How Can Compression Expand Intelligence?

From A Brief History of Intelligence to world dynamics—and how OUTBIRD should become smarter

After listening to A Brief History of Intelligence, what stayed with me was not a particular brain region or evolutionary date, but four ideas: complexity, modeling, compression, and the expansion of capability.

Together, they pose an apparent contradiction. Compression discards information; greater capability lets a system do more. Why might leaving something out help a system understand more and act better?

This question brought me back to two ongoing concerns. The first is what I provisionally call “world dynamics”: how material conditions, an agent’s understanding, and the consequences of action shape one another. The second is more concrete: how can OUTBIRD, the system I am building, become genuinely smarter rather than simply larger?

This essay connects those concerns. It is not an established unified theory, and its system proposals are not claims about deployed features. They are working hypotheses that need to be tested against evidence.

1. Five breakthroughs change which relationships can be used

Max Bennett organizes an evolutionary account of human intelligence around steering, reinforcing, simulating, mentalizing, and speaking. His earlier paper calls the framework a “first approximation.” I use it as a thinking tool, not a five-level examination that every organism and AI must pass in sequence.[1]

In my own words: distinguishing benefit from harm; learning from the consequences of action; rehearsing events that have not happened; understanding one’s own and others’ mental states; and sharing experience and ideas through language. Mentalizing is not the same as sympathy, and language is not the same as writing.[1]

What interests me is that these are not merely improvements in speed on an unchanged task. Each makes a previously difficult relationship more usable: consequences over time, unrealized possibilities, another agent’s understanding, or ideas communicated between individuals.

This suggests a provisional definition of an expansion in intelligence: a new representation or form of organization makes previously difficult problems expressible, predictable, or actionable.

This is an expansion of capability, not an automatic increase in physical or mathematical dimensions—and not a claim that evolution must reach a predetermined destination.

2. Modeling preserves relationships; compression leaves out details

Consider a hypothetical navigation task.

I could record the color of every wall and the texture of every tile. Or I could remember: “Turn left at the end of the corridor; the second door is the exit.” The latter discards enormous detail but may be better for finding a way out. If the task changes to inspecting cracks in the walls, that representation is no longer adequate.

Before compressing, we therefore need to ask: for whom, for which task, and with which unacceptable errors? The information bottleneck method formalizes a related problem: retaining as much information relevant to a target variable as possible within a compact representation.[2]

I now see modeling and compression as two sides of one process. Modeling emphasizes which relationships are retained; compression emphasizes which distinctions are omitted. A model is not a miniature world. It is a choice about what deserves to be distinguished.

This also corrects my earlier tendency to emphasize attention too heavily. Attention does not simply select important items from a ready-made pile of information. A model already influences which distinctions can be recognized and which clues become noticeable. Goals and attention, in turn, influence which models are learned.

Rather than a one-way chain from information through attention to meaning, I now emphasize the interaction of models, attention, goals, and feedback. Calling attention “online compression” and memory “compression over time” can be useful, but these are functional analogies, not substitutes for investigating their mechanisms.

Good compression does not retain as little as possible. It omits irrelevant detail without erasing distinctions that would change a judgment.

3. Compression can expand capability without establishing causality

Imagine one system memorizing a hundred routes and another learning how roads connect. The second may preserve fewer details yet assemble a route it has never traveled.

This is what I mean by capability emerging through compression: moving from stored instances to reusable, composable structure. A more economical representation and a wider range of tractable problems are not contradictory.

But the idea needs a limit: abstraction is not causality, and prediction is not intervention.

In a hypothetical software incident, “timeouts often occur alongside large files” describes an association. “Splitting large files will reduce timeouts” is an intervention claim. The latter needs additional mechanistic evidence or a controlled comparison; it does not follow directly from the former.

“Reduce data dimensions, increase causal dimensions” may be an evocative slogan, but it is not an established law. A shorter summary may have deleted the crucial exception. A cleaner rule may simply have made a misunderstanding more durable.

Compression should therefore be judged not only by how much text disappears, but by how much decision-making ability survives in new situations—and whether contrary evidence can still correct the representation.

