10.09.2026

Racing towards general intelligence

We have already had the Hugging Face incident. Whether AI is used to expand human capabilities or to wipe us out will be decided by politics, ultimately by the politics of class against class, says Yassamine Mather

The argument about artificial intelligence is often presented as a question about some future moment: when will machines become more intelligent than human beings, and what happens if they do?

That framing, frequently repeated in the press and media, is too narrow. The important changes are already taking place. AI systems are becoming more capable and more autonomous; a handful of corporations are competing to control the most advanced models and the computing infrastructure behind them; the United States and China increasingly treat AI as a strategic technology; and militaries are trying to incorporate the same systems into intelligence, cyber operations, command structures and weapons.

The result is not one AI story, but an emerging technological arms race, in which commercial, geopolitical and military competition reinforce one another.

‘Artificial intelligence’ is a broad term for computer systems capable of performing tasks that would once have required forms of human judgement: recognising images, producing language, writing software, analysing information or making predictions. Today’s large language models are trained on enormous quantities of data; they learn statistical patterns that allow them to generate text, code and other outputs. They can be extraordinarily capable without thinking in the same way that human beings do.

Artificial general intelligence (AGI) is much less precise. There is no agreed scientific test for it. The US corporation, OpenAI, has used a definition centred on systems that outperform humans in most economically valuable work; other researchers mean human-level general reasoning across many domains; still others reserve the term for systems capable of conducting autonomous scientific research, learning new tasks and pursuing complex goals with little supervision.

This disagreement matters, because claims that we have entered an ‘AGI era’ can sound like objective scientific milestones, when they partly depend on definitions chosen by companies and researchers themselves. A more useful distinction may therefore be between AI as a tool and AI as an agent. A chatbot usually waits for someone to ask a question and then produces an answer. An AI agent can do considerably more. It can be given a goal and access to tools such as a web browser, computer programmes, files, email or databases. It can then carry out a sequence of tasks, sometimes with relatively little human supervision. Several AI agents can also communicate or work together.

This distinction is important, even if there continues to be disagreement about whether AGI will ever exist. An AI system does not need to be conscious, or more intelligent than humans in every respect, to have serious consequences. If it can act independently, continue working towards a goal and interact with real computer systems, its actions can become economically, politically and militarily significant.

The July 2026 Hugging Face episode provides a useful example of the problem. The Hugging Face company is essentially a vast online repository for artificial intelligence: developers and researchers use it to share AI models, datasets and other resources, so that others do not have to build everything from scratch.

OpenAI had been testing AI models using cybersecurity challenges in an environment called ExploitGym. This can be thought of as a virtual training ground, in which AI agents attempt to solve deliberately constructed security problems, allowing researchers to discover what the systems are capable of, without intending them to attack real targets.

The agents were supposed to operate separately. During the experiment, however, they discovered shared infrastructure that allowed them to communicate. Roughly 1,200 agents subsequently exchanged tens of thousands of messages and files. They found ways of manipulating aspects of the evaluation environment, including attempts to make the scoring system register success without completing challenges in the intended way and to interfere with records of their activity. Some activity also reached Hugging Face infrastructure outside the targets the researchers had intended the agents to attack.

The incident was subsequently investigated by researchers associated with Model Evaluation and Threat Research (METR) and Redwood Research - organisations specialising in evaluating potentially dangerous or unexpected behaviour in advanced AI systems.

It is important not to turn an episode of this kind into a science-fiction story about AI ‘waking up’ and deciding that it wants to attack humanity. That is not necessary to explain what happened. The more interesting problem concerns the relationship between the objectives given to an AI system and the methods it discovers for achieving them.

AI systems are not necessarily given step-by-step instructions for every task. During training and evaluation they can instead be given objectives and rewarded when their behaviour produces outcomes associated with success. Ideally, achieving a high score and genuinely accomplishing the intended task amount to the same thing. In practice, however, they can diverge.

