Thinking with LLMs: A Reflection

A praxis for thinking with AI through recognition, resistance and recursive form-finding

I increasingly think that the way I work with AI resembles carving into marble.

But it is not the marble of Michelangelo, with a single statue already waiting inside the stone. It is a generative marble containing an indeterminate number of possible forms. Each cut reveals something, but each cut also changes what can subsequently be seen. The object is not simply uncovered. It is progressively discovered and constituted through the act of working on it.

This is different from how human–AI interaction is commonly described. The dominant model remains one of instruction and production: the human specifies a task, the machine generates an answer, and the quality of the result depends on the precision of the prompt.

That is not how much of my work with AI happens.

I often begin without a fully formed proposition. I begin with a hunch, a tension, a fragment of language, a causal relationship I can sense but cannot yet explain, or a structure whose significance I recognise without being able to describe its full architecture.

The purpose of the dialogue is not merely to express an idea I already possess. It is to discover what the idea is.

Recognition before specification

The central mechanism is recognition before specification.

A human being may be unable to specify an object in advance while still being able to recognise whether a provisional version is faithful to it. We frequently know more than we can initially articulate. We can sense that something is too shallow, too conventional, falsely resolved or organised around the wrong centre of gravity before we can explain precisely why.

AI has an inverse capacity. It can often articulate more than it can judge. It can produce structures, arguments, distinctions and formulations at great speed, but it does not carry the situated stakes through which their significance can finally be assessed.

This creates a productive asymmetry:

The human can judge more than they can initially articulate. The machine can articulate more than it can finally judge.

The first machine-generated response is therefore not best understood as an answer. It is an epistemic surface: a provisional object made visible enough for human judgment to encounter it.

That encounter produces information.

I can see that the argument has become too managerial. That the language has domesticated the original insight. That an apparently neat distinction is actually false. That the causal structure is missing. That the text has become fluent by removing the very tension that gave rise to the inquiry.

The next input is then not simply an instruction to improve the prose. It is a further act of discovery. In recognising what the object is not, I become more capable of specifying what it might be.

The prompt as a cut

Different prompts perform different sculptural operations.

Sometimes I narrate the skeleton. I name a small number of relationships that I think must be load-bearing, without yet knowing the complete body they will support. The machine externalises those relationships into an inspectable structure. Once visible, the skeleton can be tested: are the joints in the right places? What is missing? Which apparent branch is actually the spine?

Sometimes I ask the machine to “think harder” or “deepen this.” This is not a request for greater length. It is a test of generativity. Does the originating hunch produce consequences across different scales and domains? Is it the seed of a coherent theory, or merely an attractive phrase?

Sometimes I ask for an audit. This introduces resistance. The object is tested for causal leaps, hidden assumptions, category errors, internal contradictions and unacknowledged exclusions. The question is no longer simply whether the idea is interesting, but whether it can bear weight.

And sometimes the audit itself has to be audited. I might ask: what norm or gravity is driving this assessment? This matters because conventional standards of reasonableness, evidence or institutional legibility can quietly pull a genuinely new idea back into an old ontology. What appears to be analytical rigour may actually be the enforcement of inherited categories.

The work is therefore not only to improve the object. It is also to examine the field of judgment within which the object is being formed.

Fighting gravity

Every model has a gravity.

Large language models are pulled towards statistically familiar language, established categories, recognisable genres and widely available patterns of argument. They tend towards smoothness, symmetry and resolution. They are exceptionally capable of producing language that feels complete.

But completion can be a failure.

An idea may become more fluent by becoming less true. A radical proposition may become more legible by being reduced to a familiar policy category. A living tension may be translated into a framework that organises it neatly while removing its force.

Working well with AI therefore requires the human to identify and resist the model’s default attractors.

This is what I mean by fighting gravity. It is the work of replacing the model’s statistical centre of gravity with the inquiry’s actual centre of gravity.

This resistance cannot be one-directional. The machine must also be used to fight the gravity of the human’s own narrative. It can identify seductive claims, missing boundary conditions, weak causal links and concepts that have been extended beyond their legitimate range.

The relationship is most valuable when neither the human’s intuition nor the machine’s articulation is allowed to pass without resistance.

Without human resistance, AI domesticates the idea.

Without machine resistance, AI merely decorates the intuition.

The intelligence is produced through the friction between them.

Tone is not surface

Changing tone is another way of rotating the object.

