Masters of the Metaverse

The philosopher David Chalmers believes we may be on the cusp of a great migration, one likely to overshadow every wave of human migration in history: the move to a virtual world or worlds, as the real one continues to degrade. In centuries to come, instead of asking: ‘should we move to a new country to start a new life?’ we may ask: ‘should we shift our lives to a virtual world?’. As with emigration, often the reasonable answer may be ‘yes’.¹ With sufficient computing power, virtual reality will no longer consist of pre-packaged individual experiences. It will become a concurrent and real-time platform where billions of people can meet to conduct business, shop and have fun. In one of his online interviews, the journalist Dean Takahashi called the metaverse ‘the most difficult and important thing humanity will ever build’.

In October 2021, when we were halfway through the pandemic, Facebook announced it would be changing its name to Meta and altering its mission to the elusive pursuit of the metaverse, a new virtual reality platform that will play host to ever more of our future lives. In February 2024 Apple launched the Vision Pro, a device allowing us to run apps and watch videos in the virtual space around us while granting us access to the real world not directly but through a perfect simulation of the real world, built from cameras on the front of the device, as if the user is seeing right through it. The interface is strikingly new: you look at things you want to control, as if your eyes had become the mouse, and you tap your fingers to click. Perhaps the belief that hand gestures are used by witches and wizards was a prophecy of the future or a holdover from some lost technological civilisation.

Chalmers wants to reassure us that we can and should embrace the migration to new virtual worlds. There is nothing to fear because the worlds we are about to enter are no less true than the world we inhabit today, and our lives will be no less valuable on account of being virtual. One of his main arguments is the famous simulation hypothesis. Now that humanity is approaching the point where it can create powerful computer-based simulations, it is only natural to ask whether a more advanced civilisation, perhaps even our own future generations, might not have created a simulated world, and whether we exist within that simulation – the Matrix from the eponymous film – taking it for reality. In that case we have lived in a virtual world all along and there is nothing to fear in moving to new ones of our own making.

As virtual technologies improve and grow, we are already seeing how they are starting to erode the structures aimed to contain them. One could say the whole point of these technologies is to reach deep into the ways we experience reality and to radically transform them. Cryptocurrencies have not just dematerialised money; they promise to create social structures built around new forms of decentralisation, where control over the network is equivalent to the network itself. In the blockchain – a programmable computer with no physical existence – contracts stop being documents or pieces of paper and become computer programs that enforce or execute themselves, without the need for human involvement. Within these networks, artificial intelligences – arbitration bots – viciously compete to capture ‘maximal extractable value’. The ultimate source of truth is the blockchain, not the physical world.

Plastic surgeons are now asked to match the settings on Instagram picture filters, and people redecorate their houses in order to create suitable backgrounds for guests to pose against. Social media platforms have yanked us out of the normal rhythms of life and created an artificial world where conversation and politics happen at superhuman speeds and with such gripping intensity that many find it difficult to log off. In early 2024 a large multinational company lost HK$200 million in a scam after employees at its Hong Kong branch were fooled by deepfake technology, with one incident involving a digitally recreated version of its chief financial officer ordering money transfers in a video conference call. The scammers applied deepfake technology to turn publicly available video and other footage into convincing digital copies. The company employees in the call looked and sounded like the real people whom the targeted employee recognised. They were all fake.²

Habitat, created by Lucasfilm Games in 1985, was one of the first large-scale commercial virtual environments. The system was developed to support a population of thousands of users in a single shared cyberspace. The players were represented by animated figures the game creators called ‘avatars’. It was the first modern use of the term to refer to the player characters in a virtual environment. The player issuing commands to the avatar exists outside the virtual world of the character. The distinction between the two worlds immediately raises the question of transcendence, making it natural to look for references in religious experience. Since ‘avatar’ is the Hindu term for the incarnation of a god in transient reality, the choice was a felicitous but perhaps obvious one. Chip Morningstar, one of the creators of Habitat, explained that his background as a bookish kid helped him remember the word from writings on Hinduism. Humans become like deities, or at least external souls, with respect to a virtual world that exists only inside a computer simulation.

In his 1992 novel Snow Crash, Neal Stephenson defined avatars as the ‘audiovisual bodies’ existing in the virtual reality internet he seminally called the ‘metaverse’. A lot of our life has been taken over by different avatars already. My Twitter persona is certainly very different from my real one. As we approach the final point of a fully virtual existence, the question of how to live in a world of avatars becomes an urgent one. Will the metaverse create a wholly new sense of self?

