This paper proposes a conceptual architecture for self-developing artificial intelligence.The central claim is that intelligence should not be identified only with one-shot performanceunder a fixed objective. Rather, intelligence is understood here as the capacity to learnand examine through experience by means of language in a broad sense, and to grow byupdating a cognitive loop and its embodied languages. In this view, learning is not merely theaccumulation of knowledge or the adjustment of parameters. It is also a linguistic adaptationto an environment, where the environment includes objects, users, dialogue histories, externalmemories, domain plugins, auxiliary bodies, evaluation logs, media, and social contexts.The architecture proposed in this paper is called BrainOS. It is not a single large prompt,nor is it an external add-on placed outside a language model. Rather, it is a cyclic cognitivearchitecture in which an LLM, BrainOS, external memory, domain plugins, auxiliary bodies,self-audit, development history, and dialogue history participate in processes of generation,observation, explanation, evaluation, memory update, plugin update, and redesign. Promptsare treated not as mere instructions, but as media that describe and generate cognitivestances. However, BrainOS is not reducible to prompt engineering. Prompt generationcontrol is connected to domain plugins, auxiliary bodies, embodied languages, higher-orderplugins, and, in the long run, plugins for LLM design and LLM-generation research.The paper introduces the notion of explainable embodied compression. Experience is notonly stored; it is compressed into bodily and linguistic organs that make future cognitionlighter while keeping the reasons for selection, rejection, success, failure, discomfort, andrevision traceable. Domain plugins are not repositories of expert knowledge. They areorgans that accumulate observations, success conditions, failure classifications, candidatecompressionrules, auxiliary-body usage, and update histories. A higher-order plugin thenreduces such domain experience into a general and reusable form with other plugins in view,and returns it to BrainOS.This paper also reads the No Free Lunch theorem as a limit on universal optimality bya single fixed algorithm, and positions BrainOS not as one universal procedure but as amultiple-activation variable algorithm that selects, combines, deepens, and stops cognitiveorgans according to the situation.As an exploratory diagnostic example, the paper analyzes a simple route-counting problemon the multiplication table, under the condition that programs are not allowed. Theexample shows that even elementary problems may expose implicit assumptions, prematurereduction to a standard problem, confusion between value and position, failure to extractgenerative rules, erroneous compression, and memory residue. The point is not that BrainOSguarantees correct answers. Rather, it makes visible which cognitive organs are missing orimmature. The paper also discusses personal-level optimization as a problem of translationamong thought languages, expressive forms, emotional diversity, domain languages, andauxiliary-body formats. Finally, it revisits philosophical zombies and the Chinese Roomfrom the standpoint of multilayered cyclic cognitive systems, arguing that consciousness,subjectivity, and life should not be fixed in advance as human-only categories.
Yoshiki Ueoka (Mon,) studied this question.