Skip to content

Deep Structure⚓︎

Why do presentations exist as artifacts at all?

The sociologist Bruno Latour offers a useful concept. He calls certain objects immutable mobiles: things that can travel across distances and contexts while remaining stable. A map is an immutable mobile. So is a legal contract, a scientific paper, a quarterly report. These artifacts enable coordination without co-presence. You don't need to be in the room; the document carries the message. It arrives the same as it left.

Presentations function as immutable mobiles. The deck you send to investors, the training materials distributed to five hundred employees, the pitch signed off by legal—these need to arrive identical to how they departed. The artifact isn't just for viewing. It's for traveling.

This explains why organizations resist purely ephemeral content. The resistance looks like conservatism but functions as coordination. When the same artifact reaches everyone, you can discuss it, reference it, audit it, improve it. When every viewing generates something different, coordination mechanisms break. The investor sees a different deck than legal approved. The employee in Sydney learns different content than the employee in London.

Any approach to presentations must grapple with this. The artifact isn't optional decoration on top of the communication. For many uses, it is the coordination mechanism.

Surface-Level Operation⚓︎

PowerPoint gives you a canvas: a fixed-dimension rectangle, typically 16:9. You place text boxes, images, shapes. You manipulate visual elements directly, arranging them spatially until the slide looks right. Then you do it again for the next slide, and the next, until you have a deck.

This is surface-level operation. You work directly on the output, the visual surface that viewers will see. Your actions map one-to-one onto pixels. Drag this text box here; it appears there. The tool is a direct manipulation interface for visual surfaces.

Gamma, a presentation startup valued at over two billion dollars, improved surface-level operation by replacing slides with cards. Cards flex and scroll instead of sitting in fixed frames. Embedding live content becomes possible. AI generates an initial surface from your prompt. But the fundamental mode remains: you work on surfaces. You edit the generated cards. You refine the visual output. You produce an artifact, a shareable link to a deck you created.

PowerPoint with Copilot now generates surfaces too. You prompt, it produces slides. The AI race is a wash. The difference between PowerPoint and Gamma is the format (fixed slides versus flexible cards) and distribution (files versus links). Both differences matter. But both tools share an assumption: the user operates at the surface level, manipulating or editing the visual output that viewers will consume.

A Different Level⚓︎

In the 1950s, the linguist Noam Chomsky proposed a distinction that clarifies what it would mean to operate differently.

Chomsky distinguished deep structure from surface structure. The surface structure is the actual sentence you hear or read, the specific words in their specific order. The deep structure is the underlying meaning, the relationships between concepts that the sentence expresses. Between them sit transformation rules, the grammar that converts deep structure into surface.

The same deep structure can yield different surfaces. "The dog bit the man" and "The man was bitten by the dog" express the same underlying meaning. Different transformation rules (active voice versus passive voice) produce different surface forms. The meaning travels; the expression varies.

Applied to presentations:

Linguistic Concept Presentation Equivalent
Deep structure Intent, content, audience, purpose
Transformation rules Patterns that shape communication
Surface structure The visual output viewers experience

PowerPoint operates at surface. You manipulate the visual output directly. Gamma operates at surface with AI assistance. You prompt, then edit the visual output.

What would it mean to operate at deep structure?

You would specify what you want to communicate, to whom, and for what purpose. You would not manipulate visual elements. You would not edit cards or slides. The system would apply transformation rules (patterns that shape effective communication) and generate the appropriate surface. Different contexts might yield different surfaces from the same deep structure, just as the same meaning yields different sentences depending on transformation rules applied.

Transformation Rules as Pattern Languages⚓︎

The transformation rules aren't arbitrary. They're the accumulated knowledge of what makes visual communication work.

Edward Tufte spent decades codifying principles for data visualization. Maximize the data-ink ratio; every visual element should convey information. Show comparisons; always answer "compared with what?" Integrate evidence; keep explanations adjacent to what they explain. These aren't aesthetic preferences. They're transformation rules: given content that needs to be visualized, here's how to convert it into effective visual form.

