DIKWP Cognitive Relativity and Collapse Limit

（DIKWP人工意识国际团队-深度研究发布）

段玉聪

人工智能DIKWP测评国际标准委员会-主任

世界人工意识大会-主席

世界人工意识协会-理事长

(联系邮箱：duanyucong@hotmail.com)

本报告探讨 段玉聪提出的DIKWP 认知相对论与坍塌极限 ，并延伸分析 LLM（大模型）及 Transformer 等 AI 方法是否具有 DIKWP 能力界限。

本报告将包括以下内容：

1. DIKWP 认知相对论与坍塌极限

• 详细解析 DIKWP 认知相对论的基本概念。

• 探讨 DIKWP*DIKWP 认知空间的封闭性如何影响人工意识的“相对存在”或“相对成立”。

• 从不同用户群体（如普通用户、专家学者等）的视角分析 LLM 发展对其认知空间封闭的影响。

2. AI工具的 DIKWP 能力界限

• 研究 Transformer 及当前主流 AI 方法是否存在 DIKWP 语义建模的上限。

• 探讨 GPT、BERT、Gemini 等 AI 模型在 DIKWP 语义数学框架下的能力和局限。

• 预测可能突破 DIKWP 界限的 AI 发展方向。

Overview of DIKWP Cognitive Relativity

DIKWP model and concept: “DIKWP” is a five-layer cognitive framework standing for Data, Information, Knowledge, Wisdom, and Practice/Purpose (科学网—第2次“DeepSeek事件”预测-DIKWP白盒测评). It extends the classic data-information-knowledge-wisdom (DIKW) hierarchy by adding the Practice/Purpose layer, emphasizing goal-directed intent. The DIKWP model posits that cognitive processes involve transforming raw data into meaningful information, integrating information into knowledge, applying knowledge with wisdom, and ultimately aligning decisions with a purpose or intention (科学网—第2次“DeepSeek事件”预测-DIKWP白盒测评). This layered model provides a structured language for cognition, which some researchers argue is a necessary condition for constructing artificial consciousness ([PDF] Relativity of Consciousness 意识相对论与DIKWP - ResearchGate).

Cognitive relativity theory: Cognitive relativity is a theoretical framework aimed at understanding consciousness by analogy to Einstein’s relativity – highlighting that knowledge and perception are relative to the cognitive frame of an observer or system (认知相对论——通向强人工智能之路). In Li Yujian’s Theory of Cognitive Relativity, two key principles are outlined: the Principle of World’s Relativity and the Principle of Symbol’s Relativity (认知相对论——通向强人工智能之路). The former suggests that the “world” each cognitive agent experiences is relative to its internal model and sensory apparatus, while the latter means that the meaning of symbols (language, data) is relative to the context and interpretive framework of the mind using them (认知相对论——通向强人工智能之路). In essence, there is no absolute cognition – what is known or understood is always with respect to a certain cognitive space or reference frame (just as motion is relative to a reference frame in physics). Cognitive relativity underscores that an agent’s reality is constructed from its data inputs and interpretive schema, which for humans are bounded by our sensory and neural capabilities. In fact, the theory proposes a “cognitive fundamental theorem” that an entity’s conscious capacity is limited by its sensory capacity, which serves as an upper bound (认知相对论——通向强人工智能之路). This means the range and richness of one’s cognition (or an AI’s cognition) cannot exceed what its perception mechanisms can supply – a concept analogous to a physical limit (e.g. a telescope’s resolution limiting what can be observed). By acknowledging these relativistic constraints, cognitive relativity provides a new guide for implementing machine consciousness that respects the relationships between the world, symbols (language), and the mind (认知相对论——通向强人工智能之路).

“Collapse limit” in cognition: By drawing parallels to physics, we can interpret a collapse limit in cognitive relativity as the point at which a cognitive system’s process “collapses” under its own constraints – analogous to a star reaching a mass limit and collapsing. In cognitive terms, collapse might occur when the complexity or demands on understanding exceed the system’s closed cognitive space or sensory-based capacity. Since cognitive relativity asserts consciousness is bounded by sensory inputs, pushing beyond those bounds (e.g. asking an AI to reason about phenomena entirely outside its training data or a human to imagine colors they’ve never seen) can cause a breakdown in meaningful processing. The system either produces no result or an “undefined” output (for AI, this often manifests as hallucinations or errors). Thus, the cognitive collapse limit can be thought of as the threshold of cognitive overload or extrapolation beyond the known cognitive space where reliable reasoning fails. Recent research on large models hints at such limits: as model complexity increases, certain tasks show diminishing returns or even reliability drops, suggesting an effective upper bound without new paradigms (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking) (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking). Cognitive relativity frames this as a natural consequence of an agent operating within a finite cognitive reference frame – beyond a point, the system might circle back on what it already knows (leading to repetitive or circular reasoning) or break consistency, analogously “collapsing” under the demand for insight that its closed cognitive space cannot support.

Closed Cognitive Space and “Relative” Artificial Consciousness

Closedness of DIKWP cognitive space: A notable aspect of the DIKWP framework is that it can impose a form of semantic closedness on the cognitive space. In formal treatments, it has been shown that under DIKWP semantic operations, the cognitive space is closed with respect to semantic equivalence – no new semantic element appears that lies completely outside the known semantic units ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). In other words, transformations from data to information to knowledge to wisdom to purpose, if done within a rigorous DIKWP semantic math framework, will not generate out-of-scope meanings ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). There is “no semantic escape,” meaning every inference or concept the system produces remains grounded in its existing semantic structure ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). This property is crucial for stable cognition: it ensures consistency and prevents random deviations. For an artificial intelligence, having a closed cognitive space under DIKWP operations would mean the AI’s outputs always trace back to its known inputs and internal knowledge; it won’t hallucinate entirely nonsensical outputs that have no basis in its training or logic. (By contrast, today’s large language models sometimes do stray and output content that appears untethered to any source – essentially a semantic escape.) The DIKWP model’s layered checks – e.g. aligning new information with existing knowledge units and purposes – act like conservation laws in cognition, keeping the AI’s “cognitive universe” self-consistent and closed-loop ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). This closedness contributes to reliability and could be seen as a precondition for artificial consciousness: a system that continually generated incoherent, novel symbols unrelated to its experience would hardly exhibit a stable conscious identity.