One immediate effect is the bifurcation of the self between a higher command centre and a multitude of created characters inhabiting one or several virtual worlds. Individual life experiences the miracle of multiplication. What if an avatar could attend online business meetings and calls in your place so you can engage in more productive activities instead? It would be easy to train a model on recordings from hundreds of your previous meetings and communications so it knows exactly how to react and what to say. ‘It actually has an even better memory than yourself.’³ Users of the dating app Volar can already train a chatbot to go on virtual first dates for them with the bots of potential matches, ‘opening with an icebreaker and chatting about interests and other topics picked up from the person it is representing’.⁴

No longer sent into the world understood as a garden of forking paths, the obvious temptation will be to live many parallel lives at once. Biography becomes biographies. There will be nothing to lose because no one will feel too invested in one particular avatar. In that way our sense of self might dissolve to a considerable extent. A man or a woman might feel at ease adopting a different gender in the metaverse, at least as one of their many avatars. Seeing the world from other points of view will give us a richer experience and a more tolerant outlook. Or so one hopes.

If these speculations remind you of the different personalities adopted by the Portuguese poet Fernando Pessoa, that is no coincidence. In a letter to Adolfo Casais Monteiro, Pessoa described his heteronyms as products of his taste for simulation. ‘Be that as it may, the mental origin of my heteronyms lies in my organic and constant tendency towards depersonalisation and simulation.’ Novelists create a gallery of characters, but they leave them in their novels, reluctant to give them reality in their lives. Fiction stops short of virtual reality. But why should it stop short? Why should a novelist care more about his or her novels than about his or her life? The obvious culmination of these developments is to use your corpus of social media activity to train an artificial intelligence capable of impersonating you after death and live in virtual eternity. ‘When your heart stops beating, you will keep tweeting,’ one company once promised.⁵

The great migration to digital during the pandemic showed the enormous advantages of being able to work and live within an artificial, secondary universe. In this universe, the laws of space and time no longer apply, or at least they can be bent, enhancing human powers in ways that remain to be explored: an end to long commutes and the achievement of measurable increases in productivity; the ability to participate in meetings and conferences on different continents on the same day; and children still able to attend school even amid the worst public health emergency in a century. Unfortunately, the limits of digital experience were no less apparent. A lot gets lost when human interaction takes place on a screen. The results of remote schooling have so far proved mixed at best. A digital work environment soon revealed itself as considerably more exhausting than the real counterpart. Human beings are built for the kind of immersive interaction that takes place in the physical world, where all of our senses are involved. Some of our mental abilities, including memory, suffer markedly when we are reduced to disembodied egos on Zoom. And as for entertainment, digital experiences are still so far from the actual fun of going to a restaurant or a music concert that nothing one tried on the internet during the lockdowns measured up.

The immediate appeal of the metaverse is that it promises to marshal the virtues of digital life while addressing many of its shortcomings. Instead of business meetings on Zoom, imagine entering a digital room and talking to your colleagues around a virtual table, or even walking together in an electronically conjured garden. I think of the metaverse as a virtual world with some of the characteristics of a city. There will be virtual malls, where users can move from store to store and buy the products that will later be delivered to their physical homes, a big improvement in digital shopping over the flat web page. There will be virtual beaches where we can meet our friends to chat and play, much like what Fortnite already offers today, only much better. There will be concerts and art galleries. Is there a reason to travel physically to Venice to visit the Biennale instead of jumping into the metaverse and enjoying all the art and video installations with the latest fully immersive technology? Travelling to exotic locations could happen while we sit in our own living room. One area where I expect the metaverse to have a profound impact is television news: after reporters travel to a war zone or a political rally, viewers will be able to experience those events as if they too were present on location by using immersive technology. It is easy to imagine the growth of a new virtual economy, where creators will be less dependent on mediators of all kinds. Barriers to entry will likely be lower – and audiences potentially much larger – than in the real world.

