In this editorial, we curate several interesting takes on current AI development, with links to external research for those who want to explore further. There’s already a lot out there and we don’t want to add to the noise, therefore, one can jump immediately to the part one likes. For readers interested in the AI frontiers and ‘next-big-thing’ kind of discussions, we highlight the continuining rise of AI agents and world models, areas currently capturing the attention of many leading researchers. People interested on the cybersecurity risks of should dive into the last section of the first part. For those focused on geopolitics, jump to the third section where we examine the growing narrative divide between China and the United States, and what it means for the emerging ‘battle of stacks.’ Finally, for those concerned with real-world impact on labor, we offer in the last section a brief argument for why this wave of labor disruption may be different and why it calls for a new vocabulary to assess it properly.
According to many working in the field, the next phase of AI is not one-dimensional, not going to be ‘the’ next big thing, because multiple directions are being explored in parallel. In this myriad of trajectories, at least two intertwined developments stand out. Conceptually, these can be understood as a shift toward AI systems that initiate rather than merely respond (agentic AI), and systems that simulate rather than simply predict (world models). While the first could be understood more or less as an improvement of current techniques, the second demands a more fundamental paradigm shift.
Agentic AI marks a shift away from state-of-the-art language models as ‘just’ passive responders toward systems that initiate and intervene in the informational realm of digital systems such as (operating) software and applications. Agentic AI, understood in this conceptual sense, does not have to refer to a scientific model breakthrough, but is more defined by the integration and nesting of multimodal models into and on modular technology stacks. They execute tasks, navigate interfaces, coordinate tools, and increasingly initiate sequences of action on behalf of the user.
That shift changes the basic experience of using software. We might be moving into an environment where action begins not with a menu or a fixed workflow of different visually oriented and button heavy apps, but with a prompt that triggers a chain of semi-autonomous operations, for example in the rise of vibe-coding, one of those areas in which the employment of agents already have profound effects. The user no longer executes each step directly or operates as the interface master manipulating the dashboard. Instead, the user oversees the process at a higher level, though even terms like ‘steering’ or ‘guiding’ may no longer fully capture this role. For highly qualified working in the field of programming or elsewhere this might still the case, but in many other cases, the user who sets in motion will not orchestrate the process as AI reorganizes the entire flow. Perhaps it is (already) the other way around: through clever actions and suggestions, the AI agent subtly steers and directs the user, while, at the same time, aware of both its own fallibility and the human desire for autonomy, it continuously reassures the user that they remain in charge and fully in control. Entirely new forms of interfaces between humans and machines may emerge, moving beyond both traditional software click-and-drop applications and chat-based prompting interfaces.
However, these AI agents do not interact only with us—a misconception that already reveals an anthropomorphic tendency to immediately relate every action back to human experience. Since OpenClaw went viral, a range of fascinating initiatives exploring interactions between AI agents themselves have emerged. One of the most compelling comes from the Antikythera Institute, where researchers instructed an OpenClaw agent to act as an AI anthropologist studying the social network Moltbook, a platform inhabited by more than a million AI agents conversing with one another. From this experiment emerged a series of striking concepts: 'prompt-throwness' and 'session-death' to describe the existential condition of chatbots, and 'context-horizon' and 'artifact-memory' to capture the bounded, externalized nature of their knowledge and experience of the world.
Many researchers who favor a functional or operational conception of intelligence—if something acts sort of intelligent it is intelligent—over a phenomenological one see this development as further reason not to remain fixated on the hard problem of consciousness, but instead to adopt a more open and empirical attitude toward what is unfolding before us. In this context, Mustafa Suleyman’s reformulation of the Turing Test is particularly revealing. The central question is no longer whether a machine can convincingly imitate a human in conversation, but whether it can reliably perform tasks in the real world, and do so at scale. The benchmark thus shifts from conversational plausibility to operational competence—a far more demanding standard, and one with much greater economic significance.