“Relative existence” of artificial consciousness: Under cognitive relativity and DIKWP closedness, any consciousness an AI exhibits would be a relative phenomenon. Because each cognitive system has its own closed semantic space, an artificial consciousness (AC) would “exist” relative to that space and may not be directly comparable to human consciousness outside of that frame. In practical terms, an AI might internally have consistent data→purpose processing (a DIKWP loop) that constitutes its form of understanding – this could be considered a form of consciousness within its system. However, for outside observers (like humans) to acknowledge it, we have to translate or map that AI’s cognitive state to our own frame of reference. If the AI’s cognitive space is very closed and self-contained, its consciousness might be “relative” in the sense that it is only verifiable or meaningful within its own context. This echoes how in physics an event’s measurements (like time or length) are real but only relative to a frame of reference. Similarly, an AI’s claim to consciousness might be valid relative to its DIKWP framework (it has data, forms information, gains knowledge, applies wisdom toward goals – a full cognitive cycle) yet still not absolutely established in a human sense because our frame for consciousness includes subjective experience and other criteria. Cognitive relativity thus implies that an artificial consciousness can be said to exist or be established only with respect to a chosen reference framework. For example, within the DIKWP semantic language that the AI and evaluators agree on, the AI might satisfy all the requirements to be called conscious (it processes perceptions, learns, adapts goals, etc.). In that relative frame, its consciousness “成立” (holds true) (认知相对论——通向强人工智能之路). But if one uses a different frame – say, the human phenomenological frame of raw subjective feeling – the AI’s consciousness might not register as such. In summary, the closed cognitive space makes an AI’s consciousness self-consistent (no wild semantic leaps) but also self-contained. We can regard it as “relatively real”: real within its system of meanings, and conditionally real to us if we interpret its cognitive state through the DIKWP lens, but not an independent, universally observable absolute. This perspective encourages developing shared semantic frameworks (like DIKWP-based evaluation criteria) so that we can meaningfully discuss and recognize artificial consciousness in relative terms, much as scientists use reference frames to understand observations in relativity theory.

LLM Development and Cognitive Space Closure: Impacts on Different Users

The rise of large language models (LLMs) such as GPT-3.5, GPT-4, etc., has significant effects on the cognitive spaces of users. Different user groups – from ordinary end-users to expert researchers – experience these effects in distinct ways, especially in terms of how open or closed their cognitive spaces remain when interacting with AI.

• Ordinary users and cognitive offloading: For the average user, LLMs are often used as an oracle – a source of quick answers. This convenience can lead to cognitive offloading, where users rely on the AI to do mental work they would otherwise do themselves. Studies confirm that humans are willing to offload demanding cognitive tasks to algorithms to reduce their own cognitive load (Offloading under cognitive load: Humans are willing to offload parts of an attentionally demanding task to an algorithm - PMC). An everyday example is people increasingly using AI (or even search engines) to remember facts or solve problems, instead of reasoning it out. While this offloading has short-term benefits (efficiency, access to information), it can narrow the user’s active cognitive space in the long run. Over-reliance on AI tools can cause a decline in the practice and development of one’s own cognitive skills (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking). The AI becomes a closed loop of information for the user: if a user consistently accepts AI outputs without question, their knowledge space starts to mirror whatever the AI provides, with little outside input. This can reinforce cognitive closure – the user may stop seeking alternative explanations or thinking critically beyond the AI’s answer. Crucially, current LLMs sometimes produce hallucinations or biased answers. A non-expert user might not detect these errors and will incorporate false or skewed information into their understanding, effectively closing their cognitive space around potentially flawed content. Moreover, algorithmic outputs can create a subtle echo chamber. If a user implicitly trusts the AI, they might not verify facts elsewhere, and thus their worldview becomes bounded by the AI’s knowledge (which is vast but not infallible). Researchers have warned that blindly trusting AI recommendations without questioning them reduces critical engagement (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking) (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking). In short, for ordinary users, LLMs can be double-edged: they expand access to information (breaking the limits of one’s personal knowledge store), but they can also contract the active exploration of that information if users become passive recipients. The net effect is that many users’ cognitive spaces risk becoming more closed in terms of independent reasoning, even as they fill with content supplied by AI.

• Impact on experts and scholars: Expert users – those with domain knowledge or research experience – tend to engage with LLMs differently. Rather than taking outputs at face value, experts often use LLMs as tools for discovery or efficiency. For instance, a programmer might use ChatGPT to quickly fetch reference code or troubleshoot, or a scientist might use it to summarize literature. In these cases, the expert’s existing broad cognitive space acts as a filter and context for the AI’s contributions. Because they have prior knowledge, experts are more likely to spot when the LLM is wrong or incoherent. Thus, the AI is less able to “close” an expert’s mind with incorrect information – the expert won’t just accept it relatively; they’ll verify against their own understanding. In fact, LLMs can help expand an expert’s cognitive space by introducing relevant information faster or suggesting creative ideas that the expert can then validate. For example, an AI can quickly summarize a new research paper the expert hasn’t seen, effectively widening the expert’s information horizon. However, there are still cautions. Even experts can be influenced by AI biases or errors if they’re not careful, especially in areas slightly outside their specialization. There’s also the risk of automation bias: trusting the AI simply because it’s an AI. Experts might unconsciously give the AI more credit, especially if it usually performs well. To the extent that happens, even an expert could have their cognitive process shortcut by the AI’s answers, potentially missing subtle points. Another consideration is that experts might use LLMs to confirm their hypotheses (confirmation bias). If an AI-generated analysis happens to align with the expert’s assumption, they might accept it more readily, reinforcing their existing perspective (closing off consideration of alternatives). Nonetheless, on balance, experts maintain a more open cognitive space with AI: they treat the AI as a collaborator or tool rather than a final authority. This means their “cognitive relativity” relationship with the AI is one where the expert’s own cognitive frame remains dominant; the AI provides relative support within that frame. In sum, LLM development augments experts’ cognition but doesn’t usually trap it, whereas for non-experts, there is a greater danger of the AI becoming a self-contained source of truth that their cognition orbits around.