A reflection, a provocation, a strategic framework, a concept note, a technical requirements paper and a public essay are not simply different packages for the same content. Each genre makes different properties of the thought perceptible.

A reflection can preserve ambiguity and interior movement. A provocation can reveal the fracture an argument is trying to open. A framework exposes components and relationships. A technical paper forces the articulation of conditions, responsibilities and verification. A public essay tests whether the thought can travel beyond the context in which it was formed. A visual can reveal topology that remains almost impossible to perceive in linear prose.

Tone therefore changes more than how an idea sounds. It changes how the idea behaves, what it is required to account for and what relationship it establishes with its reader.

Changing tone is like placing the sculpture under a different light. Contours that were previously invisible become apparent. But it can also be more fundamental than relighting. Sometimes changing the genre changes the material law under which the object is being formed.

The method

The process is recursive rather than linear:

Hunch → provisional form → recognition or dissonance → new cut or counterforce → rotation → audit → temporary stabilisation

It usually begins before the thesis exists.

The hunch is externalised early, rather than protected until it appears complete. The machine produces a provisional structure. That structure is read diagnostically: not “Is this well written?” but “What has this version revealed about the thought?”

The object is then rotated. Its implications are explored. Its scale is changed. Its causal structure is tested. It is rendered for different audiences and through different genres. Missing information is introduced. False resolutions are reopened.

Eventually, the work moves towards compression. Temporary scaffolding is removed. Repetition is stripped away. The argument is required to carry its own weight.

But this is only a temporary stabilisation. The object is complete enough for its present purpose: to circulate, to invite challenge, to organise action or to become the starting material for another inquiry. It is not treated as the final form of the thought.

The output returns to the field as new material.

What each participant contributes

This is a form of co-discovery, but it is not a symmetrical relationship.

The human holds the stakes. The human brings lived context, directional judgment, moral gravity, taste, veto and responsibility. The human decides what matters, what must not be flattened and when a provisional form is sufficiently faithful to enter the world.

The machine supplies generative surplus. It can materialise tacit structures, produce alternative formulations, follow consequences, compare adjacent models, move between scales and rapidly recompose the whole.

The machine enlarges the field of forms that can be inspected. The human remains responsible for determining which forms deserve to become consequential.

This distinction is critical. AI does not remove the need for authorship. It relocates authorship away from the manual production of every sentence and towards the construction of direction, judgment, criteria and closure.

The human is not merely editing machine-generated text. The human is governing a field of possible articulation.

Carving the intelligence of the pair

Each iteration acts on two objects at once.

The first is the idea itself.

The second is the temporary intelligence of the human–machine pair.

Across repeated exchanges, the dialogue develops a local vocabulary, a memory of important distinctions, an understanding of forbidden simplifications and a shared standard of what adequate depth feels like. Certain concepts acquire precise meanings. Earlier arguments become available as material for later ones. The interaction becomes capable of carrying greater conceptual weight.

The pair is not simply producing more content. It is developing a situated capacity to recognise and construct particular kinds of form.

This may be the more durable asset. The final text matters, but so does the relational capacity through which such a text can be made. The history of revisions, disagreements, abandoned formulations and changes of gravity is not incidental to the result. It is the process through which the result became thinkable.

Beyond content production

As machine-generated articulation becomes abundant, producing plausible language becomes progressively less valuable.

The scarce capacity is judgment: the ability to distinguish a live form from a merely fluent one; to identify the gravity an inquiry must preserve; to know when coherence is clarifying and when it is concealing; and to remain responsible for what is allowed to stabilise and circulate.

The relevant unit of work is therefore not the prompt or the answer. It is the evolving object together with the sequence of judgments through which it was formed.

This has implications beyond individual writing. If organisations use AI only to accelerate document production, they may generate enormous volumes of articulated knowledge without increasing their collective capacity to know. The polished output will conceal the assumptions, disagreements and decisions through which its apparent coherence was produced.

A more serious practice would make the form-finding process itself available: the skeleton, the alternative models, the rejected assumptions, the counterarguments, the changes of gravity and the reasons for provisional closure. The task would no longer be merely to produce knowledge, but to construct the conditions through which people can recognise, challenge and revise it together.

The statue is not waiting inside the machine.

Nor is it already complete within the human.

It becomes available through a recursive encounter between hunch and articulation, possibility and judgment, generation and resistance. Each response makes a possible form visible. Each act of recognition changes the next cut. And through that movement, we do not simply find better words for what we already thought.

We discover what the thought is capable of becoming.