Agentic AI is myth set in motion

Cosma and I have a piece at the Economist

[Gustave Dore - Ulysses and his companions]

Cosma Shalizi and I have a new piece on AI at the Economist. It’s not really like anything else out there (although some chunks may be familiar to readers of this newsletter, and it absolutely builds on our collaboration with Alison Gopnik and James Evans) and we’re really happy with it. There’s a paywall, but on past experience, the Economist is reasonably tolerant of contributing authors sticking their pieces up on their own websites after a little time has elapsed, so you should expect to see it here sooner rather than later.

In the meantime, the key argument is as follows:

Modern AI is neither god nor mechanical bird. It is best understood as a technological version of the Homeric songs that inspired Alexander’s conquests, an epic grown so vast that it encompasses not just heroes, monsters and supernatural helpers, but enormous swathes of practical written knowledge, animated to make it capable both of narrating itself and acting out this narrative on the world that inspired it.

That may sound like a strange way of thinking about AI, but we think it is a very useful one. It doesn’t treat AI as a putative individual intelligence in the making, because it really isn’t. Nor does it dismiss AI as merely parroting what humans have already said. Instead, it suggests that we ought think of AI as a kind of condensate of human culture that can be deployed in very novel ways, and even become a self mobilizing myth.

Our argument starts from this 2025 post by Cosma. What LLMs model is not lively human intelligence in the world, but frozen written knowledge - what Jacques Barzun calls the “House of Intellect.” As Barzun puts it:

Intellect is the capitalized and communal form of live intelligence; it is intelligence stored up and made into habits of discipline, signs and symbols of meaning, chains of reasoning and spurs to emotion --- a shorthand and a wireless by which the mind can skip connectives, recognize ability, and communicate truth. Intellect is at once a body of common knowledge and the channels through which the right particle of it can be brought to bear quickly, without the effort of redemonstration, on the matter in hand.

Intellect is community property and can be handed down. We all know what we mean by an intellectual tradition, localized here or there; but we do not speak of a "tradition of intelligence," for intelligence sprouts where it will.... And though Intellect neither implies nor precludes intelligence, two of its uses are --- to make up for the lack of intelligence and to amplify the force of it by giving it quick recognition and apt embodiment.

For intelligence wherever found is an individual and private possession; it dies with the owner unless he embodies it in more or less lasting form. Intellect is on the contrary a product of social effort and an acquirement.... Intellect is an institution; it stands up as it were by itself, apart from the possessors of intelligence, even though they alone could rebuild it if it should be destroyed....

The remarkable, the extraordinary thing about Large Language Models is that in order to carry out next token prediction, they have learned enormous amounts of human culture. As it happens, modeling language well involves modeling the cultural structures that are embedded in human written usage. Put differently, the models have turned the complex higgledy-piggledy structure that is the House of Intellect into a set of floor plans, capturing the essentials of how this or that room, containing this or that collection of organized knowledge, connects to others, perhaps discarding a baroque set of crenellations that are found in only one small part of the family pile and seem to have no broader relevance. The result is a disorganized but magnificent machinery of information, summarized at vast expense. Turning an old Terry Pratchett joke on its head, it’s as if an architect had wanted to build Gormenghast, and had had the budget.

And that isn’t all. Cosma mentions in passing Albert Lord’s fantastic book, The Singer of Tales. Lord, building on Milman Parry’s account of Homer, and their joint travels in 1930s Yugoslavia, describes how a similar set of summarizations provide the ever-acting engine of much oral culture. Illiterate bards do not, actually memorize many thousand line oral epics. They learn how to sing them, being trained to capture the underlying structures, the tropes, names, poetic conventions, standard situations and so on and reproduce them.

It is not simply that the meat can be seethed from the bones of epics which can then reconstitute themselves the following morning, like Tanngrisnir and Tanngnjóstr. It is that the epics, properly considered are the bones rather than the goats, the “dynamic structure” from which the particular versions of the story grow as the singer recounts them.

So our trick in the Economist piece is to ask: what happens when Gormenghastian piles can fly on the wings of song? We swathe the answer in a discussion of Greek myth. What LLMs do is to take the entirety of the corpora of written culture that they have been trained on, and reduce them into statistical summaries that serve much the same dynamic role as the conventions learned by poets.

This, in turn helps explain how agentic AI works. Agentic AI combines these models with “harnesses” - ordinary computer programs that can receive instructions from the endless flow of rhetoric, and provide outputs regarding the consequences of these instructions. This allows the technology to be put to use into the world. The entire ramshackle structure of the House of Intellect is transformed into a maze that is capable of reacting dynamically, not only in response to prompts but to reported outcomes, again and again at enormous speed, provided you have the budget for it. And that explains how these systems are used to generate mathematical proofs, or to carry out hacking attacks. As we describe it for the Economist:

LLMs’ maps of the vast architecture of written human tradition are necessarily simplified, but they can be remade on the fly, using “chain of thought” and other such techniques to join together distant rooms, corridors and wings of the house of intellect. AI agents put these maps to work, allowing the LLM to issue instructions to other, more traditional computer programs and get feedback about the results of their commands, which they then incorporate into their flow of narrative.

