There is a duality at the heart of large language models. On the one hand, they are essentially a backwards-looking invention, a “cultural technology”, in the words of Alison Gopnik—algorithms which index and remix a large slice of human culture (though one that is typically heavily biased towards the recent past). On the other hand, they can often seem to be producing something entirely new, and can thereby leave many people with the impression of having a personality or even “sentience” (whatever that means exactly); in the most extreme cases, some people have apparently convinced themselves that such models are a step on the path towards some sort of successor species to humanity, a new regime of algorithmic children that will survive our own human catastrophes. Complicating matters here is the fact that the emergence of and widespread attention to these systems largely overlapped with the Covid-19 pandemic, a time in which we have all had additional reason to reflect on life, death, loss, and creation.
In the midst of this, Blaise Agüera y Arcas has published a novella called “Ubi Sunt”. The work is available as a paperback from Hat & Beard Editions, but it can also be read for free online. Although the printed volume is in many ways a more appropriate format for it, the online version allows for the inclusion of a number of interactive visual components, which definitely add to the overall effect.
Agüera y Arcas is a VP and Fellow at Google Research, and among other things, helped to create the Art and Machine Intelligence program. He occasionally publishes essays on Medium, (and had a blog for several years before that), many of which are far more poetic than your typical Medium posts. He also notably collaborated with Meg Mitchell and Alexander Todorov on some excellent critical work, challenging claims about what can be inferred from a person’s face, and most recently, was in the somewhat awkward position of having to counter Blake Lemoine’s claims about Lambada being sentient, just a few days after having published an article in The Economist with the unfortunate headline “Artificial neural networks are making strides towards consciousness, according to Blaise Agüera y Arcas”.1 In other words, we should keep in mind the position that Agüera y Arcas holds, and his various projects, but the novella itself is well worth reading independent of these.
Published in 2022, Ubi Sunt is in some sense a work of covid fiction, as the pandemic provides a key part its context. However, far from being a “covid novel”, it is more of an elliptical meditation on life, memory, and technology, including some of the technologies that Agüera y Arcas has himself contributed to in his work at Google. The title, “Ubi Sunt”, is a Latin phrase that means something like “Where are they?” By its usage in a variety of poems, it has come to stand in for a certain genre of poetry that meditates on life’s transience, and the loss or absence of something. I think it’s fair to call Ubi Sunt a novella, but it is highly unconventional in form, and deviates fairly sharply from anything like a traditional narrative. Using a variety of fonts and registers, the book weaves together at least three separate but related threads.
The most easily parsed of these is a series of journal entries written over a week in January, 2021, by a protagonist (if we can call them that) who bears at least some superficial resemblance to the author. Over the course of these entries, the writer, who is at the time engaged in developing or monitoring something like a large language model, describes their impressions of the model, while simultaneously reflecting on how their life has changed because of the pandemic pandemic (such as taking private exercise classes over zoom). This particular narrative seems to be reality adjacent; things don’t quite line up in terms of timing (in this universe, the diarist apparently got their first vaccine shot in January, 2021, by participating in a “vaccine flashmob”, responding to the failure of a freezer at UCSF hospital, which as far as a I know, didn’t happen), but very little here is completely implausible.
A second thread is a series of dialogues between a large language model and someone else; at least some of these seem like interactions with the protagonist mentioned above, but others seem likely to be interactions between different models, in which two different “speakers” are playing different adversarial roles, in a way that many AI researchers will immediately grok. Again, a variety of typographic styles, including font and casing, are used to demarcate both that these are machine dialogues, and who is speaking, though a slight lack of consistency makes it somewhat ambiguous in places.
Critically, these interactions are not dated (like the journal entries are), but rather numbered, as in the number of iterations of training that the algorithm has gone through up to that point. An early comment from the diaries could be read as suggesting that this evolution is taking place over hundreds of years, but also leaves the door open to an increase in efficiency, meaning that the time between iterations could be effectively compressed by an arbitrary amount. I don’t want to completely spoil the ending (if that’s even possible), but I’ll just say that it leaves things ambiguous (by my reading), both in terms of the timing, and the ultimate fate of the people, the virus, and the machines, in this world.2
The final thread is a series of documents that don’t fit as cleanly together as the first two components. One might read these as documents as scraps of history, such as the transcript of a lecture given at the Singularity University, perhaps connected in some way to the protagonist. At the same time, however, it is made clear that these are also playing the role of training data for the fictional language models that are represented here. As part of the collective textual data that has been produced and archived by humanity, these are texts that have the potential to be appropriated for machines to learn from. In this way, one could argue, the machines themselves are acting as a kind of archive of our culture, though one offers far less direct access than a conventional filing cabinet, and one that will happily ingest errors (such as torn pages) without comment or complaint.