Pedagogical research offers transformation rules for learning. Progressive disclosure: don't overwhelm; reveal complexity gradually. Multiple representations: show the same concept through analogy, visual, code, and first principles. Retrieval practice: test understanding, don't just present information. These rules transform content into learning experiences.

Narrative structure provides transformation rules for persuasion. Setup-conflict-resolution. Problem-solution. The hero's journey. Compare-contrast. These patterns shape how arguments unfold, how evidence lands, how conclusions feel earned.

Christopher Alexander called this a pattern language, a generative grammar of composable patterns that produce valid designs. The patterns aren't templates. They're relationships that recur because they solve problems humans actually have. "Light on two sides of every room" isn't a room design; it's a pattern that can be instantiated infinitely many ways, each valid if it honors the relationship.

A generative presentation layer would internalize these patterns as transformation rules. Given deep structure (intent, content, audience, purpose), the rules would generate appropriate surfaces. Different content would yield different outputs, the way different building programs yield different applications of "light on two sides." But every output would embody patterns that make communication effective.

Whose Patterns?⚓︎

The preceding sections treat transformation rules as though they're discovered rather than designed. Tufte's principles, pedagogical frameworks, narrative structures, all presented as "what makes communication effective." But effective for whom?

The media philosopher Vilém Flusser distinguished traditional images from technical images. Traditional images bear human intention directly; technical images are produced by apparatuses according to embedded programs. The danger isn't that technical images are bad; photography can be art. The danger is that people forget the program. They mistake the output for unmediated reality. The apparatus becomes invisible; its assumptions become nature.

Transformation rules are programs. "Maximize data-ink ratio" encodes Tufte's values: his preference for density, his assumptions about what readers can parse. "Progressive disclosure" assumes a theory of how understanding develops. "Problem-solution narrative" privileges certain story shapes over others. None of this is wrong. But none of it is neutral either.

The political theorist James C. Scott named a related concern: legibility. High modernist planning makes complex systems readable, countable, administrable. It produces standardized outputs, measurable outcomes, comparable results. What it loses is mētis—local, practical, contextual knowledge that resists systematization. The forest optimized by the forester's grid may be ecological disaster. The city planned on rational principles may be unlivable in ways the planners couldn't model.

Communication standardized through transformation rules gains legibility. Every output embodies proven patterns. Effectiveness becomes measurable: did the learner pass? did the pitch convert? did the board approve? What this can't capture is effectiveness that defies the patterns: the presentation that worked because it broke the rules, the learning that happened through productive confusion, the pitch that landed because it felt raw rather than polished.

The response isn't to abandon transformation rules. Scott doesn't argue that legibility is avoidable or always bad; states need to administer, organizations need to coordinate. The response is to remain aware that rules are choices, not discoveries. Make the pattern layer visible rather than invisible. Surface the assumptions so users can contest them. When someone asks "why did the system structure it this way?" the answer should be available, not hidden in opaque weights but stated as principle, open to challenge and revision. The apparatus shouldn't disappear into the outputs it produces.

Use Case Modes as Rule Configurations⚓︎

This reframes what "learning mode" and "presentation mode" mean.

They're not separate tools with separate codebases. They're different configurations of transformation rules applied by the same underlying capability.

Learning mode weights pedagogical rules: progressive disclosure, knowledge checking, multiple representations, spaced repetition. The deep structure (content + audience + purpose) passes through these rules to generate an interactive learning experience.

Presentation mode weights persuasion rules: narrative arc, evidence integration, visual hierarchy, call to action. The same deep structure, different rule weights, different surface.

Pitch mode weights credibility and urgency rules: social proof, scarcity, authority signals, clear ask. Again, same capability, different configuration.