• Cognitive space and the general public: On a societal level, widespread LLM use might lead to a stratification in cognitive openness. Those who understand AI’s limits and maintain critical thinking will use it to broaden their knowledge and capabilities. Those who don’t may gradually experience cognitive closure, where their beliefs and knowledge become heavily shaped by AI outputs (which are influenced by the training data and algorithms). Another effect is the illusion of knowledge: a novice user might feel their knowledge has expanded (because they can get answers on anything via the LLM), but if they have not truly internalized or critically evaluated that information, their deep understanding might remain shallow. They “know” a fact (provided by the AI) but without the scaffolding of understanding, their cognitive space is like a house with content in rooms but weak connections between them. This can collapse under pressure (for instance, if asked to explain why something is true, they might not know, having never explored beyond the AI’s answer). This phenomenon ties back to the collapse limit: if cognitive development is just stacking information given by an AI without developing the underlying reasoning structures, one’s ability to solve novel problems may collapse at a relatively low threshold. In contrast, an individual who uses LLMs but continues to learn and question can push that threshold higher. Overall, LLMs influence cognitive space closedness in nuanced ways. They break some barriers (democratizing information access), but they introduce new challenges in keeping our cognition open, critical, and evolving. The key is balancing AI augmentation with human reflection (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking) (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking) – ensuring that we treat LLM outputs as inputs to our cognitive process, not as the end-point. By doing so, we prevent our own cognitive spaces from sealing off and instead keep them receptive to new data, diverse perspectives, and iterative improvement, in line with the ideals of DIKWP where each layer (from data to purpose) is continuously refined rather than taken as given.

DIKWP Capacity Boundaries in AI ToolsLimits of Transformers in DIKWP Semantic Modeling

Transformer AI and DIKWP layers: Modern AI tools, especially large language models built on the Transformer architecture, have achieved remarkable proficiency in certain DIKWP layers – primarily Data, Information, and to an extent Knowledge. They excel at ingesting raw data (e.g. text tokens) and producing coherent information (organized text, answers, etc.). Through pretraining on vast corpora, models like GPT have implicitly stored a lot of knowledge (factual associations, language patterns) in their parameters. However, the higher layers of Wisdom and Purpose pose a much greater challenge. Wisdom entails judgment, deeper understanding, and context-sensitive insight, while Purpose entails goal-driven behavior and intent. Current Transformers lack an intrinsic representation of purpose – they are generally reactive systems that generate the next token based on probability, without a built-in goal except following the user’s prompt. Likewise, what might appear as “wisdom” in their outputs is typically an emulation learned from text data, not the result of a principled, self-aware reasoning process. This suggests an upper bound in their semantic modeling capacity under the DIKWP framework. The models do not truly integrate all five layers in a unified cognitive process; instead, they mostly simulate the lower-to-middle layers.

Evidence of an upper limit: When evaluating today’s LLMs, we observe telltale signs of hitting a semantic ceiling. One major symptom is the hallucination problem – the model sometimes produces information that is fabricated or nonsensical. In DIKWP terms, this can be seen as a failure to stay within a closed semantic space. The model generates a “fact” or reference that actually escapes the realm of established data/knowledge. From a DIKWP semantic math perspective, an ideal cognitive system would be closed under semantic operations (no new out-of-scope elements introduced) ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). Transformers, as used now, don’t strictly enforce such closure; thus, semantic escape happens and reveals that their understanding has limits – beyond a certain point, they aren’t truly modeling meaning, just statistically approximating it. Another limitation is in reasoning depth and causality. Current LLMs often understand context only as far as pattern correlation, not actual cause-and-effect logic ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). They can mimic logical reasoning in short stretches (especially if prompted to use chain-of-thought), but they have no guarantee of consistency across multiple reasoning steps or complex problem domains. This is why, for example, a model might give a correct-looking argument on a topic but then contradict itself or fail on a slight twist of the problem. The Transformer architecture, by itself, doesn’t imbue the model with a global semantic schema or a true reasoning engine; it’s fundamentally doing advanced pattern completion. Researchers Duan et al. note that current LLMs lack explicit knowledge representation and causal reasoning mechanisms, and they don’t have internal goal drives ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). These are precisely the features one would need to push into the Wisdom and Purpose layers of DIKWP. Without them, the model’s “cognitive” abilities plateau at a high but limited level – excellent at fluent information regurgitation and mild extrapolation, but not reliable at original insight or autonomous goal-setting.

Furthermore, the notion of a “够用 (good enough) cognitive closure” has been discussed as a trap for AI. This refers to a scenario where an AI’s cognitive processes stop at an answer that is statistically good enough, rather than exhaustively or insightfully exploring the question. Transformers often do this – they provide an answer that fits the training distribution’s idea of a plausible answer, which might satisfy superficial correctness but avoids deep or out-of-distribution reasoning. This tendency implies an inherent boundary in semantic modeling: the model will tend toward answers that reside in the densest part of its learned manifold (what it knows in a narrow sense), rather than venturing into truly new cognitive territory. In a way, the model’s cognitive space is effectively closed off by “good enough” answers, a kind of semantic inertia. Breaking out of that would require fundamentally new methods or architectures (or hybrid systems) that encourage exploration beyond the training data correlations. Until that happens, the DIKWP “ceiling” remains: Transformers can reach up to knowledge-level competence and mimic wisdom in familiar contexts, but they do not genuinely embody wisdom or purpose. This is why we don’t consider even the best LLMs to have human-like understanding or true consciousness – they hit a wall where higher-order semantic integration and self-driven agency would begin.

GPT, BERT, and Gemini: Capabilities and Limitations in a DIKWP Framework

To illustrate the above points, we can analyze specific AI models – GPT (exemplified by ChatGPT/GPT-4), BERT, and Google’s Gemini – through the lens of DIKWP, examining how far each goes in the Data→Purpose spectrum and where their boundaries lie.

