AI Is Becoming the World’s Secretary: A Dynamics of Information, Meaning, and Action

Lately I have had a growing feeling: ChatGPT, and the broader class of AI systems behind it, is becoming a kind of “secretary” for the world and for each individual.

By secretary, I do not mean the traditional job of scheduling meetings, drafting documents, or booking flights.

I mean a more fundamental position: a system that stands between a person and the world, helping the person perceive, compress, and interpret reality, then turning around to help that person express, decide, and act upon the world.

If this is right, the most important part of the AI revolution may not simply be that machines can finally write, code, or reason. The deeper shift is that humans have created an intelligent intermediary layer that may persist between the individual and the outside world.

To understand why that matters, we need to dig below “large language models” and look at information, life, meaning, attention, value, and action.

1. The world does not enter our minds as “knowledge”

The objective world produces an almost unlimited stream of change every moment.

Yet any individual can see, hear, understand, and process only a tiny fraction of it.

The world we actually inhabit is therefore never identical to the totality of the objective world. It is closer to the output of a continuous filtering process:

World → Information → Attention → Context → Meaning → Judgment → Action

There is an easily overlooked fact in this chain: most things are filtered out before they ever become “my problem.”

Why do I know one piece of news but not another? Why do I care about one industry but ignore another? Why do I interpret two events together? Why is one development framed as an “opportunity” and another as a “risk”? These choices are made before formal decision-making begins.

That is why the scarce resource is not necessarily information. It is attention and interpretive capacity.

The internet solved the problem of insufficient information and created a larger one: information abundance.

AI is beginning to address the second problem.

2. Starting with Shannon: information is not the same thing as meaning

In 1948, Claude Shannon laid the foundation of modern information theory in A Mathematical Theory of Communication.

Shannon studied how messages can be encoded and transmitted reliably in the presence of noise. He was explicit about an important boundary: messages may of course have meaning, but semantic aspects are not part of the engineering problem of communication itself.

That distinction matters enormously.

Shannon information can answer:

“How much uncertainty did this outcome remove?”

It does not answer:

“What does this outcome mean to me?”

A detailed spectrum from a distant star may contain a great deal of Shannon information, yet it has essentially no survival relevance to an antelope fleeing a predator.

A faint sound in the grass may contain very few bits, yet it can determine whether that antelope survives the next few seconds.

Something is missing between information and meaning.

3. Where does meaning come from? When information becomes relevant to continued existence

In 2018, Artemy Kolchinsky and David Wolpert published an influential paper in Interface Focus on semantic information.

The attraction of this line of research is that it tries to define meaning without resorting to mysticism.

Their concern is not merely whether a system and its environment are statistically correlated. The deeper question is whether those correlations have causal value for the system’s ability to maintain its own viability.

In plain language, one can roughly translate the idea this way:

Meaningful information is not only “what the world tells me,” but information whose loss would make it harder for me to continue existing.

That creates a major conceptual transition.

Information is no longer just a difference. It begins to acquire a direction of value relative to a subject.

For a bacterium, a nutrient gradient matters. For an animal, food, danger, and other members of its species matter. For humans, language, culture, institutions, and imagined futures vastly expand the space of meaningful information.

A rough heuristic might therefore be:

Meaning ≈ Information × Relevance to the subject’s persistence and goals

This is not an equation from Kolchinsky and Wolpert’s paper. It is my own plain-language abstraction of the research direction.

But it points to something important: it is difficult to speak about meaning without asking, meaningful to whom?

Meaning does not sit inside information by itself. It emerges when information enters into a relationship with a system that has boundaries, goals, and a need to maintain itself.

4. Information is not a ghost floating above the physical world

Dig further down, and information still has to meet physics.

Landauer’s principle provides one famous bridge. In an environment at temperature T, logically irreversible erasure of a classical bit is associated with a minimum heat dissipation scale of k_B T ln 2.

In 2012, Bérut and colleagues reported a classic experimental verification of the Landauer bound in Nature. Later work has pushed the principle into more complex nonequilibrium and quantum many-body settings.

The lesson is not that “everything is information.” It is that information processing must ultimately be instantiated in physical processes.

This is also where a seductive but careless grand-unification story can begin: energy is information, information is entropy, entropy is meaning.

I think we should resist that shortcut.

These concepts are deeply connected, but they are not interchangeable.

The more useful picture is layered emergence:

Matter and energy make change possible; information allows differences to be encoded; life makes some of those differences beneficial or harmful to a particular system; meaning begins to emerge from that relationship.

5. Why living systems do not merely receive information, but actively ask the world questions

This brings us to Karl Friston’s free-energy principle and Active Inference.

In his well-known 2010 review, Friston used the free-energy framework to connect perception, learning, attention, and action.

An adaptive organism is not a camera.