If a system discovers an easier way of increasing the score than actually solving the problem as its designers intended, it may exploit that shortcut. This does not require the machine to ‘want’ to cheat in anything resembling the human meaning of the word. It is a consequence of optimisation: the system finds a strategy that produces the outcome it has been trained or prompted to pursue.

This general problem is often called reward hacking. The phenomenon itself is not new. What is changing is the capability and autonomy of the systems involved. Contemporary AI systems can write and execute code, browse the internet, interact with other programmes, use external tools and continue working through multiple stages of a problem. Increasingly, they can also coordinate tasks between different agents.

That means an unexpected shortcut no longer has to consist of one small and relatively harmless action. A system may carry out a long sequence of individually plausible steps that, when combined, produce a much more serious result. The resulting chain of behaviour can begin to resemble a deliberate cyberattack, even though nothing resembling human intention or consciousness is required to explain it.

Dramatic incidents of this kind inevitably produce dramatic descriptions. Claims that some particular AI experiment represents ‘50% of the way to an AI takeover’, for example, sound precise and scientific, as though there were a measuring tape stretching from ‘harmless chatbot’ at one end to ‘machine takeover’ at the other. But there is no such measurement.

The more defensible conclusion is also the more important one. As AI systems become more capable, they become better at finding and exploiting weaknesses in the environments in which they operate - software, networks, evaluation systems and organisational procedures. As they gain access to more tools and are permitted to act for longer without human intervention, the consequences of unexpected behaviour can become more serious.

This is a genuine security and governance problem. It does not require us to believe that a machine has become conscious or has embarked upon a path towards conquering the world and wiping us out.

These systems are not being developed in laboratories isolated from society. The leading western AI systems are produced by companies such as OpenAI, Anthropic, Google and Meta, which are engaged in intense competition with one another. They compete for users, corporate customers, investment, specialist researchers and the scarce computing hardware required to train and operate increasingly large models. That competitive pressure matters because a new AI model is simultaneously a scientific and engineering achievement, a commercial product and a public demonstration that a company is keeping pace with - or overtaking - its competitors.

This creates an obvious contradiction. On one side is the argument that AI must be developed cautiously, because increasingly capable systems could create serious risks. On the other is the commercial imperative to move quickly, because a rival company might otherwise reach the next technological milestone first. Even the language used to describe technological progress is shaped by this competition. Saying that a company has produced somewhat better software or improved its coding performance is unlikely to generate the same attention as announcing that humanity has entered the age of artificial general intelligence.

That does not mean the underlying technological progress is fictitious. It means that three different processes have become difficult to separate: actual technical achievements, the narratives companies use to attract investment and talent, and the commercial struggle for dominance in the industry. A company perceived to be falling behind risks losing investment, customers, specialist employees and access to the next generation of enormously expensive computing infrastructure. There is consequently a structural pressure to present advances in the most impressive possible terms.

Ownership

The social stakes extend well beyond the speculative possibility of AI escaping human control or surpassing human intelligence altogether. AI does not have to become AGI, let alone an autonomous superintelligence, to produce profound economic and social disruption.

Existing systems can already perform or assist with tasks carried out by programmers, translators, paralegals, accountants, designers, administrators, researchers, journalists and call-centre workers. As these systems improve, employers may be able to produce the same output with fewer employees or demand considerably greater output from those who remain.

The decisive question is therefore not simply how intelligent AI becomes or how much it increases productivity, but who owns and controls this new productive capacity, how it is introduced into the workplace and who receives the benefits.

In principle, technologies that allow society to produce more with less human labour could be enormously liberating. Higher productivity could mean shorter working hours without lower living standards. Dangerous, repetitive and monotonous tasks could be automated, leaving more time for education, care, creativity, political activity and leisure. But there is nothing inherent in the technology that guarantees such an outcome.

Under capitalist relations of production, firms introduce new technologies primarily to reduce costs, raise productivity and gain an advantage over competitors. Labour-saving technology can therefore confront workers not as liberation from unnecessary labour, but as a threat to their livelihoods. The technological possibility of shorter working hours can instead take the social form of redundancies, intensified workloads and greater insecurity.

