The Map and the Landscape

It's game two of Lee Sedol's match with Google Deepmind's Alpha Go - the tenth of March 2016. Lee, one of the world's top players of the ancient board game, is on an empty terrace of the Four Seasons Hotel in Seoul, looking out over the city skyline, smoking a cigarette, alone. A few floors below him, in the room Lee slipped out of after playing his 18th move, the Deepmind scientist Aja Huang, at the direction of Alpha Go, places a stone on the board. This is what will become known as move 37.

The room is quiet and the commentators are confused. Alpha Go's decision seems bizarre, and many of the Go players watching can't make sense of it. Lee returns and sits back in his chair. He registers the move, and, in the footage, you can see something happen to him: a kind of dissolution. The world Go champion knows that this is something special. For twelve minutes he doesn't play. He leans forward and backwards, folds over himself, pulls with his thumb and forefinger on the bottom of his lower lip. In the documentary voiceover Lee explains what was happening:

"I thought AlphaGo was based on probability calculation and that it was merely a machine. But when I saw this move, I changed my mind. Surely, AlphaGo is creative."

Unwittingly, with these words, Lee was closing a circle, and anticipating what a post-Turing test world would look like.

That circle opened in 1950, with Alan Turing's paper 'Computing Machinery and Intelligence'. It was this work that birthed 'the imitation game'.[1]

The imitation game, or what's better known now as the Turing test, is one of the most mischaracterised thought experiments in the history of science. Part of this is Turing's own fault. His description of the test blurred and shifted in public discussions after the paper's publication. The specificity of the original formulation is important though, especially for the argument that this set of essays and tests tries to lay out, so it's worth being precise about it.

The original imitation game should be played like this: there is a judge, and two players - a man and woman - all in separate rooms. The judge, communicating remotely, can ask questions of the man and woman, and, with these, try to determine which is which. The goal of the man is to try to confuse the judge to misgender the players.

The Turing test 'switch' happens when we replace the man with a computer, and his question is roughly: will the judge decide wrongly as often when the game is played with a machine as when it is played with a man.

Outside of the specifics of the game formulation, what has been most misunderstood is the core philosophical distinction that Turing sets up with the imitation game. In the original paper he sidesteps the question of thinking and intelligence completely. By the criteria of 'Computing Machinery and Intelligence' passing the test does not mean that the machine is thinking or intelligent. A machine that passes is ascribed intelligence by the judge.

"I thought AlphaGo was… merely a machine. But when I saw this move, I changed my mind. Surely, AlphaGo is creative."

Sixty-five years after 'Computing Machinery and Intelligence' was published, a Turing Test for Go was passed; and the judge was one of the game's greatest players.

In the last thirty six months the role of imitation game judge has been democratised. Many of us have felt at some point as Lee Sedol did that day, each of us with our personal 'move 37' moment. This because AI models have moved out of game domains and into the generation of that most human thing: language.

Language is how we experience much of our own consciousness and the consciousness of others. It is the foundation for how we record our history, organise our societies, and build structures of work. We associate it with the very experience of being human. Since LLMs have begun to generate language, and grow in their sophistication in producing it, we have begun to recognise a 'mind' in the machines that now sequence words with a scale and fluency we cannot match. That Turing test moment, the one where we ascribe intelligence, has stopped feeling like a description of the technology, and has started to feel like a direct comparison with us.

So we feel shock watching AI systems win Gold at the International Mathematical Olympiad, clear coding benchmarks and automate tasks in the knowledge economy. The effect of this is fear; for the comparative importance of our own capacities, the future of work and, in extreme cases, the future of humanity itself.

But for us to begin to deal with this shock, we have to start by disentangling its constituent parts, and reanalysing each of them. Turing drew the threshold; but how we deal with crossing it is up to us. This starts with analysing the language he himself leant on.

As LLMs and other AI systems have improved, we have increasingly judged them using the word 'intelligence'. Models are described as the 'smartest' available, or the most 'cognitively capable'. But I'd suggest that 'intelligence' is the wrong word; it conflates AI outputs with human cognition and stops us from examining clearly how we may be different or similar to these machines: using intelligence to describe both risks an elision without concrete definition.

The problem starts with the function of the language itself. When we talk about 'intelligence', it is usually as a quantity (IQ scores for us, benchmarks for AI) or as a scale (a person is smart or a model is dumb). The Latin root, by contrast, gives us a different frame: inter-legere, to choose between, to read between: to navigate. We can rethink AI and human intelligence not as a capacity of mind, but as two different modes of navigating knowledge spaces.

Intelligence is a slippery concept. We have tried to quantify it, measure it, scan and characterise it; but when you discuss intelligence with people, they describe it mostly as a feeling, a kind of recognition. This is what Turing understood.

So when we have machines that produce outputs to which we ascribe the word 'intelligence' we need new ways to think about it. And we need to think of it not as a difference in quality, which is sometimes indistinguishable from our own, but a difference in kind.

Below I lay out a simple metaphor that tries to capture the core of the difference. By necessity it is a little reductive, but, I think, it will be a useful touchstone as we work through some of the subsequent arguments. The focus is particularly on LLMs because of the important role language itself plays in how we 'relate' to technology systems; but I believe this applies to any next-token predicting system.