The metaverse has to be highly concurrent: this artificial world has to be continuously updated from the inputs of its millions – or potentially billions – of users. The only way a virtual experience can rival the real thing is if it acquires the same flux and complexity. It is definitely not enough to encounter a pre-packaged version of reality. One wants to travel to a virtual Biennale in order to see what others are preparing for us in real time; ideally, too, one wants to arrive with thousands of other people, with whom it would be possible to interact and with whom one could share the event. It should be possible to acquire some of the artworks, in both physical and virtual formats. Virtual objects, once acquired, must be preserved within the metaverse. One should be able to carry them to other digital spaces, without sacrificing their virtual authenticity, in a world where ‘virtual authenticity’ is not a contradiction in terms.

Let us return to the virtual malls mentioned earlier. Users visit a virtual car dealership where they can buy a model to be delivered to their physical homes. They will want to test the car virtually, and they may also be interested in buying a virtual twin to drive in the metaverse. The virtual and physical worlds might become increasingly integrated.

What truly distinguishes the metaverse is its autonomy from the physical world. The metaverse exists on its own. It has a life of its own. It creates a genuinely alternative world. One accesses the internet. One enters the metaverse. The difference matters. To see why, consider how the relation between the user and the digital environment gets turned on its head. With the internet, the user remains sovereign, dictating when and how digital interactions take place. In the metaverse, the user finds him or herself entirely surrounded by the platform, and the quality of the experiences will frequently depend on whether he or she accepts that fact. If the ambition of the metaverse is to constitute an artificial world, capable in time of rivalling the real one, experiences there must acquire meaning with reference to other experiences in the metaverse, not just those taking place in the real world. And this means that the metaverse should never really disappear, even when a user momentarily withdraws into real life. The metaverse must persist or perdure.

Aspects of the metaverse have been around for a while already. How are we to interpret the advent of social media if not as the first stage of the metaverse? Consider the social media company formerly known as Twitter and now known as X. It has some of the characteristics of an artificial world, starting with high density and including persistence and continuity, helping to explain its extraordinary success. Technically, X is rather primitive, a predominantly verbal medium of low bandwidth. From a conceptual perspective, however, it has been a revolution: a new world that users enter rather than contemplate and that continues to exist and develop in their absence. Log out, go to sleep, and when you wake up, all kinds of developments, responses and interactions will have occurred. X feels like a game. For some, this is simply perverse, but once you realise you are playing a game rather than having a conversation, your expectations adjust. X is a series of game moves during which you take certain risks with your avatar in order to accumulate high scores in attention and approval, measured by followers, likes and retweets. It was never my impression that X is a platform, if by platform one means a medium of communication. X is a virtual world where a parallel life is always developing in many different directions, even during the time when one is logged out.

I am reminded of the story of a New York executive who posted an ill-considered tweet right before boarding a flight to South Africa a few years ago. While she was flying, presumably without internet access, the tweet was retweeted thousands and thousands of times every hour; she was fired from her job, lost all her friends and saw her life collapse like a house of cards. She had disconnected from Twitter, but the beast kept growing and moving, even in her absence. What makes the platform addictive is its actual resemblance to the physical world in the power of its autonomous operations while being relatively free of many of its constraints. For an entrepreneur like Elon Musk, nothing else can remotely compare to the concept of building a virtual world to where everyone else might one day migrate. Transportation or energy networks might be foundational, but they still exist in the real world. The metaverse is a new virtual world to replace the real one. And everyone is supposed to live in that world, a world you built and, therefore, control.

Many current social developments are early manifestations of what the metaverse will likely bring. As social media lays down the initial infrastructure for alternative realities, we witness how conspiracies of all kinds and other social games start to erupt. Many can be better understood as analogues of video games, elaborate programs where participants take on different roles or avatars and must follow complex rules determining how the game should be played and how to perform the tasks for which social points may be awarded. As the metaverse expands, phenomena of wholesale dissociation from reality are bound to multiply. We may soon face a choice between building a rich metaverse and improving the real world, just as reality faces new crises ranging from pandemics to climate change.

One early report from a Chinese think tank predicted three immediate impacts of the metaverse. First, it will be a driver for technological innovation and, in some cases, in areas adjacent to military interests: simulation graphics, artificial intelligence, wearables, robot technology and brain–computer interfaces. The second will be the moving of the digital ecosystem and the digital economy to new technological platforms. Third, the metaverse will start to integrate the needs of the virtual and physical worlds, as we saw earlier, realising an old ambition of the internet age.⁶

The report anticipated that the metaverse could have deep consequences for the global distribution of power. It ‘will trigger a new round of reshuffling’ in the global technological order. Some companies and countries will lose out; others might have an opportunity to rise. That happened already with the internet economy, when Europe could not help falling behind the United States, and China was able to resist the imperium of the large American platforms but only from a defensive posture.

