First-hand information for everyone
Main discovery Supported finding (see F1, F11): a model’s context and trained weights do not, by themselves, demonstrate the presence of deliberately constructed systemic representations of the world of activity. These systemic representations are schema-like models of processes, structures, timelines, and connections that a recipient uses to form or understand messages (a schema is a structured model of a situation). (F1, F11)
Method and architecture The paper proposes a five-level content organization for strong language-based artificial superintelligence (ASI). The levels are: (1) deliberately constructed systemic representations of the world of activity, (2) a situation cut out from those representations (a situation’s schema-model), (3) a semantic or narrative description of that situation, (4) the full sense of a particular message (the intended content and its essential conditions), and (5) text (the linguistic realization). The author links this organization to existing theories in systems methodology, Korzybski’s structural differential, and Meaning–Text theory. (F2, F4)
What ASI means here ASI, in this account, is not just a model that writes text or edits code. ASI means a model that can reconstruct the systemic organization of activity: schemas of systems and subsystems, processes and structures, synchronized timelines, and the connections from which a situation is cut. Mere fluent output is not sufficient evidence of this ability. (F5, F11)
How understanding and generation proceed Understanding is bidirectional. Production goes from systemic foundations down to text. Comprehension proceeds in reverse: from the full sense of a message to its extended description, to the situation’s schema-model, and back to the systemic foundations. This “reverse work” is central to the proposed architecture. (F3)
Concrete empirical points The paper cites empirical findings that illustrate current model behavior. In one study (“Elias in the Lighthouse, Again?”), 88.3% of 20,000 stories generated by four language models contained at least one of 11 persistently recurring words, while only about 3.8% of the fine-tuning data contained such words. This suggests repetitive or persistent patterns in generated text that do not by themselves show deeper systemic reconstruction. Other experiments show models perform substantially better when an explicit description of a structure (for example, a topology) is provided than when they must infer it from context. These results separate the presence of material (what is given), its internal representation (what the model stores), and its functional use (how it grounds meaning). (F7, F8)
Limits of next-token training The technical pretraining task used in GPT-style models is next-token prediction. Predicting likely token sequences explains why a model selects a probable formulation. But token probability alone does not ensure the preservation of an intended sense or the permissibility of a formulation. Therefore, moving from token prediction to reliable linguistic realization requires separate checks that the chosen text preserves the intended content. (F6)
Engineering precedents and the paper’s contribution The separation of substantive message organization from linguistic realization has precedents in natural language generation engineering (document planning, microplanning, surface realization). The paper supplements these engineering divisions by adding explicit systemic foundations and a clear distinction between description and the full sense of a message. (F9, F2)
Significance and practical implications To claim a model has ASI in this sense requires evidence that it reconstructs and uses systemic schemas, not just that it produces high-quality text or can modify its code. The fourth level—the full sense of a message—is distinct from text. The full sense must be checked separately. Text may be abbreviated only when recipients can reconstruct the full sense using jointly constructed representations and joint activity. Text must not introduce new assertions or change essential connections. (F5, F13)
Supported caveats and limitations The presence of trained LLM weights or available corpora does not, by itself, prove that a model has acquired deliberately constructed systemic representations. Determining that requires showing the model distinguishes processes, structures, and connections and uses these distinctions to cut out situations and form or understand messages. Reconstruction from a limited message does not mean recovering every detail of the original situation. It concerns only the connections essential to understanding. Connections neither expressed nor confirmed must remain unknown and must not be filled in as established facts. Opacity in the sources of model answers—entanglement of parametric (stored) and contextual knowledge—makes it hard to establish where an answer comes from and motivates proposals to generate answers from these sources separately. (F11, F12, F10)
Bottom line The paper argues that current language-model architecture and evaluations do not by themselves show that models possess the systemic, schema-based foundations needed for full, verifiable utterance formation. The paper proposes organizing content across the five levels and developing methods to check the full sense of messages beyond token-probability metrics. (F1, F2, F5, F13)