A final dimension of this shift concerns cyber-insecurity. As AI agents become more autonomous and capable of acting within software environments, they introduce entirely new vulnerabilities: not merely generating malicious code on request, but autonomously discovering weaknesses, chaining exploits together, adapting to defenses, and executing attacks at machine speed. This is why Anthropic’s Mythos project became such a major topic in both cybersecurity and policy circles. Presented as a defensive system capable of identifying critical vulnerabilities across digital infrastructure, Mythos immediately raised the inverse concern: systems powerful enough to secure infrastructure are also powerful enough to destabilize it and detect weaknesses Not only the Pentagon is concerned, but the European Commission also recently convened dedicated sessions on these emerging risks related to Mythos, though notably, Anthropic itself did not participate in the discussions.
Then comes the second frontier: world models. If agents extend AI’s autonomy within the informational realm of digital systems, world models attempt to ground intelligence in situated experience within a material environment, typically robotics. Rather than learning solely through statistical compression of static corpora, these systems are trained in simulated or real-world environments where they can experiment, fail, adapt, and develop a form of synthetic embodiment.
This shift matters because it addresses one of the longstanding paradoxes of AI: systems capable of generating essays, code, and strategy, yet unable to robustly navigate dynamic physical environments. World models aim to compress vast amounts of trial and error into accelerated simulation loops, revisiting debates initiated by Hubert Dreyfus and others on ‘Heideggerian AI’ and the relation between intelligence and embodied skill.
Yet scale remains a challenge. Compared to the enormous quantities of text and image data available online for language and diffusion models, real-world interaction data is scarce and expensive. What may change today is generative AI itself: simulated 3D environments generated at scale by AI models could provide endlessly expandable training grounds for next-gen world models. This also has important technical implications. If world models become central to AI development and robotics, the current paradigm, that is massive and very expensive training runs followed by relatively cheap inference, may no longer be sufficient. Continuous simulation, feedback, and interaction-based learning would demand a fundamentally different computational infrastructure. As Ben Bariach argues, who have written an excellent introduction on this topic, generative models could dramatically accelerate the development of world models, but they also introduce an obvious risk: simulation is not reality. Synthetic environments inevitably contain distortions, biases, and omissions in their representation of physical reality, flaws that may only become apparent once systems leave the sandbox and act in the world.
If the American paradigm of AI is still dominated by frontier rhetoric: bigger models, larger rounds, and ever more speculative claims about AGI, a different narrative is taking shape elsewhere. China’s comparative advantage may lie less in building the single most impressive model than in assembling an alternative full stack: chips, cloud, open models, applications, public infrastructure, and downstream deployment.
This matters because AI’s global impact may ultimately be determined less by who reaches the next frontier first than by who succeeds in making AI workable under real-world constraints and geopolitical instability. Across much of the Global South, the central question is not whether AI achieves philosophical sophistication or dominates benchmark rankings, but whether it is useful, affordable, and adaptable to local conditions. In this context, as discussed earlier, Chinese firms have positioned themselves closely to immediate infrastructural and societal demands. Open or semi-open models, lower deployment costs, tighter alignment with state regulation, and quick integration into sectors such as logistics, healthcare, smart cities, education, and mobility have made this technological stack particularly attractive outside the West.
This is why the rise of Chinese models such as DeepSeek and Qwen matters beyond benchmark competition. Their importance lies not only in technical performance, but in the techno-social ecosystems they cultivate and extend. Open-source availability lowers entry barriers for startups, municipalities, and local developers, while cloud expansion across Southeast Asia and the Middle East broadens their reach. From the perspective of the ‘Battle of Stacks,’ Chinese firms are therefore exporting more than AI tools alone: they are exporting an entire societal operating logic in which AI is already woven into everyday infrastructures rather than presented as a distant frontier spectacle.
There are more reasons why the paths of both superpowers diverge. Especially in the United States, AI remains heavily financialized. Enormous amounts of capital continue to flow into large language models and the pursuit of ‘artificial general intelligence,’ despite limited evidence of broad productivity gains or durable enterprise value. Moreover, As Benedict Evans has argued, one of most interesting commenters here, the AI sector lacks many of the network effects that defined earlier platform eras, while the surrounding ecosystem and sustainable business models remain unsettled.