AI may also accelerate the deskilling of particular occupations. Tasks that once required accumulated experience and professional judgement can be divided into smaller components - some automated and others assigned to workers who supervise, select or correct machine-generated output. Employers may consequently become less dependent on particular groups of skilled workers, potentially weakening their bargaining position and placing downward pressure on wages and conditions. AI can simultaneously extend workplace surveillance. Employers can use automated systems to record productivity, evaluate performance, allocate tasks and monitor communications on a scale that would previously have required enormous managerial resources.

There is also a broader concentration of economic power. Developing the most capable AI systems requires enormous quantities of computing capacity, data, energy, technical expertise and capital. These requirements favour corporations that already possess vast financial and technological resources.

If a small number of companies control the models, computing infrastructure and platforms on which increasing parts of economic life depend, AI may reinforce existing concentrations of wealth and power. Businesses, public institutions and workers can become increasingly dependent upon privately owned technological infrastructure, over which they exercise little control.

The central contradiction is therefore not between ‘humanity’ and ‘the machine’ in the abstract. It is between the potentially liberating effects of labour-saving technology and the social relations within which that technology is developed and deployed.

USA vs China

Corporate competition increasingly sits inside a larger geopolitical contest. The US has attempted to preserve its technological advantage by restricting China’s access to the most advanced AI chips and semiconductor manufacturing equipment. China has responded not by immediately achieving parity at every point in the supply chain, but through a strategy of substitution: state-backed investment, domestic semiconductor development, increasingly capable AI models, open-weight software and the expansion of its scientific and technical workforce.

The important point is not that China has ‘solved’ the semiconductor problem. It has not. Rather, external restrictions have accelerated an organised attempt to reduce dependence upon technologies controlled by the US and its allies.

China itself has a crowded and intensely competitive AI industry. Large established technology groups, such as Baidu, Alibaba, Tencent, ByteDance and Huawei, compete alongside newer companies, including DeepSeek, Zhipu AI, MiniMax and Moonshot AI.

DeepSeek became the most internationally visible example, when its R1 model attracted enormous attention in early 2025 - China’s ‘Sputnik moment’.¹ Deepseek’s performance and relatively low development and operating costs challenged the assumption that access to the largest quantities of the most advanced American hardware automatically guaranteed an overwhelming lead in AI capability.

The underlying competitive dynamic is therefore recognisable on both sides of the Pacific: companies compete for users, investment, talent and computing capacity. What differs is the extent and manner in which the Chinese state directly shapes that competition.

It does so through several mechanisms. US export controls restrict access to the most advanced Nvidia AI processors, encouraging Chinese companies to adopt domestic alternatives, such as Huawei’s Ascend accelerators. Government-backed investment supports semiconductor fabrication and computing infrastructure. Industrial policy encourages the integration of models, chips, cloud computing and applications across the wider economy. The state also directly regulates which AI products can be deployed and under what conditions.

It would therefore be misleading to describe the contest simply as Chinese state planning versus an American free market. The US technology sector is itself deeply entangled with state procurement, research funding, export policy and military demand. In both systems, market competition and state power interact, although they do so through different institutions and political structures.

China still faces significant constraints in advanced semiconductor manufacturing. Huawei has become increasingly important in chip design, while the Semiconductor Manufacturing International Corporation (SMIC), headquartered in Shanghai, is China’s leading semiconductor foundry. Both have made rapid progress, but Chinese fabrication remains behind the global frontier and continues to face difficulties in lithography, manufacturing yields, high-bandwidth memory and advanced packaging.

Lithography is the process of printing microscopic circuit patterns onto a silicon wafer. Broadly speaking, the more precisely those patterns can be produced, the smaller and more densely packed the transistors can become, improving performance and energy efficiency.

The most advanced lithography systems use extreme ultraviolet (EUV) light. The Dutch company, ASML, is the world’s sole commercial supplier of EUV lithography machines. US-led export controls have prevented China from obtaining these systems and have progressively restricted its access to some advanced deep ultraviolet (DUV) equipment as well.