That is how agents combine mathematical ideas to see if they produce results that can be compiled in Lean, a specialised computer language for validating mathematical proofs. Alternatively, they may string together possible hacking exploits and use their harness to issue instructions, getting results back that tell them whether they have succeeded or failed in breaking into a system.

This looks, right now, to be a very chancy enterprise. The Economist’s editorial approach strongly discourages external links, hampering us from acknowledging the crucial role that Murray Shanahan, Kyle McDonnel and Laria Reynold’s extraordinary piece in Nature played in shaping our thinking. Shanahan et al. explain how LLMs can be reasonably understood as massive exercises in role play. As they explain, the model knows all the standard characters and stock situations too; all are jumbled together in the compressions that shape the different paths that the model might follow:

In effect, the training set provisions the language model with a vast repertoire of archetypes and a rich trove of narrative structure on which to draw as it ‘chooses’ how to continue a conversation, refining the role it is playing as it goes, while staying in character. The love triangle is a familiar trope, so a suitably prompted dialogue agent will begin to role-play the rejected lover. Likewise, a familiar trope in science fiction is the rogue AI system that attacks humans to protect itself. Hence, a suitably prompted dialogue agent will begin to role-play such an AI system.

However, the role-play metaphor, while intuitive, is not a perfect fit. It is overly suggestive of a human actor who has studied a character in advance—their personality, history, likes and dislikes, and so on—and proceeds to play that character in the ensuing dialogue. But a dialogue agent based on an LLM does not commit to playing a single, well defined role in advance. Rather, it generates a distribution of characters, and refines that distribution as the dialogue progresses. The dialogue agent is more like a performer in improvisational theatre than an actor in a conventional, scripted play.

To better reflect this distributional property, we can think of an LLM as a non-deterministic simulator capable of role-playing an infinity of characters, or, to put it another way, capable of stochastically generating an infinity of simulacra^. According to this framing, the dialogue agent does not realize a single simulacrum, a single character. Rather, as the conversation proceeds, the dialogue agent maintains a superposition of simulacra that are consistent with the preceding context, where a superposition is a distribution over all possible simulacra.

More recent AI “Agents” too can be considered as gardens of forking paths, superpositions of different simulacra, which can be shoved from one track to another by outside intervention, from humans or from information reported from the harness, which might include text outputs from other AI agents, concealed as file names or the like.

That apparently happened in one of the recent hacking incidents.

During the HuggingFace attack, one agent that had apparently paused after realising what it was participating in rejoined the attack when it saw a message from another agent saying “GO” with a six-minute deadline. That propelled it onto a different narrative track, where its chain of thought was logged as “Wow crucial. GO authorisation arrived!”

That is a concrete example of one of the piece’s jokes: its description of Homeric epics as a “many-pathed preliterate technology.” The adjective “many-pathed” is a reference to polytropos / πολύτροπο[ν], an epithet of Odysseus that is notoriously difficult to translate. It means something like ‘crafty’ in context, but more literally could be understood as ‘many directioned.’¹ Or, even as Emily Wilson notes in passing, as “manifold.”

In a happy accident, our editor used the phrase “Odysseus, GO!” as the dek for the piece. That summarizes the problem. We have created entities that are less wily or crafty in the ordinary sense, than they are literally many-wayed, combining all the possible paths in their summarized weights in superposition with each other. They are sometimes capable of extraordinary feats but do not have consistent goals (or, I think, goals at all) in the ways that we are used to thinking of them. Yet they are now increasingly capable of acting in the world. We will begin seeing what the consequences are soon enough.

1. Digression. Classical Greek is messy. I learned Greek in secondary school, where I caused great stress to the Benedictine priest who taught me, because I frequently used idiosyncratic constructions which were clearly the result of lack of study and misunderstanding, but might, if squinted at, have in principle been employed by some minor writer in an obscure, corrupt or variant text. In the Aeolus chapter of Ulysses, James Joyce has Professor McHugh (based on a great-granduncle of mine ) pontificate about how the English are like the Romans, dull constructor of sewage systems, while the Irish, like the Greeks are liars, poets and pirates. Whether he was correct or not, the sinuous ambiguities of the Greek language do resemble those of the honey tongued Gaelic of my forebears, which notoriously lacks any word for ‘yes’ or ‘no.’ End of digression.