There is a new goal for tech start-ups: wipe out $1trn in stock market value in a single day. That’s what the Chinese AI company DeepSeek achieved on 27 January. DeepSeek employs about 200 people and its hiring practices follow the much-touted Silicon Valley model (a model that remains far from reality in Silicon Valley itself): hire young talent, disregard qualifications, focus on sheer intellect and eschew fixed hierarchies. Notably, all of its research team was educated domestically.
What happened next is like a Hollywood plot. A team of young renegades from a hedge fund stuck it to tech giants, large companies flush with money. By 28 January, the previously unassailable Nvidia – the American chip manufacturer that powered Silicon Valley’s AI boom and had just passed Apple as the world’s most valuable company – saw its stock plunge 17 per cent. Meanwhile, Sam Altman, the hubristic CEO of American AI company OpenAI, tweeted something so self-effacing I had to wonder if his account had been hacked: “We will obviously deliver much better models and also it’s legit invigorating to have a new competitor,” he posted about DeepSeek.
The implications seemed momentous, though it was far from clear what they were. DeepSeek claimed it had trained its large language model – an AI model, like OpenAI’s ChatGPT, which can generate and understand language – for just $6m, pocket change when compared with the expansive budgets of its Western rivals. The company also claimed to have used only 2,048 older and slower Nvidia chips, a necessity imposed by US sanctions, as opposed to the 50,000 cutting-edge chips reportedly used by Microsoft and OpenAI to train GPT-5. Given these numbers – which many find hard to believe but have not been disproved – it was plausible to conclude that Nvidia’s valuation, which peaked at over $3.6trn, was unsustainable. Perhaps advanced AI models could be trained with only a fraction of the hardware once deemed necessary. The notion of spending $100bn on data centres to train leading models that are likely to be commoditised before one has even found much use for them suddenly seemed unwise.
The threat to Nvidia and other US tech companies is even greater. Computing power might not lose value in a world where it can be used more efficiently with superior algorithms: imagine what DeepSeek’s algorithmic brilliance could achieve with $100bn in chips. As the French economist Olivier Blanchard put it, DeepSeek may turn out to be the greatest positive productivity shock in the history of the world. But if the underlying models are Chinese rather than American, the surrounding ecosystem of hardware and energy sources is likely to be Chinese as well. Nvidia should be less worried with how cheap DeepSeek was to develop and more concerned with the news that for using the model rather than training it – what is called inference – the start-up turned to Chinese firm Huawei’s Ascend chips.
Though DeepSeek’s impact has been widely felt, it might still be underestimated. If a small Chinese company could marshal this kind of engineering talent and achieve a breakthrough apparently beyond its American rivals, what other reservoirs of talent might be hidden in the bas-fonds of Hangzhou and Shenzhen? New doubts have surfaced that Washington will be able to deny China access to cutting-edge chips. Are Chinese engineers close to surprising breakthroughs in creating chips as well? No one knows. Yet.
DeepSeek also made its model freely available for everyone to inspect, use and even instal locally in their own hardware. Both the company and China itself stand to gain from that choice, which contrasts vividly with the closed-source approach – where the software is not publicly available – adopted by most of its American competitors. DeepSeek will likely be able to continue hiring the best talent, which tends to be attracted by the impact they can have. China now has AI models so cheap and powerful that young entrepreneurs all over the world may choose them as the foundation for new applications.
I happened to be travelling in India the week after DeepSeek’s debut and heard of young people there downloading the model to their hardware and discussing ideas for revolutionary applications in healthcare and financial services. Every company in the world can use DeepSeek without sending data to any specific country.
DeepSeek should make us pause for yet another reason. Rather than relying on supervised learning with human feedback, the company’s engineers relied on what is called reinforcement learning. This is a technique in which the model is left to learn on its own, receiving positive reinforcement when it reaches the right answer. Pure reinforcement learning is the Holy Grail of machine learning because it creates opportunities for genuine insight rather than rote memorisation, but until now no one had managed to make it work. And if the DeepSeek model was also able to learn from other models, including its main rivals (as OpenAI has alleged), through a process called distillation, then we seem to have reached the point where AI has begun to improve itself – rather than having to wait for human engineers – and to do so at computer rather than human speed. Buckle up.