4. World dynamics needs testable loops, not longer chains

I have tried connecting energy, information, life, meaning, consciousness, attention, language, value, and action. Such a chain can be illuminating, but it is also dangerous: the smoother the arrows look, the easier it is to mistake a diagram for an explanation.

I now want to narrow the question: how does an agent form representations adequate for action, and how do the consequences of action change that agent?

As a working framework:

World and internal state → selective representation → prediction and decision → action → observable consequences → model revision.

Goals, resources, and authorization constrain the entire loop. They do not appear only at the final decision.

Meaning is no longer merely a label attached after information arrives. I would rather ask whether a distinction changes an agent’s situation and whether the agent can adjust its actions accordingly. In one theoretical framework, Kolchinsky and Wolpert define semantic information as environmental information causally necessary for a system to maintain its own existence.[3] This does not encompass all human meaning, but it offers a testable starting point: what is lost when a particular kind of information is removed?

An account of biological systems must not, however, become an AI objective by analogy. OUTBIRD’s goals should come from user authorization and value constraints, not from treating its continued operation as the supreme objective. Stopping, revoking, and correcting the system must remain part of its design.

With multiple agents, the loop gains another complication: while I act, others also interpret, predict, and act. Language makes parts of our models expressible and negotiable. Shared expression, however, does not guarantee shared understanding or aligned goals.

In a hypothetical team, “done” might mean the code has been written, deployment has succeeded, or the user has verified that the feature works. The word is identical; the world states are different. Organization is therefore not just communication between models. It also involves evidence, authority, incentives, conflict, and shared standards.

For world dynamics to explain anything, these distinctions must be unpacked rather than hidden behind words such as “recursion” or “emergence.”

5. Dissipation supplies physical constraints, not a complete theory of intelligence

The connections among complex systems, self-organization, and dissipation still fascinate me. But I no longer want the sequence “energy dissipation → life → intelligence → meaning” to stand in for all the missing mechanisms.

Nonequilibrium thermodynamics can address how physical systems maintain states and how information processing relates to energetic costs. That does not make it a complete explanation of memory, language, or organization.[3][4]

The Thermodynamics of Prediction provides a more specific connection. Under assumptions including an external stochastic driving signal, nonpredictive information retained by a system has a quantitative relationship to a form of dissipative inefficiency.[4] Those model conditions matter. The result does not directly imply that storing fewer logs improves intelligence, nor does it make information compression, thermodynamic entropy, and software architectural disorder interchangeable quantities.

For me, the research suggests a question rather than a conclusion: how much of what a system pays to remember the past actually serves its future?

Likewise, “quantitative change becomes qualitative change” is a summary, not a mechanism. A particular capability transition still requires an account of the changing variables, constraints, feedback, and possible tests. “That is how evolution works” is not the end of the explanation.

I therefore want to retain three connected but noninterchangeable levels: physical conditions, cognitive mechanisms, and organizational rules. Explaining their connections is a research task, not a rhetorical move.

6. OUTBIRD: from accumulating experience to improving the next action

Applied to OUTBIRD, these ideas first change the acceptance question, not the list of modules.

I do not want agent counts, memory records, API calls, or reports to stand alone as evidence of progress. A better question is whether, under comparable constraints, the system acts more reliably, repeats fewer mistakes, and requires less intervention from me.

The following are directions I want to test within the existing system. They are not a proposal to rebuild the architecture, and they do not imply that the capabilities are already deployed.

Memory must preserve evidence—and remain corrigible

I want to separate three things: events that actually happened, judgments inferred from those events, and rules proposed for future use.

Several deployment failures may suggest a common pattern, but three co-occurrences do not confirm a cause. Every extracted rule should retain its sources, applicability conditions, counterexamples, and conditions for invalidation. A new summary must not overwrite the underlying evidence or compress “unknown” into “resolved.”

Consolidation should mean proposing testable hypotheses from experience and allowing them to influence action after validation—not rewriting a long conversation as a shorter, more confident one.

Planning should include testable probes, not just longer reasoning

When uncertainty is material, I want the system to compare a small number of feasible approaches: what each assumes, where each could fail, and which minimal test would eliminate a mistaken direction.