It is easy to see why many in China might regard the concept of the metaverse as excessively influenced by Western ideas. The metaverse is, before everything else, a method of escapism. Each individual gains the freedom to pursue his or her own most personal fantasies in the metaverse. It is as if the common world fragmented into millions or billions of private universes. Science fiction writer Liu Cixin argued that the metaverse is like a drug, one so powerful as to break our connection with the world around us. As he put it, humanity is now at a fork in the road: in one direction lies the exploration of outer space; the other leads inward towards virtual reality, the ‘dead end of entertainment’.⁷

How could the Chinese Communist Party not regard the project as a threat? The historical mission of a revolutionary party is to call the masses to a common project and to transform the physical world as needed to satisfy this objective, not to encourage an escape to an artificial universe. Renowned blockchain expert Yu Jianing, author of a book in China called Metaverse, has insisted that the metaverse must never become a virtual economy but only a tool ‘empowering the real economy’.⁸ And yet, the metaverse offers a range of intriguing possibilities to the Chinese regime. By forging different levels of virtual reality, it could establish different levels of access for different groups while ensuring a much tighter grip on how access is distributed. The possibility reminds me of phenomena I have written about in a different context. Even in its formative stage the Belt and Road is an exercise in the opacity of power. There is an exoteric doctrine of the initiative and then an esoteric practice where deals are agreed upon, often with no written evidence, and where the hierarchy resembles that of levels of security clearance: confidential, secret and top secret. Some of the participants know only the broadest strokes of the plan, sufficient to defend it and to communicate with lower levels, others know nothing, and only a few can see months or years in advance.⁹

I return to this question in the next section. The metaverse represents the most recent battle between human freedom and the constraints of reality. As recent developments in artificial intelligence amply illustrate, it could also become a battle to define reality itself.

Second Genesis

Large language models 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 the corpus of 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 are part of the same family and explained from the same general patterns or regularities. The deeper or more abstract the patterns identified by the model, the greater the room for variation.

The interface is not a search interface but a generative one: what you ask the model is to recreate or rephrase the internet. After all, the model does not contain any content but rather a set of rules for generating content. While the internet is full of content, a large language model is empty of content. It is, in this sense, a virtual engine. Using the organising principles or statistical regularities extracted from the current iteration of the internet, it is possible to recreate it, or then to create new iterations, provided they follow the same fundamental rules. The method employed is distribution learning: how to extract from training samples the rule or algorithm governing the distribution from which they have been drawn. At which point an obvious question arises: why, if we already have one internet, do we need to recreate it?¹⁰

The answer is that this recreated internet runs virtually on machines, and a virtual internet extends to infinite variations of the internet we now have. Many people using a large language model for the first time are attracted by these virtual possibilities. For example, they want it to create a Shakespearean sonnet about Taylor Swift or a Taylor Swiftian song about Shakespeare. Or footage of a Siberian tiger eating Chinese hotpot with chopsticks in the case of a video generator. This does not mean that anything goes: the world must obey the same structural rules that are valid for its current state, since the design rules are actually extracted from this current state. What is recreated is not the current internet but a virtual one, one possible game state in a general game world. Rendering a game state is equivalent to finding rules predicting what will happen in this virtual world in response to every possible prompt that could be performed. These rules are exactly what is contained in the model and training. Given a sequence of tokens, a single token is drawn from the distribution of possible next tokens. Each such token represents a possible continuation of the sequence. From the most recently generated token, a tree of possibilities branches out. ‘This tree can be thought of as a multiverse, where each branch represents a distinct narrative path or a distinct world.’¹¹

This also means that different realities can be produced rather than discovered or reasoned from the available evidence. The difference between searching the web and prompting a large language model is the difference between the real and the virtual. Instead of having to fill in gaps or resolve contradictions between different search results, a user in the virtual internet can simply ask for the result he or she wants. This is not an already existing world one discovers or investigates. It is world building in response to our prompts, nothing short of a personalised internet.

