Two Experiences of the Same Intelligence

I recently found myself moving between two very different experiences of artificial intelligence. In one, I was having an extended conversation about ideas. We moved across philosophy, technology, work, and personal experience. An intuition that had begun as a vague feeling acquired language and structure. New connections appeared, not because the machine delivered a final answer, but because the exchange gave me another mind against which to develop my own thinking.

The conversation felt like a pursuit of knowledge and wisdom. It could take place through typed text or voice, and the difference hardly mattered. The important element was the continuity of thought. I could introduce an observation, reconsider it, challenge the response, and carry the idea further. AI was not replacing my intellectual activity. It was giving that activity greater range, speed, and expressive power.

The other experience began when I asked AI to make small corrections to Excel and PDF files. The intellectual requirements were modest. A title needed to be adjusted. Some formatting had to remain unchanged. A minor visual problem had to be repaired. These were actions I could have completed manually within minutes, yet the process became a succession of instructions, revisions, inspections, and new corrections. One problem was solved and another appeared. A file that looked correct in one preview behaved differently when opened in its original application.

The contrast was hard to ignore. The same AI that could discuss the nature of intelligence with fluency had difficulty changing one element of a document without affecting something nearby. Its knowledge seemed broad enough to contain whole libraries, while its practical control over a familiar office file could feel clumsy. In one setting, artificial general intelligence appeared close. In the other, the machine still seemed unable to perform a basic clerical task reliably.

Both impressions contain truth. They do not describe two different technologies, but two regions of an uneven capability landscape. AI can be remarkably strong when it works directly with language and meaning. Its reliability falls when understanding must be converted into exact action inside an external environment. Recognizing that division changes how we interpret both the intelligence of the system and our frustration with it.

When a Digital Environment Becomes Physical

The familiar comparison is Moravec’s paradox. Tasks that humans regard as intellectually demanding, such as calculation or formal reasoning, became accessible to computers earlier than ordinary abilities such as recognizing objects, walking through a room, or folding clothes. What appears easy to us is supported by layers of perception and physical coordination acquired through years of living in a body. We rarely notice the complexity because the skill has become automatic.

Folding a shirt requires more than knowing what a folded shirt should look like. A person must recognize its orientation, distinguish one part from another, feel how the material responds, keep track of overlapping surfaces, adjust pressure, and correct each movement as the shape changes. The action is guided by a continuous loop between perception and movement. There is no clean separation between understanding the task and sensing whether it is going well.

An Excel workbook or PowerPoint presentation is not a physical object, but operating it creates a similar problem. The AI must perceive the current state, identify the correct element, select an action through an available tool, preserve everything that should not change, and inspect the result. Its environment is digital, yet the task requires a form of situated competence. The system needs more than knowledge about spreadsheets or slides. It needs dependable control within them.

This becomes clearer when we compare an application file with Markdown. In a plain-text document, the words visible to the reader are close to the underlying artifact. A heading is marked as a heading. A paragraph exists as a paragraph. When a sentence changes, the modification can be located and compared directly. The medium in which the AI interprets the request is also the medium in which it performs the work.

Office documents contain additional layers. What appears as a line of text may depend on a theme, a style hierarchy, an embedded font, a text box, a slide master, coordinates, grouping, and the rendering rules of a particular application. A spreadsheet cell may contain a value, a formula, a number format, conditional formatting, data validation, comments, borders, and partial text styling. A person looking at the screen sees one object. The software contains a collection of related states that must remain consistent.

Modern Office formats are not necessarily opaque binary blocks. Many are packages of XML files and other assets. That technical detail does not make them transparent to either the user or the AI. They remain complex document models whose final appearance depends on software interpretation. A tool can modify the internal structure correctly according to one library and still produce an unexpected result in Excel, Word, or PowerPoint.

For the AI, tools function as a temporary digital body. The quality of its action depends on what that body can perceive, what it can manipulate, and what feedback it receives afterward. A screenshot provides only a momentary view. A software library may expose document properties without reproducing the application’s rendering. A conversion tool may preserve content while changing layout. Understanding the instruction is only the beginning.

The deeper division, then, is not between the digital and physical worlds. It is between symbolic activity and action within an environment. Language is close to the AI’s native habitat. An application interface is an external space that must be observed and controlled. In that space, moving a text box can become the digital equivalent of folding a shirt.

