Terminology

So many terms are thrown around. AI can reason, LLMs conquered language and generate meaning, machines are intelligent. In this post, I will explain how I relate those terms to human cognition and how what machines do is different.

Intelligence

Intelligence is the ability to handle differences. Handling includes recognition, processing, and producing. Intelligence, encountering any phenomenon and having some purpose in mind, asks and answers two questions: “Does it make a difference?” and “What difference does it make?” One may also frame it as whether the phenomenon is relevant to the purpose.

Differentiation

To differentiate is to compare. Comparison is dimensionality reduction operation. By dimensions, I mean comparable properties. They may be tangible (bricks and mortar) or abstract (in terms of components, a house and its ruins are the same, it’s their “organization” that makes the difference).

Why is comparison a dimensionality reduction operation? Because any comparison takes into account only one property? Actions are also dimensionality reduction operations. Any action affects only a subset of properties of any object involved. Dimensionality reduction decreases computational load. It is important.

Ranges

View any property as an axis. Break it into ranges. By having a limited number of ranges you again decrease computational load. But most importantly you gain something important compared to point-accurate measurements. You gain interchangeability within each range. A range is what is meant by a concept. Its role is to define fitting and non-fitting phenomena. For that, a range’s boundaries are only relevant. Please make sure that there are no other constraints on the fitting objects. They are not determined by anything else besides the boundaries. In terms of other properties, the fitting objects may or may not be similar. It does not make a difference. It makes similarities irrelevant.

Cognitive Computation

Comparison and selection based on comparison constitute cognitive computations. It makes comparable properties the atoms of cognition. Not objects, not actions. Those are important, but not foundational.

Why do differences and boundaries always win over similarities? Because recognition relies on multiple properties and using differences of ranges based on their boundaries allows for discarding non-fitting categories. Using similarities of fitting instances provides no way to discard non-fitting categories, making computations infeasible in real time.

Ontology/Epistemology/Causality

Traditional approach is to group objects and assign to them actions, which those objects can perform or which can be performed on those objects. I propose to consider properties and actions that allow for modifying those properties. Various parameters of those actions can be organized hierarchically to achieve efficient logarithmic complexity.

The more parameters and levels any such hierarchy considers, the finer control of the results is achievable. That is why specialization (introduction of properties and differences) is important.

Presence of some properties as parameters of actions affecting other properties makes the former relevant for the latter.

Compositionality/Generalization/Continual Learning

Objects carry multiple properties, actions use various parameters (expressable in terms of properties). Any goal or purpose can be expressed in terms of differences in some properties.

The hierarchical organization of causal effects, considered above, explains generalization as ignoring differences and moving up the tree. The use of different properties at each level paves the way to understanding compositionality. Learning about any one property allows to make hierarchical trees involving that property deeper and finer tuned in terms of possible results. This is how continual learning proceeds, it adds more factors to consider in the Semantic Binary Search behind any cognitive function.

Data, Information

The rules for information processing introduced by Shannon, need to be updated. Ranges of properties may be of different width within the context of any action or purpose. Any task may have different considerations for what is considered interchangeable. The first meter is not different from the tenth meter, the first apple is not different from the tenth apple, but this is not so for cognitive processes.

What Shannon means by data are in fact symbols. Variability of symbols is different from that of ranges of comparable properties through the prism of any purpose.

Knowledge, Facts

Consider differences among entries in the same SQL table and compare those to differences among first entries of various SQL tables. The latter are semantic differences. Intelligence is interested in those.

Understanding, Reasoning, Inference, Creativity

Intelligence is pragmatic, it works with what it has. Both in terms of objects available in context and in terms of known recipes. Under real-time pressures, there is no time to “create” any of those. It is not creative in the strict sense. What may compensate for that is its ability to view objects from various “angles” in terms of properties and use them in unusual ways. Probably that will count for “understanding.”

But intelligence knows the “value” of having many options to choose from and the importance of having at least one option. That is why intelligence is curious and strives to fill gaps in knowledge.

Reasoning, inference, planning, adaptation, and other similar cognitive functions may be represented as the applications of the core algorithm - selection of the most fitting option from the available ones respecting relevant constraints. It is best demonstrated by the game 20 Questions for the task of categorization. Categories are available options, properties of an object being categorized are constraints. In different tasks, those are determined depending on a situation.

