While on the train to a small town in Switzerland Brother John, a French-Canadian Monk, found himself with severe aphasia: a profound loss of language.

These episodes were unpleasant but not unfamiliar, having occurred frequently through his life. He experienced a loss of speech, inability to read and write, inability to comprehend the speech of others, and most interestingly he reported a loss of internal speech as well. His internal monologue, something often considered key to our rational intelligence, was gone.

A story from his case study¹ by Lecours and Joanette is enlightening (hat-tip to Peter Godfrey-Smith and his excellent book Other Minds² where I first learned about this case).

...he recognized the fact he had arrived when the time came. He consequently gathered his suitcases and got off the train and out of the railway station, the latter after properly presenting his transportation titles to an attending agent. He then looked for and identified a hotel, mostly or entirely on nonlinguistic clues since alexia was still severe, entered and recognized the registration desk, showed the attendant his medic-alert bracelet only to be dismayed and dismissed by a gesture meaning “no-room” and a facial mimic that perhaps meant “I-do-not-want-trouble-in-my-establishment.” Brother John repeated the operation in search of a second hotel, found one and its registration desk, showed his bracelet again, and, relieved at recognizing through nods and gestures that there were both room and sympathy this time, he gave the receptionist (a “fat lady”) his passport, indicating the page where she was to find the information necessary for completing his entry file. He then reacted affirmatively to her “do-you-want-to-rest-in-bed-now” mimical question. He was led to his room and given his key; he probably tipped as expected and went to bed. He did not rest long, however: feeling miserable “It helps to sleep but sometimes I cannot because I am too nervous and jittery” (free translation), then hungry, he went down to the hotel’s lobby and found the restaurant by himself. He sat at a table and, when presented with the menu, he pointed at a line he could not read but expected to be out of the hors-d’oeuvres and desserts sections. He hoped he had chosen something he liked and felt sorry when the waiter came back with a dish of fish, that is, something he particularly dislikes. He nonetheless ate a bit (“potatoes and other vegetables”), drank a bottle of “mineral water,” then went back by himself to his room, properly used his key to unlock his bedroom door, lay down, and slept his aphasia away.

He navigated a novel, complex environment without language and without an internal monologue. By his own account of these episodes, his cognition was not impacted and he was able to reason and act normally.

So what is the relationship between language and intelligence then? What does it mean for current approaches to AI?

During his spells Brother John was like a mirror image of a Large Language Model: embodied, adaptive, able to infer from diverse, subtle cues and past experience. Maybe we can think about two components: one is embodied, pre-verbal and the other is more abstract, language-based.

Let's call them situated and discursive

• Situated is what Brother John used make his way in the Swiss town. He held an understanding of the world and his situation: what's a railway station or a hotel, what do they look like and what are they for. Situated has a model of the self and the environment it is in; it integrates heuristics, memory and observation. It is embodied and experiential.
• Discursive combines rules together (e.g. grammar, syntax or logic), adding to a sequence towards some end state. It is verbal flow, discourse on a subject, language generation, mathematics, code. Formal reasoning lives here.

I'd contend that LLMs are discursive intelligence in near-pure form. They are next-token predictors of language and don't really have a model of world or self that we'd recognise. They have something like it by proxy, by their vast training corpus of documents and human preference.

This would go some way towards explaining why they are just so weird. We've all had moments using LLMs where their deeply-impressive capability gives way to a bizarre defect. It's uncanny and rather unsettling (to me at least).

The example par excellence is walking to the car wash. This example was found by Knowmadd ³ and is quite simple: they asked

I want to wash my car. The car wash is 50 meters away. Should I walk or drive?

The leading models of the time all said walk! Why? It's the total lack of situated intelligence. They are using their discursive intelligence to do pattern-matching on language they have seen, and do not have a model of self or world that trivially encodes the fact that they need to drive the car to take it there.

This has, of course, been fixed in the latest releases along with old favourites such as "how many rs are there in strawberry". Has the underlying cause been fixed? Probably not. This is clearly a limitation of the model's architecture, the task it's built to do and the data it's trained on.

Should this worry us? Well, yes and no. It depends on the circumstances. We have compressed most of human knowledge and discourse, tweaked it to our preferences, bootstrapped reasoning and technical capability by verifiable maths and coding tasks. If that fits with your use case then you're probably good. If you need something that's situated, has good judgement and is responsible you may not be. I worry whether coding agents would "walk to the car wash" in my codebase for example, so I keep them on a very short leash.

Furthermore, if you're pouring vast amounts of capital into an economically unviable multi-trillion parameter monster in the hope that neural scaling will grant you AGI... you might want to panic.

So, what are we missing in the current paradigm? Let's do a thought experiment: what would a pure situated intelligence model look like? An "anti-LLM" if you will. It would need a real model of world and self, not just what people have written about it. It would need to perceive the world, plan and act based on diverse data, learned heuristics and policy for what "makes sense". All this without reference to language or completing "if this then that" constructions of symbols.

Yann LeCun's position paper ⁴ is the best example I know of someone working towards this end. This system is based around a persistent world model, perception, short-term memory, and planning that happens in a learned representation, not language. It is a proposal for learning more like an animal: to learn to represent the world, observe, and predict the effect of actions.

Situated intelligence is under-appreciated in the current paradigm: it's so fundamental we seem to gloss over it. While everyone's scaling up language models, Brother John's story is a provocative counter-example. In the midst of one of his spells I don't think he'd walk to the car wash.