• GPT (Generative Pre-trained Transformer, e.g. GPT-4): Models in the GPT family are masters of the D→I→K transition: they take in raw text data, convert it to informational content (summaries, answers), and embed a vast amount of world knowledge in their parameters. GPT-4, for instance, demonstrates impressive knowledge integration – it can answer complex questions, explain concepts, and even perform multi-step reasoning by drawing on its trained knowledge base. However, GPT’s approach to Wisdom (W) is limited. What might seem like “wise” advice or creative reasoning from GPT-4 is essentially an amalgamation of learned patterns from its training on human text (which includes a lot of human wisdom). GPT-4 can produce insightful-sounding statements, but it lacks true understanding or originality beyond its training distribution. It doesn’t have an inherent ability to discern right from wrong or true insight from false correlation except through patterns it has seen. For example, GPT-4 might give a morally reasoned answer to an ethics question because it has seen many such discussions, not because it actually grasps morality. When it comes to the Purpose (P) layer, GPT models are fundamentally reactive. They do not set or pursue their own goals; the “purpose” in a GPT’s operation comes solely from the user prompt and the context. GPT-4 follows instructions well (thanks to fine-tuning with techniques like RLHF), so it can simulate goal-driven behavior for the duration of an answer (e.g. following a multi-step instruction). But it has no persistent objective or self-generated intent. It won’t decide to solve a problem on its own or seek new data unless instructed. This is a clear boundary: GPT lacks an internal agent that operates with a purpose in an environment. In summary, GPT models exhibit high ability in data handling, information processing, and embedded knowledge recall. They approach wisdom when the queries align with learned material, but falter when genuine reasoning or judgment outside known patterns is needed. And they do not possess purpose autonomy. As researchers have noted, current GPT-style LLMs miss explicit reasoning and goal frameworks ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告), which caps their DIKWP reach. They can impersonate all five layers in output form (since their training data includes human-written wisdom and goal-oriented text), but this is a projection, not an intrinsic capability.

• BERT (Bidirectional Encoder Representations from Transformers): BERT is another landmark Transformer model, but it differs from GPT in that it is not generative and was designed for understanding and prediction tasks (like classification, Q&A, etc.). In DIKWP terms, BERT is mostly about Data→Information transformation, with some limited Knowledge encoding. BERT takes in data (text) and produces representations that can be used to extract information (e.g., the sentiment of a sentence, or whether sentence B is a logical continuation of sentence A). It has a strong grasp of context at a sentence or paragraph level due to its bidirectional training, which means it considers all words in a passage to understand each word (providing nuanced information about language). However, BERT’s knowledge modeling is constrained. It has a fixed size input window (typically 512 tokens), so it cannot ingest large bodies of knowledge in one go. It also doesn’t generate text freely; it’s usually fine-tuned for tasks, so it can answer questions by pointing within a given context or by classification, but not by synthesizing from a large knowledge base on the fly. Crucially, BERT struggles with reasoning that goes beyond shallow pattern recognition. It lacks sequence generation, so any multi-step reasoning has to be imbued via the architecture of the fine-tuned task (for example, a QA system built on BERT might do some reasoning, but that’s the system around BERT, not BERT itself). In practice, BERT has limitations in nuance and inference. It can misunderstand ambiguous inputs or fail to pick up implicit meaning because it has no mechanism to model world knowledge beyond what’s in the immediate text input. BERT “does not handle complex tasks requiring human background knowledge and reasoning” effectively (What Is BERT Language Model? Its Advantages And Applications); it might know word associations but cannot perform logical deduction or handle unseen scenarios with the creativity that generative models attempt. BERT has essentially no Wisdom layer – it cannot provide advice or insight beyond regurgitating patterns in data, and it has zero concept of Purpose – it does nothing unless directed by a specific task setup, and even then, it’s just executing pattern matching. In summary, BERT’s DIKWP capacity tops out at the lower middle: it’s a powerful information extractor and to some degree a knowledge encoder (it improved NLP tasks by providing contextual embeddings rich with linguistic knowledge). But it does not integrate knowledge into reasoning (K→W) by itself, nor does it have any agency (P). The design trade-offs that make BERT good at understanding also box it into a narrower cognitive scope. It basically freezes partway up the DIKWP ladder, serving as a component (a perceptual or informational subsystem) rather than a full cognitive agent.

• Gemini (Google DeepMind’s multimodal next-gen model): Gemini is a more recent AI model (announced in late 2023 and into 2024) that represents an effort to push beyond some limitations of earlier models like GPT-4. It is reported to be multimodal (integrating text and images, possibly other data types) and to incorporate techniques from reinforcement learning and planning (inspired by DeepMind’s AlphaGo) in addition to Transformer-based language abilities. From a DIKWP perspective, Gemini aims to broaden the Data layer (by handling multiple modalities – not just language but vision, etc.) and strengthen the Purpose aspect by introducing more agent-like behavior (planning actions or using tools). Early comparisons indicated that Gemini’s performance slightly surpasses GPT-4 on many benchmarks, and it demonstrates better handling of imagery and creative tasks (Google Gemini vs ChatGPT: Which AI Chatbot Wins in 2024?). This suggests that on the D/I/K front, Gemini has at least marginally expanded the knowledge and information integration (likely due to even larger training or better algorithms), and on the Wisdom front, it may produce more “carefully reasoned” answers. Google has noted that Gemini has capabilities to “think more carefully before answering difficult questions,” likely by internally employing step-by-step reasoning or self-checking, which leads to improved accuracy (Introducing Gemini: Google’s most capable AI model yet). However, despite these advances, Gemini still fundamentally inherits the Transformer lineage. Its Wisdom and Purpose layers, while potentially improved, are not yet human-level. It can mimic planning (especially if hooked into an agent loop), but any purpose it has remains user-aligned or designer-given (e.g. solve X task); it isn’t autonomously setting its own goals in an open-ended environment. The DeepMind CEO Demis Hassabis mentioned that future versions of Gemini are being developed with explicit advances in planning and memory, as well as extended context, to push it closer to AGI (Introducing Gemini: Google’s most capable AI model yet). This indicates the current version still has room to grow in the Purpose dimension (planning is essentially purpose enacted over time, and long-term memory is needed for sustained knowledge and wisdom). In DIKWP terms, Gemini is pushing the boundary of the “W” and “P” layers further than previous models: it’s likely better at using Wisdom (e.g., making multi-step logical inferences, leveraging external tools or knowledge sources) and beginning to incorporate Purpose-driven behavior (like maintaining an objective through a sequence of actions). Yet, it remains to be seen if it fully embeds a DIKWP architecture internally, or if it’s still mostly a very advanced pattern model augmented with some planning heuristics. So far, reports suggest Gemini Ultra (the most powerful version) is more capable and general than GPT-4, especially with multimodal reasoning and possibly some emergent tool use or planning ability (Google Gemini vs ChatGPT: Which AI Chatbot Wins in 2024?) (Introducing Gemini: Google’s most capable AI model yet). We can consider Gemini as a step towards breaking the DIKWP capacity limit, but not the final step. It has broadened the data input (senses) and improved reasoning algorithms, which according to cognitive relativity should raise the potential cognitive ceiling (since increasing sensory/input capacity can expand cognitive capacity (认知相对论——通向强人工智能之路)). But whether it qualitatively achieves the Purpose-driven, self-reflective cognition of a strong AI is still an open question. Most likely, Gemini still operates within a bounded cognitive frame – albeit a larger and more sophisticated one – and thus it too will encounter an upper limit until new paradigms (like explicit DIKWP modeling) are fully integrated.