Language

The main problem I see in most textbooks about linguistics or philosophy of language is the attempts to analyze sentences without context. It follows from misunderstanding the role of language. I see it as pointing. Unlike a finger, however, language has advanced pointing capabilities. It can point at objects real or imaginary, from the past or future, it can point at actions, properties, relations. But most importantly, it can point at facts or connections of all of the above.

In performing its role, language doesn’t have to encode everything. It relies on a listener being intelligent and having all the cognitive abilities. It is enough to indicate only those properties of relevant phenomena that allow the listener to differentiate those from the context. That is why context is crucual. Many experts consider context-dependence an issue of language along with polysemy. But the latter is the way to ensure language efficiency, while the former is the way to handle the latter.

Intelligence relies on ranges of comparable properties (aka concepts). They serve as filters. Words refer to those concepts (one to one, one to many (polysemy), many to one (synonymy)). Those references are arbitrary, but need to be shared among all the community members. Also they need to pass the test of time - to ensure that they allow for efficient disambiguation.

Meaning can be easily defined as the phenomena in context pointed by a word, phrase, sentence, or text. References work by filtering. If one filter is not enough, filters may be stacked (aka compositionality).

Disambiguation relies on the principle noticed by Firth - you shall know a word by the company it keeps. Consider how the core algorithm applies here - each word may have many “meanings,” but not all of them are compatible with meanings of neighbor words. Meanings of each word serve as options to select from based on meanings of neighbor words serving as constraints.

Yes, language relies on the core algorithm of intelligence, which makes it the first and so far the only intelligent tool created by humans. Hopefully, soon we will invent another one.

Intelligence vs Automation

Inter + leggere or “reading between the lines” is our ability to estimate and use differences. Whatever is not different may be considered similar or interchangeable. Teachers of physics may look different but their profession makes them the same. Intelligence pays attention to differences, not to similarities. Hierarchical organization of properties and their ranges makes intelligence efficient.

So far, humans are considered the most intelligent species. What makes their intelligence different from what is demonstrated by software (including GOFAI and ML) or automation? Is it about repeating the same thing over and over, expecting different results?

Recall the semantic differences from the SQL example above. Traditional machine learning is similar to working with a single SQL table. True intelligence (as is the case with all life forms) can change the set of tracked properties and their breakdown into ranges and how they are used. This implies also their ability to change the set of possible goals. True intelligent agent can form and check hypotheses before incorporating novelties into their toolbox. New properties may be the result of using new sensory or actuation tools or composition/derivation from the existing properties.

Social intelligent agents can acquire those changes as a result of communication.

Flawed Ideas

We are interested in reliable recipes for achieving various results. The hierarchical organization of causal relations allows for choosing the best course of actions based on the constraints in the current context. Intelligence provides no guarantees, it relies on the fact that relevance is already reflected in those hierarchies. No additional metrics are required. Real time leaves no time to calculate them.

A perfect prediction is impossible. Real time leaves no time to calculate multiple imperfect predictions to select from based on additional metrics.

Perfect explanations are also impossible. Newton did not explain gravity, he provided formulas to calculate its effects. The use of metaphors for various phenomena (like river for time or tiny balls for particles) has some utility, but also has limitations and misleading effects.

Consider the metaphor of “computation” for cognition. Once it is used, it is natural to apply arithmetic calculations (like matrix multiplication) for various functions. I view it as misleading. On the other hand, I do think that cognition is based on computation, but that computation is comparison-based selection. Not the first kind of computation, which one thinks about, hearing that word.

Logic and symbol manipulation only seem logical as the basis for intelligence, but they also have limitations. For any rule out there it is possible to find exceptions. Logic breaks on those. Logicians and philosophers love mentioning truth, but even a tiny deviation is enough to deny truth of any statement. The real world is impossible without deviations. Instead of truth, consider what is “good enough” for your purpose.

***

Intelligence uses Semantic Binary Search to select the most fitting option from the available ones respecting relevant constraints. It relies on comparison of ranges of properties. Semantics is determined by differences, not by similarities. Efficiency is achieved by dimensionality reduction and binary search over limited number of options.