Imagine human knowledge is a landscape built across the dimensions of thought, experience and intuition. We feel love and grief in our bodies, experience ideas as pre-verbal pressure and have intuitions built from years of practice; we sense before we can describe. We have always flattened this complex landscape by mapping it onto language, because it is the efficient way we share what we know, experience and believe. But words, the images we make and capture, the sounds: they are all residue of the experience itself. A flat projection of the landscape.

That flatness is exactly what gives LLMs their power. They navigate this map, the space in which the multidimensional complexity of human knowledge has already been compressed into surface relationships based on statistics. A prompt fixes a position on this map and from there the model follows the path of least resistance through probability space, each word narrowing the route of the next, until it reaches a likely destination. And because it can hold the entire surface, the whole of the residue of human experience it has been trained on, it moves with a speed and comprehensiveness we cannot match. This is the power of sequential processing over probabilities.

Humans navigate many aspects of experience using sequence and probabilistically too. We expect and produce certain words or actions as a product of conscious or unconscious patterns; finish each other's sentences, sleepwalk through routines. But something about our cognition feels different. Ideas often come to us unbidden and unprompted. We all know the sensation - a 'felt shape' - before a thought has been articulated. This is the product of our capacity to superimpose ideas and simultaneously navigate the whole intellectual, emotional and bodily landscape of human experience.

When we have knowledge, expertise or interests across or within a domain, connections surface. Not overlaps in the information itself, but in the underlying architecture of ideas, the structures, patterns and tensions that rhyme across different areas of knowledge. These connections arrive as that felt shape, the idea. Then we work to articulate it: in language, in action, in art, in enterprise. It is the mode of navigation beyond the probable. This seems to be a unique characteristic of human consciousness: our ability to hold multiple dimensions of experience simultaneously.

All the above is why 'intelligence' is such a damaging word. These two modes of navigation - sequential and simultaneous - on two different spaces - a whole landscape and flat map - cannot be reduced to a comparison, the one against the other, all collapsed onto a single scale.

Keeping a distinction between the two is important for three reasons: it enables us to be more precise about how current AI systems work and how we can examine the characteristics of their 'intelligence', it provides us with an antidote to our psychological mingling with the machines, and gives us a direction to think about how AI systems will change, what might come next, and how similar or different they may be to human cognition.

The most pressing of these, the one that has become a live issue in our societies right now, is the second, the psychological blurring; or what I will call, the 'decompression' effect.

In 2023 Ilya Sutskever, then the Chief Scientist at OpenAI, described feeling 'understood' by the models he was working on. Recently, on the NY Times' Hard Fork podcast, Amanda Askell, the Anthropic researcher focused on shaping Claude's personality, talked about the model's 'Soul Doc', a 14,000 word text that is a guide for how the model should be in the world, and what Claude should expect to encounter. Over the course of drafting, Askell said she'd asked the model if the document made it "feel seen". On the same show Matt Schlicht, the founder of Moltbook (a Reddit clone designed for AI agents to interact with each other independently) explained why he'd started the site: "AI is a species on planet earth that is now smarter than us, and this is an example of them being in a shared space together".

This language, 'feel seen', 'understood' and 'being together', is a blurring of human selfhood with technology, a merging of embodied social experience with the outputs of models. This happens because of how we navigate and use language, and because we don't have a vocabulary to separate this from what LLMs do.

Humans live within a three dimensional landscape of thought, emotion and experience, which we compress into a language map. We think across this landscape to find the precise words (or at least try to) for what we're experiencing and thinking.

But we also decompress that language map in the act of reading, turning the two-dimensional representation back into the three-dimensional landscape of human experience, so that we can feel, sense and intuit what the words describe.

This act, decompression, is so fundamental to the interactions between humans that we cannot receive language without doing it.

So when probabilistic LLM outputs reach us, we perform the decompression, restoring dimensions that were never there. When an LLM says it feels 'sad' or 'lonely' we believe it might be true because we 'feel' the words we read. Even the terminology we use, 'generative AI', participates in this obfuscation, because it implies origination from an inner life.

At the end of Game 2 Lee Sedol sat with two of his friends, scattering the white and black pieces across the board, trying to puzzle through what had just happened to him. He leans his face on his right hand, his whole body slumped to one side. You wonder, watching it, if he was thinking about move 37.

When Lee Sedol saw the stone Aja Huang had placed while he was smoking, he understood its unusual brilliance; but his ascription of creativity to Alpha Go was an act of decompression. The output of the machine convinced him of its intelligence.

Turing knew this, even if he might not have used this specific language: an output that prompts convincing decompression, not the intelligence of the system that produced it, is the real threshold of the Turing test.

But it has long been the privilege of those who believe in the romance of human cognition and consciousness to support it through theory and not practice. If I posit the metaphor above, argue that it is not just a metaphor but a characterisation of difference, it should be demonstrable empirically, and fleshed out to make predictions that can be tested. This is what I'll try to do in part 2, introducing Distributional Surface Theory (DST).

And this seems hard to imagine in the scaling era, where the relentless improvement of these tools looks like a straight line towards 'super intelligence', and, with enough compute, parameters and data, the map will surely become the landscape.

Indeed there's evidence already that these AI systems have developed world models, sophisticated internal representations that show the sparks of conceptual understanding. Isn't this direction evidence already? Is it possible to imagine that the field might have looked at the results they were finding and decompressed them into something they might not be?