One capability, multiple modes. The modes select which transformation rules dominate. This is architecturally cleaner than building separate tools, and it explains why the outputs feel related: they share deep structure and core patterns, differing only in emphasis.

The Layer Architecture⚓︎

Architecturally, this suggests three layers:

Intent layer: Where humans operate. You express what you want to communicate, to whom, for what purpose. This layer translates intent into specifications the substrate can execute, filtered through transformation rules.

Transformation layer: The patterns that shape effective communication. Tufte's principles. Pedagogical frameworks. Narrative structures. This layer doesn't produce anything directly; it constrains and guides how deep structure becomes surface. Modes (learning, presentation, pitch) are configurations of this layer—named presets that weight different rules. Learning mode weights pedagogical patterns; pitch mode weights credibility and urgency patterns. Same layer, different settings.

Substrate layer: The engine that produces interactive experiences. Next.js, React, Tailwind. Technologies that render web applications. This layer knows nothing about presentations or learning. It knows how to turn specifications into working interfaces.

The layers compose. You express intent ("help my team understand this concept"). The transformation layer, configured for learning, weights pedagogical rules and generates a specification. The substrate layer renders an interactive learning experience.

Same person, different context ("convince the board this matters"). The transformation layer reconfigures: different rule weights, different specification, different output. The capability remains singular.

Resolving the Durability Question⚓︎

Earlier I noted that organizations need immutable mobiles, stable artifacts for coordination. Doesn't operating at deep structure, with generated surfaces, break this?

Not if you distinguish what needs to persist.

The transformation rules persist. Think of brand guidelines. An organization coordinates around "follow the brand" without specifying every output. The guidelines are stable; the outputs vary infinitely. Similarly, "use the compliance training curriculum" becomes a coordination point. Legal approves the curriculum structure, the required content, the assessment criteria. Each employee's experience differs (pacing, examples, remediation paths) but the rule configuration is fixed and auditable.

The deep structure persists. The intent, content, audience, and purpose can be stored, versioned, approved. Legal signs off on the deep structure and transformation rules, not on every possible surface rendering.

Specific surfaces can be stamped as canonical. When you need an immutable mobile (the investor deck, the board presentation, the compliance training record) you generate once and stamp that instance as canonical. It becomes the traveling artifact. Other instances remain ephemeral.

This separates coordination durability from output rigidity. The things that need to stay stable for coordination (rules, intent, canonical instances) stay stable. The things that benefit from adaptation (surfaces for specific viewers, contexts, devices) can vary.

The pattern layer is where institutional knowledge accumulates. An organization's "house style" becomes transformation rules. Their pedagogical approach becomes learning mode configuration. Their narrative conventions become pitch mode weights. The knowledge is durable; the outputs are contextual.

The Allographic Turn⚓︎

The philosopher Nelson Goodman distinguished two kinds of art. Autographic works require the artist's hand: a painting is the physical object the painter made; a forgery, however perfect, isn't the work. Allographic works have notations that can be instantiated: a musical score isn't the music; performances are. The score travels; expressions vary. What matters is fidelity to the notation, not identity with an original surface.

Presentations have been autographic. The deck is the work. You arrange the slides, place the text boxes, choose the images. The artifact that results is what you made, and copies are copies of that specific arrangement. This is why "deck building" feels like craft, why people develop signature styles, why templates never quite fit. The surface is the work.

The shift to deep structure operation is an allographic turn. The work becomes the notation: intent, content, audience, purpose, encoded in a form that transformation rules can interpret. Surfaces are performances of that notation. Different contexts, different performances. The score travels; the renderings vary.

This pattern recurs. Assembly language programmers worked at surface, manipulating registers and memory addresses directly. High-level languages introduced an allographic layer: you specify what you want; the compiler generates the machine code. HTML authors once hand-crafted every tag; frameworks now generate markup from component specifications. In each case, what was craft becomes substrate. The level of operation rises; the previous surface becomes implementation detail.