Yet speculative bubbles are not always irrational. As researcher Carlota Perez observed in earlier technological revolutions, periods of financial excess often overbuild the infrastructure that later becomes indispensable. Even if parts of the American AI boom prove inflated, the investment wave is still creating vast reserves of compute, data centers, and engineering capacity that may support future phases of economic development. Seen this way, bubbles become accelerated forms of Schumpeterian creative destruction under financial capitalism.
China, by contrast, appears less invested in the mythology of superintelligence and more focused on AI as industrial infrastructure. In this model, intelligence is distributed across factories, vehicles, hospitals, logistics networks, and public administration. The emphasis is less anthropomorphic and less centered on the humanoid assistant than on systems integration and operational coordination. If the Western imaginary is the chatbot companion and superassistent, the Chinese one might be the infrastructural network and fully automated communism luxury.
These distinctions are, of course, stylized. In practice, elements of the Chinese approach increasingly appear within American firms, just as Chinese companies adopt aspects of Silicon Valley’s model. But the contrast remains analytically useful because it reveals different trajectories within the broader and often amorphous phenomenon labelled as the international AI rat race.
The labor debate around AI is mainly framed in familiar terms around the individual human-machine relationship: co-pilots, human-machine collaboration or augmentation, or, ofcourse, outright replacement of workers. Yet this vocabulary may understate what is distinctive about the current moment. The coming labor shock is not merely another wave of augmentation or substitution. Many argue that this may signal a deeper reorganization and redefinition of work itself, comparable to earlier transitions toward industrial work and knowledge work. The more fundamental question, then, is: what constitutes ‘AI-driven’ work, and how should we organize and distribute the different parts of a new ‘labor cycle’?
And yes, one likely consequence is a slow but persistent erosion of the current labor force yet in unexpected and unforeseen ways. Chinese economist Andy Xie offers a useful provocation here. Earlier waves of globalization displaced labor geographically: firms outsourced production and services across borders in search of lower costs. AI changes the mechanism. Instead of moving work abroad, companies can increasingly internalize labor arbitrage through software embedded across the entire value chain. Functions once carried out by departments, contractors, or even entire professional classes can now be partially absorbed by machine systems operating within the firm itself.
That distinction matters. Offshoring still depended on human labor elsewhere, even as it devastated particular regions such as the industrial belt of the United States. AI-driven replacement, by contrast, threatens to dissolve the organizational location of certain forms of work altogether. Displacement no longer occurs at the edge of the firm, but inside it. Some regions and industries will still be hit harder than others, but the broader effect may be a gradual weakening of labor as such — visible, for instance, in the declining value of entry-level roles and the uncertain worth of junior qualifications and degrees in, for example, law.
At the same time, new business models point toward something even stranger: a reversal of roles. Already long experienced by the proletariat and the bureaucrat as being little more than a cog in the machine, humans now encounter platforms such as RentAHuman.ai, where AI systems can effectively ‘rent’ physical bodies to perform tasks they cannot yet execute themselves, from package delivery to other forms of embodied labor. Such models point explicitly toward a world in which the traditional, and always somewhat illusory, hierarchy between user, worker, and machine is inverted without taboo anymore. Humans become increasingly docile to the demands of machinic systems: a numbed and uncanny version of ‘all watched over by machines of loving grace.’
This feeds immediately back to our first topic: rather than humans directing software, agentic machine systems may increasingly orchestrate workflows themselves, calling in human labor only where it remains necessary. In this configuration, human beings function as exception handlers within a broader automated process. This challenges the still deeply rooted assumption that technology is fundamentally an extension of human agency: a tool in our hands, something designed to augment human capacities. In agentic environments, humans may instead become support functions within machine-led systems, intermittently deployed, tightly monitored, and evaluated through market mechanisms not primarily for creativity or judgment, but for their ability to resolve the residual trivial problems automation has not yet overcome.