Vastness Is Easy, but One Cell Is Hard

The paradox becomes sharper because AI is not uniformly weak with complex files. Give it a large spreadsheet and it may examine thousands of rows, identify inconsistent entries, compare formulas, and notice patterns that would exhaust a human reviewer. It can read a long report, summarize its structure, find contradictions across distant sections, or compare a collection of presentations within a short period. In scale and speed, the advantage can be extraordinary.

Yet the same system may struggle to change one cell without altering its style. It can understand a two-hundred-page report but mishandle a page break. It can evaluate the narrative of an entire slide deck but allow one title to extend beyond its frame. Vastness is easy, while locality is hard.

This scale–precision paradox arises because analysis and intervention demand different capabilities. Large-scale reading asks the AI to detect patterns in information it has been given. Local editing asks it to change a particular state while guaranteeing that all surrounding states remain untouched. The first task rewards breadth. The second requires control, restraint, and exact verification.

Intellectual conversation also has room for productive variation. There may be several good interpretations of an experience or several ways to develop an argument. An unexpected association can improve the exchange. The user can respond to a partially formed idea and redirect it. Meaning develops through movement rather than through a single predetermined output.

Application editing often allows no such freedom. If the request is to change an 11-point title to 12 points without modifying anything else, the acceptable result is narrow. A creative variation is not a contribution. It is a mistake. The qualities that make a language model stimulating in conversation can become liabilities when exact preservation is required.

Human beings also bring a powerful feedback system to small visual tasks. When I adjust a spreadsheet manually, my eyes remain on the screen while my hand performs the action. I see immediately if the row height changes or the text no longer fits. I can reverse the action before it becomes part of a longer chain of errors. The correction is so fast that I may not experience it as a separate reasoning step.

AI commonly works through discontinuous cycles. It inspects, acts, saves, renders, and inspects again. Each cycle may use a different representation of the file. The system does not always maintain the continuous visual awareness that makes human manipulation feel effortless. A minor correction therefore becomes a sequence of uncertain translations.

This difference helps explain why public judgments of AI vary so widely. A person who uses it for research, writing, interpretation, and intellectual dialogue encounters its highest capabilities. Another person who relies on it to update slides, complete forms, organize application windows, or preserve complicated formatting repeatedly encounters its weaknesses. One sees the approach of general intelligence. The other sees a system that cannot be trusted with routine work.

Neither judgment is invented. Each is based on a real part of the technology. Trouble begins when either experience is treated as the whole. Fluency in thought does not guarantee reliability in action, while poor control of an interface does not erase the system’s ability to extend human understanding. Intelligence is not a single quantity that rises evenly across every kind of task.

The High Cost of Trivial Work

The unevenness of AI capability has an economic consequence. A sustained philosophical conversation may consume many words, but the path from input to value is direct. Language enters the system, reasoning takes place through language, and language returns to the user. A single response can clarify an intuition, introduce a useful distinction, or alter the direction of an essay. The intellectual value per operation can be high.

A small PowerPoint correction may involve far more machinery. The system has to open or parse the presentation, identify the relevant object, modify its properties, save a new file, render the result, and examine the image. If the title has shifted or another element has changed, the process begins again. The task is trivial in human terms, but computationally it becomes a chain of tool calls, representations, and verification steps.

The user bears another cost that token accounting does not capture. Each round requires attention. The file must be opened, the result inspected, and the next correction explained. What would have been one direct manual gesture becomes a linguistic negotiation about a visual state. Even if AI saves a few seconds of manipulation, it may create several minutes of supervision.

This is the token-value paradox. Some of the least intellectually demanding tasks can become expensive because they require repeated interaction with an environment. Meanwhile, an exchange that contributes to years of accumulated thought may be computationally straightforward because it remains within text. Token quantity alone tells us little about the human value being produced.

Many corporate AI programs have encountered this distinction without naming it. A company purchases access to a flagship model and encourages employees to use it widely. Usage becomes evidence of adoption, so people look for opportunities to insert AI into their existing routines. They ask it to answer emails, modify spreadsheets, prepare slides, summarize meetings, fill in forms, and interact with the applications already governing their work.

These uses are understandable. Employees begin with the tasks that occupy their day. Yet many of those tasks are bound to human-centered interfaces and inherited procedures. AI does not enter a clean field of knowledge. It enters a collection of applications, templates, approval chains, and formatting conventions. The organization may be paying for advanced reasoning while directing much of it toward digital clerical work.