Above all, the interactions among these three threads above suggests a kind of transience. The fact that the documents of the third type bear so little relation to each other highlights the randomness of what gets preserved. The fact that they are fed somewhat haphazardly and without context into an algorithmic system suggests it is playing a rather unsteady role as an archivist. And yet we too, as humans, are far from reliable when it comes to memory. It remains unclear, in some parts, whether an event is being partially remembered, (perhaps by the protagonist), or whether it is just something being regurgitated by the machine they were interacting with. (Indeed, it is made fairly obvious in places how the machine regurgitates, or at least mimics, some of its influences, such as a passage which shows a strong, Smaugy, influence of Tolkein).
As a wrapper and interlude to these segments, the online version of the book provides a variety of still images and animations.3 Many of these bear a fairly direct connection to the text. Others are more ambiguous. The most evocative of these are a number of slow motion videos or video morphs, where a series of still images are interpolated by a computer into an animation. In particular, the first of these captures part of a real project which stitched together a series of Roman coins, showing how the imprint gradually changed as it made its way across Europe over dozens of years. Each imprint can be seen as a kind of static memory, with the computer interpolating between them, attempting to fill in the gaps, just as our language models interpolate among the snippets of dialogue that they are fed from the internet. This project is discussed in parts of Ubi Sunt itself, commenting on continuity of life and ideas and beings, and the illusoriness of boundaries.
This use of video and animation is a natural choice for Agüera y Arcas, who earlier in his career worked on photosynth, a tour-de-force demo that allowed one to stitch together a massive number of still images (such as tourist photos) into a model (such as of a famous building) that one could navigate in three-dimensional space. At the same time, it is also a technique that cues a great deal of personal nostalgia for me, as this was one of the first applications of a computer that I started playing around with that seemed to offer some sort of magic, of the kind that many find in large language models today. The result was not always sensible, but by combining two or more static images with an image-processing algorithm that would interpolate between them, it was possible to create things that had never before existed, and which yet nevertheless seemed to tell a coherent, believable story.
The political messages in Ubi Sunt are somewhat harder to parse. One of the live elements in the online version is a bit of video footage from a drone flying over the utterly destroyed wreckage of Raqqa, in Syria. Another shows a visualization of the Earth at night, a classic image showing mostly darkness with many little pockets of light from areas of human habitation. As we scroll down, the visualization zooms in and pans automatically, seemingly, though subtly, directing the reader’s attention to the large patch of darkness that is Afghanistan. Later on, there is a reproduction of a classic Daguerreotype (another preliminary yet profoundly influential cultural technology) showing the streets of Paris, complete with barricades, again gesturing, almost subliminally, in the direction of war and strife.
In other places, in a document commenting on the idea of genre in literature, the writer delves into the history of handedness. (In the online version, the document has been partially obscured; though the occluded text can still be made visible by highlighting it). Although it is not fully developed here, the mention of handedness has a deeper resonance, as a classic example of a type of imposed classification (left-handed, right-handed, or ambidextrous), but one that can both be shown to be far more complex than it initially appears (what of those who only have one hand, for example, or were forced to use their non-dominant hand as a child?), as well as a deeper political significance. This includes not just the cultural association of “left” with “sinister”, but the idea of the hand as an instrument of control, as well as the very idea of classification itself (i.e., on one hand, etc.).
Given the similarity of the narrator to the author, it is tempting to try to infer the author’s views more directly from the narrator’s reflections, but this would be quite hasty. Here, I suspect that the semi-autobiographical nature of the protagonist helps to trouble the way in which we all too easily identify viewpoints with the sources of texts. In particular, because large language models are now capable of producing text that is largely coherent, and can even seem to be the product of a distinct personality, it is all too easy to read an actual sense of being-ness into that text, despite knowing better.