What are these reasoning models that are being constantly born? Or maybe, who are they? Proper introductions are in order.
Until now the guiding image of an AI was that of a supremely intelligent oracle to whom one might in time direct any question and get the answers human beings have always dreamed of obtaining. But what if this image is wrong? What if the more accurate image is not of a mind but a world? Minds need a world in which to exist and operate. To get an AI agent to accomplish a task, you need to give it examples of what success looks like. The reward function in reinforcement learning does this by telling the model it is on the right path, creating a world picture. Even autonomous vehicles must operate in a virtual environment, such as a digital map of a city used by the driving algorithm. What is intelligence if not a comprehensive model of reality?
Large language models (LLMs) are exercises in world-building. When we give them a prompt, we are asking the following: given the regularities in the world of human language, how would a text or a sentence starting with these words most likely continue? How should this text be generated in such a way that it respects the same patterns contained in existing texts? The patterns are those of the language corpus, but because the text is being recreated from scratch, there is always the possibility of creating new realities, provided they respect most general patterns. Many people using an LLM for the first time were attracted by these virtual possibilities. For example, they might want it to create a Shakespearean sonnet about Taylor Swift or a Taylor Swift song about Shakespeare. Or footage of a Siberian tiger eating Chinese hotpot with chopsticks in the case of a video generator. The AI model was creating virtual worlds and inviting us in, while also getting inside our minds, redefining our sense of the real.
The introduction of the first highly effective video-generation models in February 2024 underlined the deep connections between AI and world-building. While language models seemed to track human intelligence, which is indeed expressed in language, the generation of virtual worlds seemed more like a divine power. OpenAI’s text-to-video model, Sora, is able to create complex scenes with multiple characters, specific types of motion and accurate details of the subject and background.
Li Zhifei, one of the entrepreneurs driving the Chinese AI boom, said soon after Sora appeared that large language models emulate the world of virtual thought, while video-generation models emulate the physical world. “Once the physical and virtual worlds are constructed,” he asked, “what exactly is reality?”
The march towards these comprehensive models of reality vividly illustrates the main thesis in my new book World Builders: that the world will be divided into two levels. On one is the programmer creating the engine; on the other are the users taking the constructed world as their singular and inescapable reality. Masters and slaves.
Researchers have studied how models can be built to exhibit patterns acquired through training, fine tuning, reinforcement learning or a specialised data set. Selecting a certain data set gives the game away: one could easily select a Chinese or American corpus. Users in China can log on to the website of the country’s cybersecurity association and download the Chinese basic corpus for LLMs. Last year China even produced one trained from a fixed pool of texts on the ideological doctrine of Xi Jinping. As such, every model using that corpus expresses a specific vision. There will never be a neutral model.
Concerns that bias in data could result in bias in model outputs have long plagued the industry. Asked to create an image of a nurse, an image generator might produce an image of a female one, simply because it was trained on a database of images in which nurses tend to be female. When Google offered image generation through its Gemini platform in February 2024 it had to find a way to address these diversity and bias issues, but the purported cure was worse than the disease. The solution was to modify user prompts before feeding them into the image generation model. The prompt injection process might add words such as “diverse” or “inclusive”, or even specify ethnicities or genders not included in the original input. This type of prompt engineering helps guide the model’s behaviour by adding context or constraints, in the process subtly modifying our sense of reality. DeepSeek refuses to answer questions on Taiwan or Tiananmen Square, and, when pushed, gives answers that follow China party lines. US models are no different: some refuse to answer any questions about Trump; LLMs I tested answered questions about the rights of Israelis and Palestinians very differently.
As Meta admitted in a white paper from September 2023, prompt engineering “works behind the scenes”, but it occasionally has a way of becoming apparent. When users first started experimenting with Gemini, the results were both comical and catastrophic for Google: the model responded to a prompt for a “portrait of a Founding Father of America” with images of a native American man in a traditional headdress, and when it was asked to create the image of a pope, it inexplicably returned a picture of a female pope.
