Intelligence that can regain its bearings

When remembering an answer is not enough

Imagine an assistant preparing a business proposal. During the work, a constraint emerges that changes the solution. The next day, the system finds the proposal, but it also needs to understand the source of that constraint, why it changed the decision and whether it remains valid. Preserving this relationship makes it possible to continue from the project’s present state.

Orientation concerns what makes an answer relevant: people’s needs, sources, conditions and the consequences of the work. When new information arrives, the system needs to be able to reassess its interpretation and recognize the competences required. Memory, tools and models participate in this function through different roles.

A representation is a point of view

A table, summary or plan makes work legible by selecting aspects of it. For example, a timing table may show that a draft is produced faster. Yet the receiver may need to reconstruct sources or correct an exception: understanding the outcome requires returning to the process to which that measure belongs.

The Dual-Non-Dual perspective, abbreviated D-ND, holds useful distinctions together with the field that gives them meaning. Goal and means, model and source, person and system can be distinguished without exhausting the situation. The capability to form is recognizing when a partial interpretation is taking the place of the actual work and recomposing that relationship.

Orientation that forms again in the present

A gyroscope helps us imagine a reference that enables orientation during movement. In this research, “logical gyroscope” names a cognitive function: recovering source and reason as context, tools and possibilities change. Alignment thus takes the form of dynamic orientation. When a consequence changes the project, the reference forms again too: the new starting point preserves how it was reached without exhausting subsequent possibilities. D-ND calls this relationship a “moving zero”.

In practice, this means being able to ask: where does this choice come from? Which information corrects it? Which consequence have we observed? The system needs room to explore possibilities and return to the source. Concrete actions then require the access, responsibilities and authority relevant to the case. The gyroscope analogy makes the proposed function legible; its behaviour needs study in the environments where it is exercised.

Dissonance can change what we are able to see

In a document workflow, almost every request follows the same sequence. A note outside that pattern may be an error, or may indicate that a client needs a different step. Discarding it immediately means deciding before understanding the relationship; treating it as a new rule may distort the whole process.

The D-ND perspective allows that difference to remain as potential while the system looks for the sources and conditions that make it relevant. If it truly changes the case, it enters the decision and the continuation. If it does not participate, it can fade. Sufficient understanding enables action: openness serves to form the choice and includes the point at which that choice becomes practicable.

Today’s kernels: a structure within the work

kernel_chat 1.0.0 connects conversational work to sources and competences. MAIOS Project Kernel 5.1.0 brings state, knowledge, competences and operating rules into the project. Setup AI, through Form 0.16.1, prepares an initial configuration from the described context. These are the adoption entry points available on the home page.

The external kernel makes references explicit for the AI system to use in its work: where to resume, which sources to consult, which method to make relevant and where to return learning. Continuity takes shape when these relationships are actually exercised and corrected through consequences. The System Semantic Kernel, or SSK, describes this semantic organization in the research paper.

Bring orientation into model formation

The next direction is to study how these relationships can enter the formation of a large language model, also called an LLM. Formation material should preserve how an interpretation arises, which distinctions it introduces, how it returns to its source and how a consequence changes the continuation. A useful example therefore includes the path that made an answer appropriate.

Part of the research concerns distinctions already present in model formation material: categories and representations help interpretation, yet may also make a different field difficult to recognize. The D-ND hypothesis is to form the ability to use and reassess these points of view. The technique remains to be determined: data, curricula, learning objectives, architecture, post-training and inference can participate in different ways. Corpora, compute and comparative experiments are work to design and perform; published kernels currently provide an external foundation for exploring the relationship.

Learning also changes how capabilities are formed

Suppose the system misinterprets a request because it used a summary before the original source. Correcting the answer resolves that case. Changing the competence means preserving how to recognize and recompose that substitution. In later, different work, such as a technical review, the capability can become relevant before the same form of error repeats.

A further step concerns how the system recognizes which capabilities are needed and combines them. If experience changes this function, it changes the way learning happens: this is the level of meta-competences. Knowledge can act on the method through which further knowledge is produced; D-ND calls this relationship “autologic”. The usefulness of the change is observed in subsequent work where the new capability is exercised.

When consequences also return into the conditions through which the system preserves and transforms its own organization, the autopoietic direction opens. Here, “autopoiesis” concerns the system’s formation and continuation. The manifesto develops this trajectory and its meaning for synthetic intelligence.

The model belongs to a wider system

This direction also concerns how competences form and cooperate. Experience from a project can teach a method, make it reachable for other work and change how the system discovers relevant resources. A method can take shape from the problem, people and available means while preserving the possibility of correcting its first coherent solution.

A model can be an organ of a cognitive entity comprising memory, competences, tools, perception and relationships with people and organisations. Research assistants, engineering systems, organisations and embodied systems such as robots are directions to explore with their own environments and tests. Cognitive orientation is the thread of research across these possible forms.

From research to a concrete business case

For a business, a first discussion can address internal knowledge, workflows, decisions and AI integrations. It starts from a real case and available sources to scope a configuration, pilot or agreed development work. Use reveals which capabilities are useful, which conditions limit them and which can become repeatable. AI-SaaS products may take shape from those repeatable capabilities: this is a development direction to determine through cases and tests.

The relationship is reciprocal: research changes kernels, real work changes competences and research questions, and more native formation may open up new products. The trajectory requires people, corpora, infrastructure and compute resources. Anyone assessing a collaboration can distinguish what is available, hypotheses to exercise and the next evidence needed. The nine paths help choose a first job; consulting provides a route to discuss a case.