Imagine a future in which human beings communicate not by sound or text, but by direct presentation of their mental states to one another: a kind of high-fidelity neural intersubjectivity. Ted Chiang, in his typically understated way, gestures at this possibility in his short story, The Evolution of Human Science, where the posthuman “metahumans” converse in a language called 'Digital Neural Transfer', which I’ll refer to here as Neuralese. For them, traditional language — ours — has become obsolete: a quaint, clumsy bottleneck between minds.
It’s worth asking: if we were to develop such a language ourselves, what would it actually be? What kind of structure could support brain-to-brain communication without that cumbersome detour through words?
The usual science-fictional telepathy tropes rarely interrogate the mechanics - which are anything but trivial. To “receive” a brain state is not the same as understanding it. Brains are not USB sticks; states are not portable files. Drop my active thalamocortical configuration into your hippocampus and you’d get scrambled eggs rather than shared insight.
Language is the workaround. A lossy compression algorithm designed to serialize high-dimensional mental activity into linear utterances, transmitted as air vibrations or squiggles, and then reconstructed with great effort by the recipient.
This process requires a specialised left-hemispheric subsystem with carefully evolved grammatical faculties, symbol mapping, error correction, and a fair bit of cultural priors.
It's remarkable that it works at all. And frequently it doesn't - hence the tragedy and comedy of human miscommunication.
But now consider an analogy not from speculative biology but from machine learning. One of the more intriguing developments in modern AI systems is the process of knowledge distillation: transferring the capabilities of a large, complex model (the “teacher”) into a smaller, more efficient one (the “student”) by having the student imitate the outputs — and sometimes the internal behaviour — of the teacher.
Crucially, this isn’t done by copying weights or architecture. The teacher is queried across a dataset, producing soft outputs called logits[1] which are more informative than hard labels.
These include not only the correct answer, but the confidence profile — the degree to which the teacher favours one answer over others even when it is wrong. The student then tunes itself to match these distributions, layer by layer.
What results is not a clone, but a resonance: the student model develops an internal geometry of meaning and behaviour that mirrors the teacher's, not propositionally, but probabilistically.
Now transpose this to Chiang’s 'Neuralese'.
If metahumans — or AIs, or aliens — communicate by Neuralese, perhaps they are not transferring states, but something closer to distilled representational trajectories.
Instead of sending a sentence, they send a compressed map of how a particular stimulus would modulate their cognitive manifold. The recipient doesn’t “read” this as text. They instantiate it, simulate it, and interpolate it into their own architecture. Meaning emerges, not through syntax, but through pattern alignment in an abstract vector field.
This is a far cry from that old dream - Richard Montague's grammar. There are no logical formulae, no intensional operators, no semantic trees. There is no “language” in the classical sense at all. What exists instead is something far more alien to human linguistic intuition: a dynamic, annotated graph structure, internally consistent to a model but utterly opaque to traditional semantics.
Technically, Neuralese might resemble a dynamically evolving graph:
- Nodes as embedding vectors, attention points, or positional encodings.
- Edges as transformations — weighted and directional — capturing statistical dependencies, activation flows, or relational cues.
- The entire construct evolves dynamically throughout the computation, reshaping as new inputs arrive.
And yes, when this is transmitted — say, from teacher to student model, or one metahuman cortex to another — it is serialised not as a logical expression, a sequence of wffs, but as tensors: parameter matrices, gradients, logits, and perhaps some trace of attention priors.
To a logician trained in the Montague tradition, this would look less like a language and more like a murmuration of algebraic shadows — intelligible only in motion, and only with the right decoder.
Imagine Wittgenstein confronting an attention heatmap.
And now the final speculation: what if this is how the aliens do it? Not with an imagined intergalactic Esperanto, but Neuralese of their own evolution. Perhaps their anatomy — say, a fluidic distributed nervous system or a colony-intelligence of bio-electric cells — never required the bottleneck of linear speech. Perhaps they evolved distillation before dialect, simulation before syntax. They don’t translate ideas into words. They translate minds into probability flows - and then flow into you.
The irony, of course, is that the only way we’ll ever learn to communicate with them may be to train ourselves — or our machines — to first 'inhabit' Neuralese. And in doing so, we may no longer recognise what it ever meant to 'speak'.
[1] A logit is the raw score - a real number in the range (-∞, ∞) - which a machine learning model gives before it translates that score into a probability in [0, 1]. It’s not yet a judgment about how likely something is — it’s more like the model’s unfiltered impulse or leaning, a kind of pre-conscious hunch about what class or label applies.
When models like LLMs communicate during training or distillation, they often share these logits rather than final probabilities, because logits carry more nuanced information about their internal reasoning — including how uncertain they are, or what alternatives they’re still considering. You can think of logits as the neural murmurings of a model before its thoughts are rounded into speechlike categories.