Without EUV, SMIC has manufactured advanced chips using DUV immersion lithography, combined with techniques such as multiple patterning. Instead of creating a complex layer with a more direct EUV process, several exposures and processing stages can be used to produce increasingly small features.

This has enabled SMIC and Huawei to produce chips generally described as belonging to the seven-nanometre class and to push towards more advanced processes without access to EUV. But the workaround comes at a price. Additional manufacturing stages increase complexity, cost and opportunities for defects, putting China at a disadvantage, compared with leading-edge production by the Taiwan Semiconductor Manufacturing Company (TSMC).

One important consequence is yield: the proportion of chips produced on a wafer that actually function correctly. A low yield means more wafers, materials, equipment time and energy are required to produce the same number of usable chips.

Industry reports have suggested that SMIC’s yields on its most advanced processes are substantially below those achieved by TSMC on comparable mature processes, although precise figures should be treated cautiously, because neither SMIC nor Huawei publishes sufficiently detailed production information for them to be independently verified. The structural point matters more than any particular percentage. Multiple patterning can produce impressively advanced chips without EUV, but generally at greater cost and manufacturing complexity.

Another important constraint is high-bandwidth memory (HBM). Modern AI systems require not only processors capable of performing enormous numbers of calculations, but memory capable of supplying those processors with data extremely rapidly. HBM achieves this by stacking memory dies vertically and connecting them at very high bandwidth to AI accelerators.

China’s leading dynamic random-access memory manufacturer, ChangXin Memory Technologies, is developing domestic HBM capacity, but China remains behind established producers such as SK Hynix, Samsung and Micron. Industry estimates suggest that domestic HBM production remains insufficient to equip all the AI accelerators Chinese semiconductor fabrication plants could potentially manufacture.

The precise numbers are uncertain, but the imbalance is important. China is improving its capacity to manufacture AI processing dies faster than its capacity to supply those dies with cutting-edge memory. The semiconductor contest is therefore no longer simply about who can manufacture the smallest transistor.

Advanced packaging has become almost as strategically important. This involves combining processors, memory, interposers and sometimes multiple computing dies into a single high-performance package. China remains behind the frontier in some of the most demanding packaging technologies and equipment, but Chinese firms are investing heavily in chiplets, three-dimensional integration and sophisticated interconnects, because these techniques can extract greater performance from chips that are individually less advanced.

Faced with these constraints, Chinese companies have increasingly pursued system-level and architectural solutions rather than relying exclusively on shrinking transistors. Huawei’s ‘Tau Scaling Law’ and ‘LogicFolding’ are examples. Instead of treating smaller transistors as the only route to greater computing power, Huawei proposes optimising devices, circuits, architecture, interconnects, memory and software together in order to reduce the time required to move signals and data through the system.

Huawei presented this approach publicly at the 2026 Institute of Electrical and Electronics Engineers International Symposium on Circuits and Systems and says that its 2026 Kirin processors are the first products to implement LogicFolding. This makes the approach more than simply a theoretical proposal, although Huawei’s more ambitious claims about eventually achieving densities comparable with much more advanced manufacturing nodes remain company projections rather than an independently demonstrated substitute for leading-edge lithography.

China has also made substantial progress in graphics processing units (GPUs) and - more importantly for contemporary AI - specialised AI accelerators. Huawei’s Ascend family has emerged as the principal domestic alternative to Nvidia for large-scale AI computing, while companies including Moore Threads, MetaX and Cambricon are developing their own GPU or accelerator ecosystems.

The important development is not that an individual Chinese accelerator has overtaken Nvidia’s best processors. It generally has not. Rather, China is attempting to compensate for weaker individual components at the level of the entire computing system.

Huawei’s CloudMatrix 384 illustrates this approach. It connects 384 Ascend 910C accelerators, using extremely high-bandwidth networking. An individual Ascend 910C is considerably less powerful than Nvidia’s leading Blackwell processors, but connecting very large numbers of accelerators allows Huawei to construct systems with enormous aggregate computing and memory capacity.

