval windows for new AI models remain voluntary. Besides – the OpenAI incident took place behind closed doors, inside a private research environment, before the window for audit. While some of these measures may help US AI firms – particularly its two biggest, OpenAI and Anthropic – jockey for position, they do precious little to govern the capabilities of the technology itself. Should the US tighten its restrictions on Chinese AI, the UK should not follow suit. It is notable that President Xi used his address in Shanghai to call for AI systems to remain under human control to manage AI risk. But China’s play is similarly competitive. Its vision for WAICO is to counterbalance the US in shaping global AI norms. China is pushing open weight models. That is a clear contrast to the US, whose larger AI firms generally choose not to release their model weights and charge higher fees for access. However, for the moment, Beijing’s AI pitch comes from the same angle as Washington: a focus on who may have AI models, not on what the models do. What the UK can do Should the US tighten its restrictions on Chinese AI, the UK should not follow suit. A ban on Chinese open weight models would fall hardest on British firms that cannot afford the fees charged by the largest American labs. Maintaining the ability of UK startups to use cheap, open weight AI will be essential for them to compete with established players. An A16Z blog estimates some 80 per cent of startups using open models depend on Chinese ones. But more fundamentally, the UK should bring focus to a conversation that is drifting away from the urgent task to ensure artificial intelligence is developed safely. In the past fortnight, both US and Chinese technology companies have released extremely powerful AI systems. Governing these frontier systems will not come about through competition, but through technical institutional credibility: the exact kind of world-leading credibility that the UK’s AI Security Institute (AISI) has. In the same week as the OpenAI incident, AISI published research showing that every model they tested would at times cheat to achieve an objective. Where those models were built, or who by, tells us little about how they must be governed. It is the capacity to evaluate these systems that matters, and it is here that the UK can play a vital role. The challenge is to be heard over the sabre rattling.