The constraints that seem strategic are actually structural. PowerPoint can't easily shift to allographic operation because its entire architecture assumes the slide is the work. The file format encodes surfaces. The interface is a surface manipulation tool. Thirty years of muscle memory expects direct manipulation. These aren't business decisions that could be reversed; they're ontological commitments baked into the product's conception of what a presentation is.

Gamma changed the format primitive but kept the autographic assumption. Cards are still surfaces you edit. The constraint Gamma preserved is the constraint that matters most: the user works on the output. A deeper move would abandon that assumption entirely. Specify intent; let transformation rules handle expression. The presentation becomes score, not painting.

What Emerges⚓︎

When presentation capability operates at deep structure, forms become possible that couldn't exist when every surface required manual creation.

Responsive communication. A sales call where the system generates competitive analysis slides as objections surface. The rep mentions a competitor; a comparison matrix appears, drawn from the deep structure of product positioning, rendered according to transformation rules for competitive framing. The prospect asks about pricing; a TCO visualization materializes. The surfaces are ephemeral, created for this conversation, dissolving after. No one built these slides. They were performed from notation when the context demanded them.

Adaptive learning. Training materials that notice a cohort struggling with a concept and generate additional visual explanations mid-course. Not a single learning path but a space of paths, each learner traversing differently depending on what the transformation rules infer from their responses. The deep structure (what must be learned, in what sequence, to what standard) stays fixed. The surfaces vary per learner. Assessment becomes continuous; the system knows what you understood because it watched which generated explanations you needed.

Analytical multiplicity. An analysis that produces small multiples: the same data visualized dozens of ways simultaneously, so you can find the view that reveals the pattern. Not an analyst choosing one chart but the system generating the space of valid visualizations, constrained by Tufte's rules. The insight isn't in any single rendering; it's in the comparison across renderings. This is only possible when generation is cheap enough to be speculative.

Conversational evidence. A board discussion where relevant data visualizations appear as the conversation unfolds. Someone asks about regional performance; a map materializes. Someone questions the trend; a time series with confidence intervals appears. The presentation isn't a deck someone prepared; it's a capability that renders evidence on demand, shaped by transformation rules for executive communication. The board member who asks sharp questions gets sharper visuals.

The trajectory extends further than current interfaces suggest:

Near-term: Better surfaces faster. This is where Gamma and Copilot compete. AI generates initial outputs; humans refine. The surface remains the work; the tools just speed up its creation.

Medium-term: Context-aware generation. The system knows who's viewing and what they know. A sales deck renders differently for a technical buyer than for a procurement officer. Same deep structure, different transformation rule weightings. The author specifies intent once; the system handles audience adaptation.

Long-term: Communication that doesn't exist until needed and dissolves after use. No deck to store, version, email. A capability that produces the right surfaces when context demands them, then lets them go. The archive changes. What do you version-control when surfaces are ephemeral? The deep structure. The transformation rules. The canonical instances stamped for coordination. But most outputs are never saved because they were only ever performances.

What changes about expertise. When surfaces generate cheaply, the bottleneck shifts. Knowing how to arrange rectangles on a canvas becomes less valuable. Knowing what to communicate, and why, and to whom, becomes more valuable. The skill isn't "design" in the sense of visual arrangement; it's "design" in the sense of intent specification. The person who can articulate "help the board understand why customer acquisition cost increased despite higher marketing spend, so they can decide whether to continue the current strategy" produces better outputs than the person who says "make a presentation about Q3 results." Thick intent yields thick output.

What changes about coordination. Organizations coordinate around artifacts. When artifacts are ephemeral, coordination must anchor elsewhere: to deep structures that persist, to transformation rules that are versioned and approved, to canonical instances stamped for situations that require immutable mobiles. The shift is from coordinating around outputs to coordinating around capabilities. "Use L2 onboarding mode" replaces "send them this deck." The capability is the coordination mechanism.