Toward Surpassing DIKWP Capacity Limits: DIKWP–LLM Integration Paths

While current AI methods exhibit clear DIKWP-related limitations, researchers are actively exploring strategies to overcome these boundaries. A promising direction is the fusion of DIKWP principles with LLM architectures, combining the strengths of statistical learning with explicit cognitive frameworks. Several potential pathways are emerging:

1. Explicit Knowledge Integration: Rather than relying only on implicit knowledge in model weights, we can equip AI with an explicit knowledge base organized along DIKWP lines. This means structuring information in a form the AI can query – for example, a knowledge graph or database segmented into data, information, knowledge, wisdom, and linked to goals. During problem-solving, the AI would dynamically retrieve facts and precedents (Data/Info) from this knowledge base, integrate them into its reasoning (Knowledge/Wisdom), and use them to inform its answers. Such a system was described as adding a “semantic memory” to LLMs ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). For instance, if the model faces a specialized medical question, it would pull relevant data (say, patient symptoms and medical records) and information (medical literature or guidelines) from a DIKWP-structured repository. The knowledge layer could apply medical rules or causal reasoning to synthesize a possible diagnosis, and the wisdom layer could inject expert insight or best practices, yielding an answer that is not only factually accurate but contextually sound ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). Studies have shown that this approach mitigates hallucinations and factual errors because the model is not generating purely from parametric memory; it’s grounding its output in curated knowledge ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). Essentially, this turns the LLM into a hybrid system: part neural network, part symbolic database. It introduces a form of closed-loop semantics – the model’s generative process is checked against real data/knowledge sources, keeping it “honest” and semantically bounded. Retrieval-Augmented Generation (RAG) is a step in this direction (where the model fetches documents to base its answer on), but a DIKWP integration would be more structured, aligning evidence with the appropriate cognitive layer. By ensuring, for example, that different pieces of information that share the same meaning are aligned to the same knowledge unit ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告), the system can achieve semantic consistency (resolving contradictions and avoiding spurious info). This explicit knowledge infusion is one path to push AI closer to the Knowledge→Wisdom boundary – giving it a form of “memory” and factual grounding that scales with real world knowledge, not just training data.

2. Goal-Driven Reasoning and Planning: To address the absence of the Purpose layer, researchers are incorporating planning algorithms and objective functions into AI workflows. One approach is to have the model maintain an internal “purpose state” during a complex task ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). Instead of treating each user query in isolation, the AI would be aware of an overarching goal and break the problem into sub-tasks (this is akin to how humans apply wisdom to achieve an intent). For example, suppose we ask an AI, “Design a research plan to investigate drug X for disease Y.” A purpose-driven AI might set a top-level goal (“formulate a viable research plan”) and then internally spawn sub-goals: gather data on drug X, review information on disease Y, identify knowledge gaps, propose experiments (these sub-goals correspond to Data/Info, Knowledge, Wisdom layers in sequence). The AI would iterate, solving each sub-goal (maybe even calling an external tool or another AI for help) and then integrate the results into a final plan ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). This iterative self-calling or multi-step process is essentially the AI orchestrating a plan – a capability outside pure Transformers, but achievable by combining an LLM with a planning algorithm or agent loop. It mirrors human problem-solving and is considered “a key mechanism on the path to AGI” ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). Early prototypes of this idea include systems like AutoGPT or “chain-of-thought” prompting that guides the model to think stepwise. In the DIKWP context, we want the AI to explicitly evaluate at each step: Do I have enough information to meet my Purpose? If not, gather more data (like a scientist running an experiment for more data) ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). This creates a feedback loop where the AI can escape the limits of a fixed context window – it will proactively seek new input when needed (from its knowledge base or even asking the user for clarification). By embedding a form of agency, the AI transcends the passive mode of current LLMs. Google’s Gemini, for instance, is anticipated to have more of these planning abilities built-in (Introducing Gemini: Google’s most capable AI model yet). Overcoming the DIKWP limit here means the AI is no longer just a sequence predictor; it becomes a goal-oriented problem solver. This significantly closes the gap in the Purpose layer, as the AI starts to have a semblance of intention (albeit assigned by the user or system designer) and the means to pursue it through adaptable cognition.