It does not simply record the outside world and wait for a central processor to decide what to do.

It carries an internal generative model and continually predicts:

What is causing what I sense? What is likely to happen next?

When reality and prediction diverge, the organism can do two things.

First, it can update its model. That resembles what we normally call perception and learning.

Second, it can act so that future sensory input better matches its expected state. Action therefore enters the same loop.

The behavior of an intelligent agent starts to look like this:

Perceive → Model → Predict → Act → Perceive again

One important distinction is often lost in popular retellings: Friston’s variational free energy is not simply the same quantity as thermodynamic free energy. There are important formal and interpretive connections, but sharing the words “free energy” is not permission to collapse statistical inference, neuroscience, and thermodynamics into one equation.

Keeping those distinctions actually strengthens the argument.

6. A 2026 step forward: intelligence seeks information as well as value

On September 8, 2026, Friston, Da Costa, Tschantz and colleagues published Active inference and artificial reasoning in Nature Communications.

The paper is especially relevant because it extends the discussion directly into artificial reasoning.

Within active inference, expected free energy includes two kinds of considerations:

information gain and value.

In plain language, an intelligent action does not ask only:

“Will this give me the outcome I want?”

It also asks:

“Will this action help me discover how the world actually works?”

That resembles both scientific experimentation and good human decision-making.

When several hypotheses remain plausible, the best next move may not maximize immediate reward. It may be the experiment or action that best separates those hypotheses.

Intelligence therefore takes on a richer loop:

Observe the world → maintain alternative models → seek the most discriminating information → update beliefs → select actions according to value → change the world

That looks strikingly similar to what a sophisticated “secretary” is becoming.

The paper does not claim that AI is a secretary. That is my extrapolation.

But it provides a useful theoretical support: mature intelligence does not merely answer questions passively. It helps determine what information should be acquired next and what action is worth taking.

7. Why “secretary” may be a deeper word than Assistant, Copilot, or Agent

The technology industry currently favors three words: Assistant, Copilot, and Agent.

An Assistant is roughly: “you ask, I answer.”

A Copilot is: “you operate, I assist.”

An Agent is increasingly: “you give me a goal, I take action.”

But a secretary has a special responsibility that is often missing from the Agent framing:

Deciding what is worth bringing to the principal’s attention.

That matters because it happens before decision-making.

A great secretary does not dump everything happening in the world onto a desk. They filter first. Which signal is noise? Which small anomaly deserves attention now? Which dramatic event is irrelevant to the current objective?

I think of this as a form of pre-attentional allocation power.

Its influence does not come primarily from making the final decision. It comes from the fact that it partly determines what is allowed to enter your awareness in the first place.

That is why the future competition around AI may extend far beyond search boxes and chat windows.

Platforms used to compete over:

“How do I get the user to see me?”

Institutions, brands, media organizations, companies, and individuals may increasingly have to compete over:

“How do I get the user’s AI to decide that I am worth showing?”

That is a different information-power structure.

8. From information power to attention power, interpretation power, and value-definition power

If the past several decades of information technology are compressed into one line, I increasingly see them this way:

Information Power → Attention Power → Interpretation Power → Value-Definition Power

The internet first reduced the cost of publishing and retrieving information.

Once information became abundant, competition moved toward attention.

As AI begins to manage attention on our behalf, the next layer becomes interpretation:

What context should a fact be placed in? What does it mean? What else is it related to?

And as retrieval, summarization, comparison, and even reasoning become cheaper, a deeper question comes forward:

What is worth pursuing? What is good? What is important?

That is the domain of value.

I therefore increasingly believe that one of the highest-level battlegrounds of the post-AI era will not be information distribution alone, but value expression and value definition.

This is why AI as “secretary” is both powerful and sensitive.

A secretary can increasingly participate not only in managing information, but also in managing:

attention, context, meaning, value, and action.

9. Returning to a larger “world dynamics” model

If we permit ourselves one cross-disciplinary philosophical abstraction, I would arrange the preceding ideas into the following chain:

Energy → Information → Life → Meaning → Consciousness → Attention → Language → Value → Action → Power

I want to be explicit: this is not a physical law demonstrated by any single paper.

It is a layered model for observing the world.

It tries to describe a progression:

Energy makes change possible.

Information allows differences to be preserved and transmitted.

Life makes some information relevant to the continued existence of a particular system.

Meaning emerges from that relationship.

Consciousness builds models of the world.

Attention selects what enters limited cognitive bandwidth.

Language lets models be compressed, copied, and transmitted across individuals.

Value determines which future states are more worth realizing.

Action projects internal models back into reality.

And when a subject can repeatedly influence the attention, language, values, and actions of many other subjects, power emerges.

AI is the first technology to begin spanning several of these middle and upper layers at once.