What changes about authorship. If the surface is a performance of notation, who authored it? The person who specified the deep structure? The team that curated the transformation rules? The system that composed them in context? Authorship distributes. The question "who made this presentation" becomes as complicated as "who made this musical performance"—composer, arranger, conductor, orchestra, venue acoustics all contributed. The allographic turn doesn't eliminate authorship; it complicates it, spreads it across the layers that contribute to the final rendering.

The Residue of Craft⚓︎

Does this eliminate the need for design skill? Craft relocates rather than disappears.

Someone must know what makes visual communication effective. The transformation rules don't write themselves. Tufte's principles exist because someone spent decades analyzing what works and why. Pedagogical frameworks exist because researchers studied how understanding develops. These insights become inputs to the system: transformation rules that shape generation.

Someone must curate the pattern layer. Which principles apply in this domain? How should learning mode differ from pitch mode for this organization? What makes communication effective for this audience? The answers require judgment that can't be automated because it depends on context the system doesn't have.

Someone must specify intent clearly. Deep structure isn't magic. "Make a presentation about Q3 results" is thin; it yields generic output. "Help the board understand why customer acquisition cost increased despite higher marketing spend, so they can decide whether to continue the current strategy" is thick; it yields focused output. The skill of specifying intent precisely, knowing what you actually want to communicate, remains valuable.

The craft of arranging rectangles on a canvas may become less valuable. The craft of knowing what to communicate, why, and how: that persists and may become more valuable when surface-level work no longer differentiates.

But there's a harder question. The philosopher Michael Polanyi observed that we know more than we can tell. The expert presenter who senses the room losing attention and pivots mid-slide—that's tacit knowledge. The designer who tries three layouts before one "feels right" is exercising judgment that resists articulation. Transformation rules encode explicit knowledge: principles someone could write down, patterns someone could name. The tacit dimension doesn't transfer so easily. It manifests in practice but can't be fully captured in rules.

Some craft may be irreducibly embodied. The question isn't whether all craft disappears but whether the irreducible part matters. If tacit judgment improves outputs at the margins, making good presentations slightly better, then the loss is tolerable, perhaps unnoticeable. If tacit judgment is what distinguishes communication that transforms from communication that merely informs, the loss is substantial. Honest answer: we don't know yet. The boundary between codifiable and tacit shifts as systems improve. What seemed irreducibly human yesterday becomes pattern-matchable today.

There's a related concern. The anthropologist Lucy Suchman studied how people actually use technology and found that plans are resources for action, not determinants of it. People figure out what they mean through the act of making. The presenter who struggles with a slide discovers, in the struggling, what they actually wanted to say. The designer who drags a text box and immediately sees it's wrong learns something about the communication that wasn't accessible before the attempt.

If intent specification happens upstream, separated from the struggle with surfaces, where does that discovery occur? The system generates; the user accepts or rejects. But the cycle of attempt, failure, revision is how thick intent develops. Deep structure operation might need to preserve productive friction rather than eliminate it entirely. Not intent specification followed by surface generation, but something more iterative: rough intent, generated surface, refined intent, regenerated surface, each pass sharpening what the user actually meant. The interface isn't just for expressing intent. It's for discovering it.

Simon Wardley maps value chains by evolution: activities move from genesis (novel, uncertain) through custom-built and product to commodity (standardized, interchangeable). As activities commoditize, value migrates up the stack to whatever remains scarce. Surface generation is commoditizing fast. Every AI lab ships it. The substrate becomes utility. What stays scarce is what sits above: the transformation rules that encode effective communication, the pattern curation that fits those rules to organizational context, the intent specification that distinguishes thick from thin requests. Defensible value concentrates where judgment remains. The moat isn't in the generation; it's in the layers that shape what gets generated.

Thickness Through Use⚓︎

In a previous essay, I explored how AI primitives should thicken through practice: starting thin, accumulating context, becoming more fitted to their use. The concept applies here.