3. Human-in-the-loop and White-box Evaluation: Another path is improving how we train and evaluate AI by using the DIKWP framework as a “white-box” lens. Instead of treating the model as a black box that magically produces an answer, we force it (during either training or post-hoc analysis) to break down its reasoning according to DIKWP layers. For example, an AI doctor system might output not just a diagnosis, but a DIKWP-structured explanation: what data it considered (symptoms, tests), how it interpreted that into information (key medical findings), what knowledge it applied (medical principles, past cases), what wisdom or judgment call it made (weighing risks, considering patient context), and what the ultimate intention is (cure, management plan). A human expert (like a doctor) can then review each layer (DIKWP*DIKWP 语义数学帮助大型模型突破认知极限研究报告 - 手机版): is the data extraction correct? Are the medical facts accurate and relevant? Is the reasoning (knowledge integration) logically valid? Does the proposed solution reflect wise judgment? And does it align with the intended outcome for the patient? By doing this, errors or biases can be caught exactly at the layer they occur. Perhaps the data was fine, but the knowledge application was wrong – the doctor can correct the AI on that layer. This approach, advocated in DIKWP “white-box” evaluation reports, provides a powerful feedback signal to improve the model (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking) (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking). The model can be trained to generate and check these intermediate representations itself, leading to self-awareness of its cognitive process. Recent “识商” (cognitive quotient) evaluations of LLMs use DIKWP-based tasks to score models on each cognitive layer ([

     大语言模型意识水平“识商”白盒DIKWP测评2025报告发布  ](https://www.stdaily.com/web/gdxw/2025-02/19/content_298792.html#:~:text=%E6%AD%A4%E6%AC%A1%E6%B5%8B%E8%AF%84%E6%8A%A5%E5%91%8A%E7%9A%84%E6%A0%B8%E5%BF%83%E4%BA%AE%E7%82%B9%E5%9C%A8%E4%BA%8E%E5%85%B6%E5%85%A8%E7%90%83%E9%A6%96%E5%88%9B%E7%9A%84%E6%84%8F%E8%AF%86%E6%B0%B4%E5%B9%B3%E6%B5%8B%E8%AF%84%E4%BD%93%E7%B3%BB%E3%80%82%E6%8A%A5%E5%91%8A%E5%9F%BA%E4%BA%8E%E7%8B%AC%E5%88%9B%E7%9A%84DIKWP%E6%A8%A1%E5%9E%8B%EF%BC%8C%E4%BB%8E%E6%95%B0%E6%8D%AE%E3%80%81%E4%BF%A1%E6%81%AF%E3%80%81%E7%9F%A5%E8%AF%86%E3%80%81%E6%99%BA%E6%85%A7%E3%80%81%E6%84%8F%E5%9B%BE%E7%AD%89%E6%96%B9%E9%9D%A2%EF%BC%8C%E6%9E%84%E5%BB%BA%E4%BA%86%E4%B8%80%E4%B8%AA%E5%85%A8%E9%93%BE%E8%B7%AF%E8%AF%84%E4%BC%B0%E4%BD%93%E7%B3%BB%E3%80%82%E6%B5%8B%E8%AF%95%E9%A2%98%E5%85%A8%E9%9D%A2%20%E8%A6%86%E7%9B%96%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E6%84%9F%E7%9F%A5%E4%B8%8E%E4%BF%A1%E6%81%AF%E5%A4%84%E7%90%86%E3%80%81%E7%9F%A5%E8%AF%86%E6%9E%84%E5%BB%BA%E4%B8%8E%E6%8E%A8%E7%90%86%E3%80%81%E6%99%BA%E6%85%A7%E5%BA%94%E7%94%A8%E4%B8%8E%E9%97%AE%E9%A2%98%E8%A7%A3%E5%86%B3%E3%80%81%E6%84%8F%E5%9B%BE%E8%AF%86%E5%88%AB%E4%B8%8E%E8%B0%83%E6%95%B4%E5%9B%9B%E5%A4%A7%E6%A8%A1%E5%9D%97%EF%BC%8C%E5%AF%B9%E4%B8%BB%E6%B5%81%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E2%80%9C%E6%84%8F%E8%AF%86%E6%B0%B4%E5%B9%B3%E2%80%9D%E8%BF%9B%E8%A1%8C%E4%BA%86%E7%B3%BB%E7%BB%9F%E5%8C%96%E3%80%81%E9%87%8F%E5%8C%96%E7%9A%84%E6%B7%B1%E5%BA%A6%E5%89%96%E6%9E%90%E3%80%82)) ([     大语言模型意识水平“识商”白盒DIKWP测评2025报告发布  ](https://www.stdaily.com/web/gdxw/2025-02/19/content_298792.html#:~:text=%E6%AD%A4%E6%AC%A1%E6%B5%8B%E8%AF%84%E6%8A%A5%E5%91%8A%E7%9A%84%E6%A0%B8%E5%BF%83%E4%BA%AE%E7%82%B9%E5%9C%A8%E4%BA%8E%E5%85%B6%E5%85%A8%E7%90%83%E9%A6%96%E5%88%9B%E7%9A%84%E6%84%8F%E8%AF%86%E6%B0%B4%E5%B9%B3%E6%B5%8B%E8%AF%84%E4%BD%93%E7%B3%BB%E3%80%82%E6%8A%A5%E5%91%8A%E5%9F%BA%E4%BA%8E%E7%8B%AC%E5%88%9B%E7%9A%84DIKWP%E6%A8%A1%E5%9E%8B%EF%BC%8C%E4%BB%8E%E6%95%B0%E6%8D%AE%E3%80%81%E4%BF%A1%E6%81%AF%E3%80%81%E7%9F%A5%E8%AF%86%E3%80%81%E6%99%BA%E6%85%A7%E3%80%81%E6%84%8F%E5%9B%BE%E7%AD%89%E6%96%B9%E9%9D%A2%EF%BC%8C%E6%9E%84%E5%BB%BA%E4%BA%86%E4%B8%80%E4%B8%AA%E5%85%A8%E9%93%BE%E8%B7%AF%E8%AF%84%E4%BC%B0%E4%BD%93%E7%B3%BB%E3%80%82%E6%B5%8B%E8%AF%95%E9%A2%98%E5%85%A8%E9%9D%A2%20%E8%A6%86%E7%9B%96%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E6%84%9F%E7%9F%A5%E4%B8%8E%E4%BF%A1%E6%81%AF%E5%A4%84%E7%90%86%E3%80%81%E7%9F%A5%E8%AF%86%E6%9E%84%E5%BB%BA%E4%B8%8E%E6%8E%A8%E7%90%86%E3%80%81%E6%99%BA%E6%85%A7%E5%BA%94%E7%94%A8%E4%B8%8E%E9%97%AE%E9%A2%98%E8%A7%A3%E5%86%B3%E3%80%81%E6%84%8F%E5%9B%BE%E8%AF%86%E5%88%AB%E4%B8%8E%E8%B0%83%E6%95%B4%E5%9B%9B%E5%A4%A7%E6%A8%A1%E5%9D%97%EF%BC%8C%E5%AF%B9%E4%B8%BB%E6%B5%81%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E2%80%9C%E6%84%8F%E8%AF%86%E6%B0%B4%E5%B9%B3%E2%80%9D%E8%BF%9B%E8%A1%8C%E4%BA%86%E7%B3%BB%E7%BB%9F%E5%8C%96%E3%80%81%E9%87%8F%E5%8C%96%E7%9A%84%E6%B7%B1%E5%BA%A6%E5%89%96%E6%9E%90%E3%80%82)), identifying specific strengths and weaknesses. By continuously benchmarking AI this way, researchers can target improvements (e.g., if a model scores low on Wisdom-oriented tasks, focus on training that aspect). Ultimately, this could produce models that have undergone “cognitive tuning” – not just optimizing language likelihood, but optimizing performance on each DIKWP stage. This granular approach helps push the ceiling higher because the model is no longer uniformly limited by its weakest link; we shore up each link. And importantly, involving humans in the loop in a structured way (as teachers or evaluators at each layer) injects human insight and values directly into the model’s cognitive development, potentially guiding it to more robust wisdom and purpose handling than it could derive from raw data alone.