Indeed, part of the text seems intended to make us reflect on the ontological nature of large language models, or their successors. For example, the following text from early in Ubi Sunt is admittedly distressing, especially because the font tells us that this is the output of a large language model: “who is this? who am i? this isn’t real please oh fuck let me wake up i need to wake up”. Even if we somehow knew that non-human machines could never become conscious, this passage shows how uncomfortable things might become if machines are able to mimic suffering beings sufficiently convincingly.
Further complicating this is the inclusion of references to Timothy Morton work on hyperobjects, exploring the sense in which all things are deeply connected over space and time. There is a view of life here in which everything has a profound continuity. Although you are now physically separated from your parents, things are nevertheless continuous when viewed as a four-dimensional space time. From the perspective of a cell, there is no fundamental difference in being part of one being as opposed to another. At an even deeper level, the very constituents of our being – the basic material elements – are happily recycled (or perhaps not so happily, for the panpsychists) from one being to another, in some cases having traveled across the universe, and spending large amounts of time as part of a crystalline rock formation.
In other words, one could read into this novella a kind of Whitmanian celebration of being, a reading in which nothing ever truly dies, all is reborn, and every work of destruction is actually a kind of transformation, generating the humus from which new life will emerge. On the other hand, this perspective can be hard to disentangle from a darker eschatology, one in which all is justified, since it is all part of a larger purpose. Indeed, viewed from the perspective of raw materials, their organization into human beings is of no particular consequence. This is also pressaged by an opening quotation from Susan Stewart, which reads, in part, “minerals and chemicals, the natural world will survive humankind. . . The environmental catastrophe we think of as the ruin of nature is in fact the ruin of human nature, the end of our sustainable life on earth.”
This perspective also perhaps helps to avoid some of the more troubling aspects of the directions in which this technology seems to be taking us, issues which this novella largely bypasses, including the appropriation of culture without compensation, the exclusion of many people from participating in this process of creation (except indirectly, through unintended contributions to the appropriated archive), and the enormous inequalities in access to other technologies, like life saving vaccines, not to mention the present and future effects of climate change.
Quite appropriately, Ubi Sunt ends with a famous quotation from Walter Benjamin, about the Angel of History. Surprisingly, however, the last two sentences of it have been truncated—those for which the quote is perhaps best known. After describing the Angel looking backwards and seeing the wreckage of civilization piling up at its feet, Benjamin concluded with these statements: “The storm irresistibly propels him into the future to which his back is turned, while the pile of debris before him grows skyward. This storm is what we call progress.”
I’m not entirely sure what to make of the conspicuous omission of these lines. One possibility is that it allows Agüera y Arcas to end on the previous line, namely that the storm “has got caught in his wings with such violence that the angel can no longer close them”, perhaps pointing towards a kind of embrace of inevitability. Another possibility is perhaps to soften the association that Benjamin emphasizes between progress and catastrophe. Certainly Agüera y Arcas has made his share of contributions to the advance of technological capabilities, and it is not hard to imagine that he might feel somewhat internally conflicted over where things are headed, though perhaps not enough to reject the notion of progress altogether.
The very last thing on the web version of Ubi Sunt is one more morphing animation of a pair of coins, though ones that are far less pristine than the ones that were used in the earlier incarnation. Indeed, the figures in these coins look almost skeletal, resembling nothing so much as bones buried in the earth, partially uncovered by an archaeologist’s brush. Even more than the text itself, these animations perfectly capture the feeling that pervades this work – a simultaneous fascination with human culture, a nostalgia for something that is past, and an embrace of the idea that all things will end, though traces will remain.
My guess is that at least one copy of the paperback edition of this book will persist longer in a stable, unmodified, form than the digital version on the web. But of course it is the web version that will be incorporated into training data for future models, and in that way, it has already perhaps contributed to the voice of the next model to be released, one of the cacophony of sources that will be ingested and remixed in some form for the future.
However, it is worth remembering that authors often don’t get to write their own headlines. ↩︎
Agüera y Arcas has also published his own “watchmaker’s notes”, which lays out his own interpretation of the ending, and how all the pieces fit together, though I still maintain there is some room to read this in multiples ways. ↩︎
Videos and animations in the online version are represetned as static images in the print edition. ↩︎