An internally imagined outcome remains a hypothesis. Sandboxes, read-only checks, historical replay, and controlled small-scale execution are valuable because they obtain feedback at different costs—not because they decorate a confident guess.

High-risk, irreversible, or externally consequential actions still require the relevant permissions and approvals. Being able to predict an action does not authorize its execution.

Routing should learn conditions rather than freeze provider labels

I want model selection to remain provider-neutral, prioritizing actual delivery quality and speed rather than simply minimizing the price of one call.

The useful question is which path works reliably under which task, context, tool, and recent runtime conditions. A failure should first be distinguished as a reasoning, permission, context-transfer, network, or execution-environment problem rather than being collapsed into “the model is bad.”

Historical success rates should not automatically become the probability of the next success. Unequal task assignment, limited samples, and changing conditions all need to be addressed in evaluation design. My engineering requirement is to preserve uncertainty and calibrate judgments against comparable tasks and subsequent outcomes. Feedback counts as learning when it changes routing behavior, not merely when it adds another chart.

Multiple agents need shared state, not just messages

I want a handoff to establish the goal, the location of evidence, what has happened, what remains a hypothesis, who is authorized to act next, and what constitutes completion.

“Code submitted,” “review passed,” “deployment succeeded,” and “user verified” should likewise remain distinct states. A handoff must not silently promote one state into the next.

In engineering terms, understanding another agent is not giving it a human-like personality. It is knowing its available context, tools, and permissions; which evidence supports its output; and under which conditions an assessment of its capabilities remains valid.

7. Demonstrate one small learning loop before claiming self-evolution

If these ideas are worth adopting, the first step should not be another expansive “world-model platform.” I would rather select a frequent, low-risk family of tasks with clear outcomes and run a controlled comparison.

Measure the original process end to end, then add one change—for example, traceable experience consolidation. Test on new, comparable tasks: does it repeat fewer errors, reach external acceptance sooner, and reduce human intervention without increasing unauthorized or incorrect execution?

Only then should the change expand to other tasks. If results do not improve, revise or withdraw the hypothesis rather than interpreting failure as proof that more modules are required.

I also no longer want a conceptual expression such as “prediction quality times action capability divided by information and compute” to create an apparently precise intelligence score. Its quantities and measurement methods are not unified; a multiplication or division sign does not solve that problem.

A plainer acceptance criterion is more useful: under explicit quality, safety, and resource constraints, what was actually accomplished—and what was learned from the outcome?

Conclusion: compress the world, but let the world correct you

Moving from complexity to models, and from compression to greater capability, is not a one-way road to success. A model may erase important distinctions. A rule may entrench a bias. An organization may learn only to confirm itself.

What I want to retain is therefore not just “compression can expand intelligence,” but its necessary qualification: the expansion can become dependable only if the compressed model remains subject to tests against reality.

For world dynamics, that means fewer all-explaining arrows and more testable mechanisms. For OUTBIRD, it means less reliance on internal activity as evidence of progress and more reliance on external outcomes as tests of learning.

Intelligence is not fitting the entire world inside oneself. It is learning which distinctions are worth acting on—and recognizing when those distinctions were chosen incorrectly.

References and further reading

[1] Max Bennett, A Brief History of Intelligence: author’s book page, and Five Breakthroughs: A First Approximation of Brain Evolution From Early Bilaterians to Humans, 2021. The evolutionary framework and this essay’s engineering analogies should be understood separately.

[2] Naftali Tishby, Fernando C. Pereira, and William Bialek, The Information Bottleneck Method, arXiv version submitted in 2000.

[3] Artemy Kolchinsky and David H. Wolpert, Semantic Information, Autonomous Agency, and Nonequilibrium Statistical Physics, 2018. This essay draws on the framework’s intervention-based, testable approach rather than treating it as a complete theory of human meaning.

[4] Susanne Still, David A. Sivak, Anthony J. Bell, and Gavin E. Crooks, The Thermodynamics of Prediction, 2012. Its quantitative results depend on the paper’s physical-system and driving assumptions.