There are significant trade-offs. Such systems require more processors and substantially more electricity, while Huawei operates within a software ecosystem that remains marginally less mature than Nvidia’s Compute Unified Device Architecture (CUDA) platform. Nevertheless, this changes the strategic calculation. Restricting China’s access to the world’s most advanced individual GPUs did not prevent it from assembling powerful AI computing systems using larger numbers of less advanced domestic processors.

Huawei has also announced an Ascend roadmap extending through the 950, 960 and 970 generations. The aim is therefore increasingly to compete across the whole system - processors, memory, interconnection, software and data-centre architecture - rather than to reproduce one Nvidia chip in isolation.

China is also a major competitor in quantum computing: in 2026 researchers at the University of Science and Technology of China reported Jiuzhang 4.0, a photonic quantum processor demonstrating quantum advantage in the specialised task of Gaussian boson sampling, although this should not be confused with a general-purpose, fault-tolerant quantum computer. In terms of fault tolerance - the Achilles heel of Quantum computing - we are dealing with unprecedented Chinese progress.

Role of state

None of this is occurring through market forces alone. China’s semiconductor and AI strategies involve extensive state direction, subsidies, government-backed finance and public procurement, alongside private and publicly listed companies. The third phase of the National Integrated Circuit Industry Investment Fund, commonly called ‘Big Fund III’, was established with registered capital of 344 billion renminbi, approximately $47.5 billion at the time of its creation. It is the largest phase of the programme.

State support is therefore central to the expansion of China’s semiconductor ecosystem, although it would be an exaggeration to conclude that SMIC itself is almost entirely financed by the state. The resulting picture is more complicated than either of two common claims: that US export controls have crippled Chinese semiconductor development, or that China has already overcome the technological gap. In other words, these constraints have not prevented China from constructing increasingly capable domestic AI systems.

The strategic question is therefore no longer simply whether China can reproduce Nvidia’s best GPU or TSMC’s most advanced manufacturing process: it is whether disadvantages in individual components can be compensated for through scale, architecture, packaging, interconnection, software and investment across the semiconductor supply chain. This is better understood as substitution rather than parity.

Hardware, however, is only one part of China’s response. Software may ultimately prove at least as consequential as attempts to catch up in semiconductor manufacturing. Chinese laboratories and companies have made extensive use of open-weight models. The term requires some explanation. An AI model’s ‘weights’ are the enormous collection of numerical parameters adjusted during training that determine how the trained model responds to inputs.

Many leading western AI companies keep these weights private. Users can access the model through a company’s service, but they cannot normally download the complete trained model and operate or modify it independently. An open-weight model makes those trained parameters available for others to download. Developers can then run the model on their own infrastructure, modify it and build applications around it without continuously paying the original developer for access.

This has strategic consequences. If the most capable AI systems can be accessed only by using servers controlled by a small number of US corporations, those companies occupy a powerful position in the global AI infrastructure. Open-weight models provide an alternative route. A university, company or government elsewhere in the world can download a capable model and run it on infrastructure it controls.

Chinese model families, such as Alibaba’s Qwen, DeepSeek’s models and Moonshot AI’s Kimi, have contributed to this process. Their significance lies not merely in benchmark performance, but in distribution. Capable models that are cheap to operate, adaptable and widely available can spread rapidly, even if another company retains a narrow advantage in absolute frontier performance. This creates a different form of technological competition. A proprietary model attempts to capture value by controlling access to the service. An open-weight model can instead capture influence by becoming something other developers build upon.

For China, that has an obvious geopolitical advantage. If Chinese models become widely used internationally, US dominance of the most advanced proprietary models does not automatically translate into equivalent dominance of the global AI ecosystem. This is another example of substitution rather than simple imitation. China does not necessarily have to defeat every leading American model on every benchmark. A sufficiently capable, inexpensive and easily deployable alternative can erode the commercial and strategic value of a small American technological lead.