It is much cheaper and easier to create a new virtual world than to use the real one. For example, it is easier to produce a new video from scratch using a large image model than to look for a location in the physical world in order to film it. It is easier to produce billions of weather patterns in a digital twin of the Earth than to wait for those patterns to be naturally produced in the physical planet. Fabs building semiconductors, research labs and chip equipment manufacturers connect digital twins of chips, equipment, processes and labs in a complete virtual representation. This way one can test new ideas for a few dollars of computational time in virtual labs that might cost millions in physical labs. Already seven decades ago, at the birth of the computer age, the mathematician and computer engineer Richard Hamming explained to the president and leadership of Bell Labs, which he had joined after leaving the Manhattan Project: ‘At present we are doing one out of ten experiments on the computers and nine in the labs, but before I leave it will be nine out of ten on the machines.’ It was so much cheaper to do simulations than real experiments, so much more flexible for testing, and one could even do things that cannot be done in the real world. If the notion was met with resistance and is still met with resistance, then that may be, as Hamming suggests, because we have been taught, since Bacon and Galileo, to learn from nature rather than books. Today, we are again, like the ancients, ‘looking more and more in books and less and less in nature!’.¹² The only difference is these books are written in code.

Think of the difference between the internet and a virtual internet as analogous to the difference between a machine and a computer. While most machines are developed for particular purposes, modern computers are designed to be able to simulate the operations of many different machines, such as movie projectors, music players, cameras or typewriters.¹³ Or think of the internet as you know it as a very large museum and the virtual internet as a screen or a stage where content flickers for a moment only to be replaced by new content. A large language model can simulate many different internets, even previously unthinkable types. The technological leap is bought at a cost, however. With the internet as it exists today, the user is ultimately in control, being tasked with creating and interpreting content. The virtual internet takes over these tasks. The internet changes from an object to an environment, a medium within which we move, reacting to our moves.

The model provides both the rules and the limits of the virtual world. It takes the place of the laws of nature in the physical world. This is where artificial intelligence finds its mission: in order to build a virtual world endowed with the same qualities of the real one, one needs some order or organisation. Were this order to be given by the laws of nature, it would be no more than our physical reality. Were it given by human thought and will, it could not be an autonomous world. By ‘an autonomous world’ I mean a world that does not disappear the moment we stop thinking about it.

One project at Stanford in 2023 used a large language model to create an artificial society of twenty-five members in a game environment. With traditional games or simulations it would be necessary for a coder to script different behaviours manually. With generative agents it is enough, for example, to tell one of the agents that she wants to throw a party in order to produce believable simulacra of both individual and emergent group behaviour. The key observation is that large language models encode a wide range of human behaviour from their training data. By interacting with each other in their small virtual town, the generative agents in the Stanford model exchanged information, formed new relationships and coordinated joint activities. These social behaviours are emergent rather than preprogrammed. Theorists often speculate on what the best image or metaphor for artificial intelligence might be: an agent, an oracle, perhaps a genie or tool following human commands. But they all miss the point: artificial intelligence is a virtual world – or, better put, it 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. As Chalmers notes, the agency model is misleading, and a large language model is ‘more like a chameleon that can take the shape of many different agents. Or perhaps it is an engine that can be used under the hood to drive many agents. But it is then perhaps these systems that we should assess for agency, consciousness, and so on.’¹⁴

The introduction of the first highly effective video generation models in February 2024 underlined the deep connections between artificial intelligence and the metaverse. While language models seemed to track human intelligence, whose operations are indeed formalised within verbal and mathematical languages, it is much less natural to think of the generation of virtual worlds as a faculty of human intelligence. In reality, of course, video generation and language models are quite similar. Whereas large language models have text tokens, a video generator such as Sora has visual patches. Given a compressed input video, it extracts a sequence of space–time patches that act as tokens to generate realistic videos. Sora is able to generate complex scenes with multiple characters, specific types of motion and accurate details of the subject and background. The model understands not only what the user has asked for in the prompt but also how those things exist in the physical world. Scaling video generation is a promising path towards building general-purpose generators of virtual worlds. Li Zhifei commented soon after Sora appeared that large language models are emulators for the world of virtual thought, while video generation models are emulators for the physical world. The Chinese tech founder went on to ask: ‘Once the physical and virtual worlds are constructed, what exactly is reality?’¹⁵

A genuine virtual world, or a metaverse, cannot exist before the emergence of a powerful artificial intelligence. Some intelligence distinct from both natural laws and human reasoning must govern how these worlds work. But the march towards a virtual world engine illustrates the main thesis in World Builders in a second way: it divides the world into two ontologically different levels, that of the programmer creating the engine and that of the users taking the virtual world for their singular and inescapable reality.