Costs then rise faster than expected. Limits are introduced, access is narrowed, or employees are told to use tokens more carefully. Those who experienced AI mainly as a convenient addition to familiar applications often return to their former methods. They can write the short email themselves or adjust the slide more quickly by hand. Once unrestricted use disappears, little has changed beneath the temporary layer of automation.

The comic element conceals a strategic error. The company believed that maximizing AI usage would make it AI-native. In practice, it encouraged employees to spend expensive intelligence on every available activity without distinguishing intellectual leverage from interface overhead. The result was maximum consumption rather than maximum value.

AI-Generated Bureaucracy

Corporate knowledge already passes through too many forms. An observation becomes a document. The document becomes a slide deck. The deck is presented in an online meeting. The meeting becomes a transcript, which becomes minutes, which becomes a summary. That summary may later be copied into a report or transformed into another presentation for a different group.

Across these transformations, the quantity of material grows while the underlying knowledge may remain unchanged. A one-page idea becomes twenty slides, an hour of discussion, ten pages of transcript, and a concise set of AI-generated notes. Each artifact can be competently produced. None guarantees that the organization understands more, decides better, or acts sooner.

This is output expansion without epistemic gain. AI makes the pattern easier to sustain because it lowers the effort required at every stage. A person no longer needs an afternoon to turn prose into slides. A meeting no longer needs a designated note-taker. A long transcript can be reduced to action points in seconds. The reduced cost of each artifact appears to promise efficiency.

The total effect may run in the opposite direction. When reports and presentations become cheaper, organizations can request more of them. When meeting notes are automatic, there is less pressure to question the number of meetings. The cost per item falls while the volume of items rises. This resembles a Jevons-style rebound effect applied to office work: greater efficiency in producing corporate material leads to greater consumption of corporate material.

AI can therefore accelerate bureaucracy as readily as it can accelerate thought. It does not decide on its own that a recurring report has lost its purpose or that three approval layers are unnecessary. If asked to support an inherited process, it can make that process faster, larger, and more polished. Inefficiency becomes less visible because automation reduces the discomfort that once drew attention to it.

Presentation carries a special risk because polish can imitate substance. A thin idea placed inside a professional template acquires visual authority. Large headings, restrained colors, spacious layouts, polished icons, and consulting-style phrases suggest that careful strategic work has taken place. Even minimalist design can become ornamental when it devotes most of the slide to creating an impression rather than aiding understanding.

Before generative AI, such polish required time and a degree of skill. It served as an imperfect signal of effort. Now these materials can be produced with far less effort. As polished output becomes abundant, its value as evidence of serious thinking declines. The presentation may look more authoritative at the same moment that its appearance tells us less about the quality of the knowledge beneath it.

The resulting inflation is not confined to slides. AI can produce executive summaries, key takeaways, talking points, follow-up emails, and formal minutes from the same small body of information. Each version gives the organization another object to circulate and review. People become occupied by the movement of representations rather than the development of ideas or the making of decisions.

Meeting summaries illustrate the distinction clearly. Automated notes can be valuable when a necessary discussion involves real disagreement, coordination, or commitment. They preserve decisions and reduce the burden on participants. But accurate documentation cannot rescue a meeting that had no reason to occur. A perfect summary of an unnecessary meeting remains part of the waste.

AI adoption can expose this condition if organizations are willing to look. Rising token costs reveal how often the same information is being expanded, reformatted, summarized, and circulated. What appears on a budget line as excessive AI consumption may also be evidence of excessive corporate ritual. The technology is not creating every inefficiency. It is making the existing structure easier to see.

Markdown and the Discipline of Proportion

Plain text offers a different starting point. Markdown does not eliminate weak thinking, and a long Markdown document can be as empty as a long presentation. It does, however, remove many of the devices that allow presentation to distract from content. The argument must rely on its language, evidence, sequence, and internal coherence.

This creates a discipline of proportion. A small idea can remain a small artifact. It does not need to be enlarged into a deck to appear legitimate. A clear paragraph may be sufficient for an observation. A page may be enough for a proposal. When more space is required, it can be earned by the development of the thought rather than by the demands of a template.

Markdown is also well suited to cooperation between humans and AI. Its structure is explicit without becoming heavy. Headings, lists, quotations, links, and emphasis remain visible in the source. Changes can be compared at the level of words and lines. The user and the model work on nearly the same representation, so there are fewer hidden states between intention and result.

This compatibility supports a text-first, artifact-last approach. Research, brainstorming, drafting, review, translation, and revision can take place in plain text. That file remains the source of knowledge. A PDF, Word document, web page, or presentation can be produced when the material reaches an audience that needs a particular format.