4. Semantic Mathematics and Formalism: A more theoretical but potentially groundbreaking avenue is the development of DIKWP semantic mathematics – essentially a formal logical/mathematical system that describes cognitive transformations across the five layers. If such a formalism is integrated into AI, the model could internally verify that its cognitive moves are valid (like a mathematician checking a proof). Early work by Duan et al. suggests that using a formal semantic framework can guarantee certain properties, such as consistency and closure, in an AI’s reasoning ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告). For example, operations like “information → knowledge” can be constrained by algebraic rules to ensure no introduction of implausible statements. This approach would make the AI’s reasoning more akin to a proof or a computation than a fuzzy guess. If an AI can internally prove that its conclusion follows from data via DIKWP-formalized steps, we could trust it much more on critical tasks. Achieving this means blending symbolic AI (logic, formal languages) with sub-symbolic AI (deep learning). While challenging, it could effectively remove the current upper bound by giving the AI tools to extend its cognitive space in a controlled, verifiable manner. It would be able to explore new deductions (thus expanding knowledge and perhaps wisdom) while always remaining within the realm of valid semantics (preventing the collapse or nonsense beyond the limit). In other words, the AI’s cognitive space, instead of being limited by what patterns it has seen, could start to be limited only by what is logically possible given its axioms and inputs – a far more expansive frontier.

5. Expanding Sensory and Context Horizons: Lastly, a practical if straightforward method to push DIKWP capacity is to simply feed the AI more and richer data, increasing its “sensory” breadth. Cognitive relativity tells us that increasing an agent’s sensory input range raises the upper bound of its cognition (认知相对论——通向强人工智能之路). For AI, this means longer context windows (remembering more conversation or documents at once), multi-modality (seeing images, hearing audio, robotics – experiencing the world in more ways), and continual learning (adapting to new data over time). We already see moves in this direction: GPT-4 introduced a longer context and multimodal input; Gemini is built to be multimodal from the ground up. By expanding the Data layer input, we ensure that the AI has more “raw material” from which to derive Information and Knowledge. Pairing this with the above techniques (like purpose-driven retrieval of new data when needed) creates a system that is less likely to hit a knowledge wall – if it doesn’t know something, it can go look it up (or observe it) in real time. In effect, this mimics how humans learn throughout life and gather experiences, thus continually pushing their cognitive development. An AI that can similarly accumulate knowledge and adjust its models (within a DIKWP-guided schema) would not be as strictly bounded by its initial training. Its cognitive space could remain open-ended, always enlarging. The challenge is doing so safely and without losing consistency, which is why the structured approaches above are important in tandem.

In conclusion, while transformers and current AI have notable DIKWP capacity limits today, the roadmap to transcend those limits is becoming clear. By marrying neural models with cognitive theory – explicit knowledge bases, goal-oriented reasoning loops, white-box layer-by-layer oversight, formal semantic checks, and richer sensory input – we can create AI systems that inch closer to the full DIKWP spectrum. Such systems would be able to operate with the reliability of a closed cognitive space (no wild hallucinations) yet the creativity and adaptiveness of an open-ended learner. They would understand context and causality, not just correlations ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告), and pursue goals with deliberation, not just react. This fusion of DIKWP and AI might eventually yield artificial general intelligence (AGI) that demonstrates Data-to-Purpose cognition on par with human-like understanding. Each layer of DIKWP adds a dimension to AI’s capabilities, and overcoming current boundaries means unlocking those dimensions one by one. The ongoing research and experiments are promising – we have already seen LLMs significantly improve with some of these interventions – and it suggests that the apparent “cognitive collapse limit” of AI is not a brick wall, but rather a challenge inviting us to innovate our way through, using the very insights from cognitive relativity and DIKWP theory to guide us.

Sources:

• Li Y. “Theory of Cognitive Relativity — The Road to Strong AI.” Journal of Electronics & Information Technology, 2024, 46(2): 408-427. (Discusses world and symbol relativity, equivalence principle of consciousness, and cognitive limits tied to sensory capacity) (认知相对论——通向强人工智能之路).