This episode offered the first public demonstration that AI is never impartial, though its presentation as a technical process may make it impossible to identify the human will “behind the scenes”. This is perhaps the highest form of power: human will disguised as reality.
Whichever model becomes dominant or foundational will have the singular power to shape how its users view the world and their reality. “Once men turned their thinking over to machines in the hope that this would set them free,” the Reverend Mother explains in the sci-fi classic Dune. “But that only permitted other men with machines to enslave them.”
While language models were used to power conversation bots such as ChatGPT, their impact seemed benign and limited. But within months of their public launches their use was widespread. Text generated by AI models is increasingly included in web pages, so the internet is reproducing their structure and biases. People ask the model for plans and advice for their professional and personal lives and then implement the results. Students learn from interacting with it. The model may perform tasks on your behalf and take over your personal device or computer. LLMs can be plugged in to robots, giving them artificial brains. The model is eating the world.
A model becomes more powerful as it becomes integral to society’s general infrastructure. AI is the central brain or operating system of a virtual world, orchestrating inputs and outputs across every format, writing code and processing data and memory. The point is not to create a super intelligence you keep in a giant data server. The point is to release it and watch it become the world brain. By offering its model as open source and reducing operating costs, DeepSeek seems to have an intuitive understanding of this point, which offers some democratic possibilities.
You might miss those implications, but they are present in the very idea of a race between superpowers. Those who build the world within which others act must find a place for everyone – or risk a more general framework to be built on top of their own failed construction. The growth of cities as repositories of succeeding civilisations is an appropriate metaphor to help us visualise the process. As for who gets to be a world builder, the point to stress is that each foundational model is an attempt to create a world, and the race is open, full of unexplored possibilities. Alas, this is where the democratic implications end.
Myriad intelligent agents will emerge this year – artificial intelligences that do actual work rather than merely thinking and talking, moving from “chatting” to “doing”. These agents run on a foundational model but are trained to perform tasks such as organising your agenda and booking flights and hotels, negotiating deals for you, attending meetings in your place or replacing your physician during routine medical appointments. Intelligent robots may also make an appearance and perhaps even companies or firms entirely staffed by AI agents. The foundational model is the virtual world within which these agents operate. It will remain the biggest prize.
In such a world there is no recourse to an external authority. The engineering power has set the rules in advance and alone enjoys singular access to those rules. Hackers call this God Mode: access to everything and root privileges to do everything. Those who designed the foundational model will have the ability to introduce specific “policies”. Prompt engineering – that deliberate manipulation of the prompts that users enter into the model – is just one example. Models have policies that developers train them towards, and these policies may be hidden. Users of a large language model may not know about hidden biases if they lack access to its inner workings or lack understanding of how it was trained or the data set it was developed on, which may have been highly selective texts.
Throughout history new technologies have raised the destructive potential of direct conflict. In the AI age, a new avenue has opened: states can now fight one another not by winning in direct battle but by building the world that everyone else must inhabit. Imagine a time when there truly is a global brain directing every social and economic activity. It might be possible or even easy to insert a specific policy or aim to which these activities must contribute, and to hide this policy so deep inside the model that no one – apart from those who built it – will ever know it is there.
The models available today may seem no more than a chatbot, but this is only the beginning. DeepSeek taught us two important lessons: first, that the process will accelerate, or has already started to. Second, that the real battle is about building a model of reality that can be adopted globally. It goes much deeper than physical infrastructure. A foundational model is the infrastructure of thought.
Call it a form of invisible government, a return to the myth of a hidden king ruling the world from the underground city of Agartha. Perhaps your opponents will assume the way things work are natural or given, that reality exists outside human control, but in fact you have moved one level up in the great game. Your opponent is playing a video game. You are coding it.
“World Builders: Technology and the New Geopolitics” by Bruno Maçães is published by Cambridge University Press on 13 February
[See also: The Do No Harm dilemma: what happens when a drug that can save lives could also ruin them?]
Content from our partners
Related
This article appears in the 05 Feb 2025 issue of the New Statesman, The New Gods of AI