Under this model, rich documents become compiled outputs rather than intellectual originals. The distinction resembles the relationship between source code and a finished application. People do not normally treat the rendered screen as the only authoritative form of a software system. They preserve the structured source from which the visible product can be created again. Knowledge work could adopt a comparable habit.

The approach would also make revision more coherent. Instead of correcting several slightly different presentations of the same material, the author updates the source and regenerates the necessary forms. AI can contribute where it is strongest by examining the argument, detecting inconsistency, proposing structure, and adapting the material for different audiences. Deterministic tools and templates can handle repeatable conversions.

None of this makes design unnecessary. A chart can reveal a relationship that paragraphs cannot show efficiently. A diagram can clarify a system. Slides can control pacing, direct collective attention, and support a live speaker. A formal document can satisfy legal, accessibility, or institutional requirements. Visual form deserves effort when it changes what an audience can understand, remember, decide, or do.

The relevant question is whether the presentation layer performs such a function. What becomes possible through this format that would not be possible through a page of clear text? If the answer concerns comprehension, evidence, accessibility, or coordinated action, richer presentation may be justified. If the answer concerns convention, status, or the desire to make limited content appear larger, the format is serving the institution more than the knowledge.

Markdown cannot reform corporate culture by itself. It can still provide a useful test. Strip away the template and the branded language, and see what remains. If the proposal loses its force when expressed plainly, the weakness may lie in the idea rather than the medium. If it becomes clearer, the presentation was adding distance instead of meaning.

Intelligence Before Usage

An AI-native organization should not be defined by the number of seats it purchases or the volume of tokens its employees consume. Those measures describe access and activity. They do not reveal whether intelligence is being used well. Heavy usage can coexist with shallow adoption when AI is attached to every existing task without examining the purpose of the task itself.

A more mature organization would distinguish among several kinds of work. AI can provide intellectual leverage when it compares evidence, finds inconsistencies, tests assumptions, develops interpretations, or helps people express what they have not yet managed to articulate. Conventional software remains better for deterministic transformations. Humans often retain the advantage in small visual corrections because direct perception and action make them faster and more reliable.

Judgment lies in choosing the right combination. Asking AI to detect anomalies across a large workbook may be an excellent use of its scale. Asking it to correct one border repeatedly may not be. Developing the argument for a presentation through dialogue can create real value. Delegating every final alignment adjustment may consume more attention than it saves. Generating notes from a consequential meeting can preserve commitments. Generating notes from every routine call can add another unread document to the system.

This approach does not reduce AI to a specialized tool. It respects the breadth of its intelligence while acknowledging the conditions under which that intelligence can operate reliably. It also respects human ability. The objective is not to remove people from every action, but to strengthen thought and reduce work that serves no clear purpose.

My own experience makes the distinction tangible. In text and voice conversation, AI can accompany an inquiry as it develops. It can bring distant ideas into relation, give form to an intuition, and return a question in a way that changes how I see it. The exchange augments an activity I already value. It does not ask me to abandon my intellectual habits. It extends them.

The experience becomes stressful when the same intelligence must pass through layers of file structure, rendering, and application behavior to complete a minor operation. That frustration should not lead us to dismiss AI as unintelligent. Nor should the power of its language tempt us to assume that it can act reliably everywhere. The contrast reveals that knowledge, judgment, execution, and environmental control remain different capabilities.

It also leaves companies with a more serious question than how to reduce their AI bills. Instead of asking how AI can produce all the documents, slides, summaries, and meeting records they already generate, they can ask why so many of those artifacts exist. Some will prove necessary. Others will be recognized as the residue of habits formed when information was expensive to create, difficult to distribute, or useful as a display of organizational effort.

The next phase of AI adoption may be defined by restraint rather than abundance. Plain text can remain the home of developing knowledge. Rich formats can be reserved for moments when form carries meaning. Meetings can be held when presence changes the outcome. AI can be directed toward inquiry, analysis, and articulation instead of being consumed by every available interface.

Artificial intelligence has become remarkably capable of participating in the world of meaning while remaining unreliable in the world of particulars. That unevenness will not disappear through enthusiasm or token limits alone. It asks for a culture that knows the difference between producing material and producing knowledge. An organization becomes AI-native when it learns where intelligence creates value, where tools create friction, and where the wisest action is to stop manufacturing another artifact.

Photo by ZHENYU LUO on Unsplash