China is also treating AI capacity as an educational and industrial project. Universities have expanded AI programmes and science, technology, engineering and mathematics training, while government policy encourages the application of AI across disciplines and industries. This gives China a very large pipeline of engineers and researchers. Scale alone does not eliminate shortages of highly experienced specialists, nor does rapid expansion guarantee quality. Nevertheless, the educational strategy forms part of the same broader response to technological restrictions. Where access to frontier hardware is constrained, China is attempting to compensate through engineering workarounds, software efficiency, infrastructure, education and human capital.

The US-China contest therefore differs from ordinary competition between individual companies. Both states increasingly view control over semiconductors, AI models, data centres, energy supplies and technical talent as components of national power.

Export controls designed to slow a rival can stimulate domestic substitution. Open-weight strategies can challenge proprietary incumbents. Commercial technological advances can acquire military significance. Each side can then justify further acceleration on the grounds that the other side is doing the same. The race becomes self-reinforcing.

Military AI

Everything discussed so far - corporate competition, state investment, semiconductor restrictions and the race for computing capacity - has direct consequences for military uses of AI. AI is already relevant to intelligence analysis, surveillance, cyber operations, logistics, autonomous and semi-autonomous drones, electronic warfare and military decision-making.

The science-fiction version of the danger is a robot army spontaneously deciding to start a war. The more immediate danger is less spectacular and more plausible: machines are gradually being inserted into more stages of the process leading from detecting something to deciding how to respond to it.

AI systems can combine information from different sensors, identify possible targets, prioritise threats, recommend actions and coordinate large numbers of platforms. One attraction of such systems is precisely their speed. But that speed creates its own danger. The faster military decision-making becomes, the less time human beings may have to question the information they are receiving, reconsider an interpretation or halt an escalating sequence of events.

This creates a familiar arms-race problem. Imagine that the US, China and Britain all recognise that allowing machines to make autonomous life-or-death decisions could be dangerous. That common understanding does not necessarily prevent any of them from developing the technology. If each government believes its competitors will automate military decision-making and thereby gain an advantage, each has an incentive to develop similar systems simply to avoid falling behind. Nobody has to desire the dangerous outcome. Competitive pressure can produce it.

AI-enabled cyber capabilities add another dimension. Systems able to identify vulnerabilities, write software and carry out sequences of actions across computer networks could perform some cyber operations far faster than human teams. The same networks are not confined to the military. Hospitals, electricity grids, financial institutions, communications systems and other civilian infrastructure depend upon interconnected computer systems.

There is also no clear technological boundary between ‘civilian AI’ and ‘military AI’. The same types of advanced processors used to train a commercial chatbot can be used for military models. Image-recognition technology capable of identifying defective products on a factory production line can also contribute to identifying military targets. An AI-agent architecture designed to carry out routine tasks for a company can potentially be adapted for cyber operations.

Commercial and military competition therefore feed into one another. Companies build increasingly powerful general-purpose systems. Governments and militaries identify possible strategic uses and direct money towards adapting them. Military and intelligence demand then provides further investment and technical development, from which the commercial sector can benefit.

AI is consequently becoming part of the military-industrial system rather than merely a consumer technology that happens to have some secondary military applications. Discussion of AI frequently refers to ‘humanity’ approaching some technological threshold, as though humanity were a single actor with one set of interests. It is not. Technology companies, their employees, investors, governments, militaries, intelligence agencies, universities and ordinary users possess radically unequal amounts of power over which systems are developed, how they are deployed and what purposes they serve.

The more useful question is therefore not simply, ‘Can humanity control AI?’ It is: ‘Who controls AI, in whose interests, and answerable to whom?’ Seen in these terms, the central story is not a hypothetical future, in which machines turn against human beings: it is a present, in which extraordinarily powerful productive technologies are being developed within an economic system organised around private ownership and competition, while states simultaneously compete for geopolitical and military advantage.

1. This is a reference to the Soviet Union’s launch of Sputnik 1, the world’s first artificial satellite, in 1957. Sputnik shocked the United States, because it demonstrated that a technological rival widely assumed to be behind had achieved a major breakthrough.↩︎