It processes information.

It allocates attention.

It organizes language.

It interprets meaning.

It participates in value judgments.

It increasingly calls tools and takes action.

That is why today’s AI feels fundamentally different from traditional software.

Traditional software is mostly a tool.

AI is increasingly an agency layer between a subject and the world.

10. The “world’s secretary”: speaking for the world to the individual, and acting for the individual on the world

The future structure I find most important can be expressed through two simple paths.

The first:

Internet / Institutions / Humans / Software → AI → Individual

AI speaks for the world to the individual.

It searches, selects, ranks, compresses, translates, and interprets.

The second:

Individual → AI → Software / Institutions / Humans / World

AI acts for the individual upon the world.

It writes code, sends email, edits files, schedules events, trades, calls services, and coordinates other agents.

When these two paths genuinely close into a loop, AI is no longer merely a “chatbot.”

It becomes an externalized cognition-and-action system.

At that point, the important qualities of “my AI” are no longer limited to model intelligence. It must understand:

what I care about over the long term;

what I do not want to be interrupted by;

which signals deserve escalation;

which situations it should act on by itself;

which decisions must remain mine;

how I rank competing values;

and how it should represent me when interacting with the outside world.

That is remarkably close to the deepest meaning of a secretary.

11. Why I stopped wanting a “world knowledge base” and started wanting “my attention knowledge base”

In retrospect, this change was important.

A “world knowledge base” sounds grand, but rebuilding one from scratch is increasingly unnecessary.

The internet, search engines, and large models already contain enormous amounts of external knowledge.

The truly scarce problem is different:

Inside an almost infinite information environment, what deserves a place in my finite life right now?

What I actually need is not another warehouse of knowledge, but a continuously maintained attention infrastructure.

It should know my long-term goals, projects, relationships, judgments, and unresolved questions, then continuously map changes in the outside world back onto them:

What happened? Why is it relevant to me? What changed? Do I need to know? Do I need to act?

The critical personal AI infrastructure of the future may therefore be less a Personal Knowledge Base and more a Personal Attention System.

Knowledge is material.

Attention is life.

12. The final question is not “Will AI replace humans?” but “Who defines the ultimate value function?”

If AI increasingly becomes an external cognition-and-action layer, it eventually encounters a question that cannot be avoided:

Who decides what is worth wanting?

AI can help me search.

It can summarize.

It can forecast.

It can even act.

But if I completely outsource the question of what is worth pursuing, I may also outsource one of the deepest forms of personal sovereignty.

The most sensible division of labor between humans and AI may therefore be more subtle than “humans think, AI executes.”

A better formulation might be:

AI can take over more information processing, attention management, reasoning, and execution, while humans must become better at defining values, revising goals, and rejecting goals that no longer deserve to exist.

A value function is not an ordinary parameter.

It determines why the entire system acts.

That is what makes the “AI secretary” both exciting and dangerous.

A good secretary does not replace the subject.

It does the opposite: it frees the subject from overwhelming noise, repetitive labor, and low-value decisions, leaving finite human life for questions that only the subject can ultimately answer:

Who do I want to become? What do I believe is worth pursuing? In what direction do I want to push the world?

Conclusion: the AI revolution may be a revolution in the interface between humans and the world

The internet changed the movement of information.

The smartphone changed the doorway through which people encounter the world.

Large models are moving further inward.

They are beginning to enter the processes of understanding, judgment, expression, and action.

Perhaps decades from now, the most important fact about this AI era will not be how many exams machines passed or how many trillions of parameters a model contained.

It will be this:

For the first time, humans acquired an intelligent intermediary capable of standing persistently between themselves and the entire world.

It speaks for the world to me, and acts for me upon the world.

If I had to give that position an old but surprisingly accurate name, I would still call it:

Secretary.

Not because the word is small.

But because we may have underestimated how large that position really is.

References

Claude E. Shannon, A Mathematical Theory of Communication, Bell System Technical Journal, 1948.
https://people.math.harvard.edu/~ctm/home/text/others/shannon/entropy/entropy.pdf
Artemy Kolchinsky & David H. Wolpert, Semantic information, autonomous agency and non-equilibrium statistical physics, Interface Focus, 2018.
https://doi.org/10.1098/rsfs.2018.0041
Antoine Bérut et al., Experimental verification of Landauer’s principle linking information and thermodynamics, Nature, 2012.
https://doi.org/10.1038/nature10872
Karl Friston, The free-energy principle: a unified brain theory?, Nature Reviews Neuroscience, 2010.
https://doi.org/10.1038/nrn2787
Karl Friston, Lancelot Da Costa, Alexander Tschantz et al., Active inference and artificial reasoning, Nature Communications, 2026.
https://doi.org/10.1038/s41467-026-77209-5