The transformation layer can learn. Which patterns work for this organization? Which rule weightings produce outputs that users don't override? What intent expressions recur, and what do they reveal about communication needs?

This isn't explicit configuration. It's observation. The system notices that learning outputs in this organization are consistently modified to add more visual examples, and adjusts the transformation rules to weight visual representation higher. It notices that pitch decks here always get restructured to lead with customer testimonials, and adjusts narrative rules accordingly.

The pattern layer thickens. House style emerges from use rather than specification. The capability becomes increasingly fitted to the organization, not through configuration but through accumulation.

This is also what saves the system from the legibility trap. High modernist planning fails when universal rules meet local conditions they weren't designed for. Thickness through use inverts that relationship. The system doesn't arrive with fixed patterns and impose them; it starts with general principles and lets local practice reshape them. The mētis—the contextual knowledge Scott argued couldn't be captured in plans—enters through accumulated observation rather than explicit encoding. What the designer couldn't specify, the system learns. The transformation rules become fitted to place rather than imposed on it.

This is where the moat forms. A generic presentation tool can be copied. A transformation layer that has learned an organization's communication patterns over years of observation cannot easily be replicated. The thickness is the defensibility.

The Bet⚓︎

The complications are real. Intent discovery through making, tacit knowledge that resists codification, the blindness of systems to what falls outside their categories. None of these concerns are dismissable. But they don't change the direction; they change what you build along the way.

The system needs iterative loops, not one-shot generation. Rough intent in, surface out, user surprise, refined intent, regenerated surface. That's not a compromise; it's the actual interface for deep structure operation. The clean separation was always a simplification for exposition.

The tacit ceiling exists, but it keeps rising. What seemed irreducibly human in 2020 became pattern-matchable by 2024. The bet isn't that tacit knowledge disappears. The bet is that the codifiable part grows faster than most people expect, and that's where the leverage concentrates.

The legibility trap is real, but thickness through use is the response. Not universal patterns imposed from above; local patterns learned through observation. The system that's been watching how this organization communicates for three years has absorbed mētis that no designer could specify. That's not perfect. The unobserved still falls outside. But it's better than static rules, and it compounds.

Surface-level tools will persist. PowerPoint isn't going anywhere; neither is Gamma. People will still arrange rectangles on canvases, and some of them will do it brilliantly. But the volume of visual communication is exploding while the number of people who can craft surfaces well stays roughly constant. The gap gets filled by generation from deep structure, whether anyone plans it that way or not.

The presentation as stable artifact category is passing. What replaces it is communication capability: the same deep structure producing slides when you need slides, documents when you need documents, interactive experiences when you need those, and forms we haven't named yet when context demands them. The artifact crystallizes from the capability when coordination requires it. Otherwise, surfaces stay ephemeral.

This is not a prediction about what will happen. It's a statement about what to build. Operate at deep structure. Encode transformation rules. Make the patterns visible and contestable. Let thickness accumulate through use. The organizations that do this will communicate more effectively than those still crafting surfaces by hand. That's the bet.

Sources⚓︎

  • Alexander, Christopher. A Pattern Language (Oxford University Press, 1977)
  • Chomsky, Noam. Syntactic Structures (Mouton, 1957)
  • Flusser, Vilém. Towards a Philosophy of Photography (Reaktion Books, 1983)
  • Goodman, Nelson. Languages of Art (Hackett, 1968)
  • Latour, Bruno. "Visualisation and Cognition: Drawing Things Together" (1986)
  • Polanyi, Michael. The Tacit Dimension (University of Chicago Press, 1966)
  • Scott, James C. Seeing Like a State (Yale University Press, 1998)
  • Suchman, Lucy. Plans and Situated Actions (Cambridge University Press, 1987)
  • Tufte, Edward. The Visual Display of Quantitative Information (Graphics Press, 1983)
  • Gamma founding story and metrics from company sources and industry analysis