Researchers have studied how models can be trained to exhibit patterns acquired through training, fine-tuning or a specialised dataset. Selecting the right dataset or corpus of text gives the game away. One could easily select a socialist or a liberal corpus. It is significant that users in China can log in to the website of the China Cyberspace Security Association and click the Chinese Basic Corpus link to download the relevant corpus for large language models.¹⁶ And in May 2024 the Cyberspace Administration of China rolled out a large language model trained from a fixed pool of texts and sources representing the full range of Xi Jinping Thought on Socialism with Chinese Characteristics for a New Era. The model has full command of the philosophical doctrine. It can answer questions, outline reports, summarise information and translate between Chinese and English. For example, suggested questions for the application included the differences between traditional productive forces and new productive forces, a recent but central conceptual distinction for Xi Jinping.¹⁷

As for fine-tuning, the technique uses a training set of thousands to hundreds of thousands of examples of the desired behaviour, perhaps scripted like a dialogue, in order to bias the model expectations towards specific patterns. To effectively do this with large language models is so expensive that it may well be that only states and two or three tech giants will succeed in achieving it. Because the patterns are not contained in any specific lines of code, they are very difficult or impossible to detect and remove. Nothing except the overall architecture is engineered: the model is actually learned or obtained from training. To the extent that large language models create virtual worlds, the limits of these worlds will be seen as coterminous with reality as experienced by its human users, but there is nothing neutral or natural about this reality. Virtual worlds have policies that the developer trains the model towards, and these policies may be hidden or backdoored. Users of a large language model may not know about hidden patterns or biases in the model if they lack access to the model parameters, the training process or the weights obtained from training or the dataset, which may encompass highly selective internet text. This may also create an opportunity to insert a backdoor: undesirable or unexpected behaviours that are triggered only by specific inputs.¹⁸

Concerns that bias in data could result in bias in model outputs have long been expressed. Asked to create an image of a nurse, for example, an image generator might typically produce an image of a female nurse, simply because it was trained on a database of images in which nurses tend to be female. When Google started to offer image generation through its Gemini platform in February 2024 it had to find a way to address these diversity and bias issues, but the propounded cure was worse than the disease. In the case of Gemini, the solution was to secretly 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. Prompt injection or prompt engineering is a direct modification of the text input before it is sent to the model, which helps to guide the model behaviour by adding more information, context or constraints, such as updating and randomly rotating through prompts that use different qualifiers – for example, ‘nurse, male’ and ‘nurse, female’. As Meta admitted in a white paper from September 2023, this ‘works behind the scenes’.¹⁹ When users started experimenting with Gemini, the results were both comical and catastrophic: 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 returned the image of the Pope as a woman. The episode offered the first public demonstration that artificial intelligence is never neutral, although its apparent objectivity, its appearance as a virtual world, may well make it impossible to identify the human will ‘behind the scenes’. The culmination of ideological power: a will disguised as thing. ‘Once men turned their thinking over to machines in the hope that this would set them free,’ the Reverend Mother explains in the science fiction classic Dune. ‘But that only permitted other men with machines to enslave them.’

While language models were used to power conversation bots their impact seemed limited, but within months of their public launches they were already replicating and gradually assuming the structure of a comprehensive environment for human thinking and activity. Text generated by the model is included in web pages. People ask the model for plans and advice and then implement them in their actions. Students learn from interacting with the model. The model may perform tasks on your behalf and even take over your personal device or computer. In June 2024 Apple introduced Apple Intelligence, which is supposed to use language models not as a product but a technology, operating more or less covertly to prioritise or summarise your emails, organise your calendar or automate your smartphone. ‘Apple is treating this as a technology to enable new classes of features and capabilities, where there is design and product management shaping what the technology does and what the user sees, not as an oracle that you ask for things.’²⁰ Large language models can be plugged into robots, including industrial robots, giving them artificial brains: ‘On the table sat three plastic figurines: a lion, a whale and a dinosaur. An engineer gave the robot an instruction: Pick up the extinct animal. The robot whirred for a moment, then its arm extended and its claw opened and descended. It grabbed the dinosaur.’²¹ It began with a conversation bot, but then there is probably no other way for the model to first manifest itself to a human user. One might recall the history of religion with its numerous examples of how a divine intelligence must manifest itself in human form in order then to reveal its true nature.