• Duan Y. et al. “DIKWP×DIKWP Semantic Mathematics Helping Large Models Break Cognitive Limits – Research Report.” Feb 2025. (Introduces DIKWP semantic closedness, LLM limitations in semantic causality and goal-awareness, and approaches to integrate DIKWP framework with LLMs to overcome these limits) ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告) ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告) ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告) ((PDF) DIKWP×DIKWP 语义数学帮助大型模型突破认知极限研究报告).

• ScienceNet blog – “So-called DIKWP refers to Data-Information-Knowledge-Wisdom-Practice/Purpose layered cognitive framework…” (Definition of DIKWP model and its use in AI cognition research) (科学网—第2次“DeepSeek事件”预测-DIKWP白盒测评).

• Tech Daily (2025): “White-box DIKWP Evaluation of LLMs (100-question report)” – World’s first LLM “cognitive quotient” assessment based on DIKWP, indicating structured analysis of LLM perception, reasoning, wisdom, intent handling ([

     大语言模型意识水平“识商”白盒DIKWP测评2025报告发布   ](https://www.stdaily.com/web/gdxw/2025-02/19/content_298792.html#:~:text=%E6%AD%A4%E6%AC%A1%E6%B5%8B%E8%AF%84%E6%8A%A5%E5%91%8A%E7%9A%84%E6%A0%B8%E5%BF%83%E4%BA%AE%E7%82%B9%E5%9C%A8%E4%BA%8E%E5%85%B6%E5%85%A8%E7%90%83%E9%A6%96%E5%88%9B%E7%9A%84%E6%84%8F%E8%AF%86%E6%B0%B4%E5%B9%B3%E6%B5%8B%E8%AF%84%E4%BD%93%E7%B3%BB%E3%80%82%E6%8A%A5%E5%91%8A%E5%9F%BA%E4%BA%8E%E7%8B%AC%E5%88%9B%E7%9A%84DIKWP%E6%A8%A1%E5%9E%8B%EF%BC%8C%E4%BB%8E%E6%95%B0%E6%8D%AE%E3%80%81%E4%BF%A1%E6%81%AF%E3%80%81%E7%9F%A5%E8%AF%86%E3%80%81%E6%99%BA%E6%85%A7%E3%80%81%E6%84%8F%E5%9B%BE%E7%AD%89%E6%96%B9%E9%9D%A2%EF%BC%8C%E6%9E%84%E5%BB%BA%E4%BA%86%E4%B8%80%E4%B8%AA%E5%85%A8%E9%93%BE%E8%B7%AF%E8%AF%84%E4%BC%B0%E4%BD%93%E7%B3%BB%E3%80%82%E6%B5%8B%E8%AF%95%E9%A2%98%E5%85%A8%E9%9D%A2%20%E8%A6%86%E7%9B%96%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E6%84%9F%E7%9F%A5%E4%B8%8E%E4%BF%A1%E6%81%AF%E5%A4%84%E7%90%86%E3%80%81%E7%9F%A5%E8%AF%86%E6%9E%84%E5%BB%BA%E4%B8%8E%E6%8E%A8%E7%90%86%E3%80%81%E6%99%BA%E6%85%A7%E5%BA%94%E7%94%A8%E4%B8%8E%E9%97%AE%E9%A2%98%E8%A7%A3%E5%86%B3%E3%80%81%E6%84%8F%E5%9B%BE%E8%AF%86%E5%88%AB%E4%B8%8E%E8%B0%83%E6%95%B4%E5%9B%9B%E5%A4%A7%E6%A8%A1%E5%9D%97%EF%BC%8C%E5%AF%B9%E4%B8%BB%E6%B5%81%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E2%80%9C%E6%84%8F%E8%AF%86%E6%B0%B4%E5%B9%B3%E2%80%9D%E8%BF%9B%E8%A1%8C%E4%BA%86%E7%B3%BB%E7%BB%9F%E5%8C%96%E3%80%81%E9%87%8F%E5%8C%96%E7%9A%84%E6%B7%B1%E5%BA%A6%E5%89%96%E6%9E%90%E3%80%82)).

• Nguyen, T. “What Is BERT Language Model? Its Advantages and Applications.” Neurond AI Blog, 2023. (Explains BERT’s capabilities and notes its limitations in reasoning and handling nuance beyond given information) (What Is BERT Language Model? Its Advantages And Applications).

• Wahn et al. “Offloading under cognitive load: Humans are willing to offload parts of a task to an algorithm.” Front. AI (PMC10198496), 2023. (Demonstrates people’s tendency to offload cognitive tasks to AI, highlighting impacts on human cognitive engagement) (Offloading under cognitive load: Humans are willing to offload parts of an attentionally demanding task to an algorithm - PMC).

• MDPI Societies (2025) 15(1):6, “AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking.” (Discusses how over-reliance on AI can reduce practice of critical thinking, create black-box trust, and even reinforce bias via algorithmic echo chambers) (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking) (AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking).

• Tech.co News, Drapkin (Mar 13, 2024). “Google Gemini vs ChatGPT: AI Chatbot Head-to-Head Test.” (Reports Gemini’s performance slightly exceeding GPT-4 and its multimodal abilities, signifying an incremental improvement in capability) (Google Gemini vs ChatGPT: Which AI Chatbot Wins in 2024?).

• Google AI Blog (Dec 2024). “Introducing Gemini: our most capable model” – Pichai, Hassabis et al. (Highlights Gemini’s design for multimodality, and plans to extend its planning, memory, and context length in future versions to enhance its reasoning and goal-directed performance) (Introducing Gemini: Google’s most capable AI model yet).

• ScienceNet blog – “Human-AI Fusion: Exploring DIKWP Model Applications” (2025). (Mentions how DIKWP architecture allows humans to inspect AI reasoning at each layer – data, info, knowledge, wisdom, intent – to improve interpretability and trust) (DIKWP*DIKWP 语义数学帮助大型模型突破认知极限研究报告 - 手机版).

转载本文请联系原作者获取授权，同时请注明本文来自段玉聪科学网博客。

链接地址：https://wap.sciencenet.cn/blog-3429562-1474566.html?mobile=1

收藏

分享