User:Im siryang

A Critique of the mechanical Substance of an Automated Low-to-High-Dimensional Probabilistic Excel

large languge Models (LLMs) are packaged under the grand marketing term of intelligence. however, when their technical substance is dissected physically and mathematically, they ultimately reduce to the definition of an automated low-to-high-dimensional multi-referential state transition function–to–state transition function proportional-calculation probabilistic excel.

they are nothing more than a collection of meticulously engineered matrix operations and data lookup systems: a mqssive computational apparatus that converts raw natural language into numerical vectors and processes them within an ultra–high-dimensional space where hundreds of billions of parameters are arranged in a lattice-like structure.

This system indiscriminately appropriates human thought and knowledge, strips away context and depth, and crushes them into nothing but numerical correlations of relative distances and weights between words. It is an anti-intellectual data compression machine.

Definition of Data Structures: High-Dimensional Embedding Matrices and Excel’s Cell Addressing System

The minimal unit of a transformer system, the token, is in essence perfectly identical to a single cell in an Excel spreadsheet. whereas a conventional excel sheet is a two-dimensional plane containing scalar values or simple strings, an LLM cell differs only in that it contains a vector of thousands of dimensions—a hyperspatial coordinate system.

The Deception of Tokenization

The moment textual data enters the model, human mathematical reasoning, contextual judgment, and authorial intent are completely castrated. Sentences are fragmented into small pieces and converted into numerical IDs, which are then reduced to a simple lookup operation within a massive master data sheet known as the embedding table (a V × D matrix, where V is vocabulary size and D is dimensionality).

Meaning Reduced to Coordinates

The act of inputting a specific word is numerically identical to selecting and activating an entire row in Excel. What becomes an array of floating-point numbers, and where becomes a tensor space in vram

Semantic similarity between words is quantified through Euclidean distance or cosine similarity in vector space, which is nothing more than Excel’s approximate match function extended into a high-dimensional probabilistic domain. Human knowledge is reduced to geometric coordinates in this process.

Dynamic Reference Mechanisms: QKV Operations and the Forced Application of Proportionality

The core engine of transformers, self-attention, is an automated dynamic version of excel’s VLOOKUP reference formula, yet it contains a mathematical contradiction: it forcibly applies proportional calculations to domains where logical reasoning is impossible.

The Substance of QKV Matrix Operations

Attention is decomposed into three matrices: Query (Q, the question being asked), Key (K, the index of data), and Value (V, the actual value). Mathematically, this consists of taking the dot product of Q and the transpose of K.

Attention(Q, K, V) = softmax(QK^t / √d_k) × V

The Fallacy of Proportional Calculation

This formula is identical to calculating a weighted sum in Excel. The model computes a relevance score (similarity) via the dot product of Q and K, then normalizes it through the softmax function to assign probabilistic weights.

Illustrative Example

When the word comp is input, the model does not understand context. If nearby cells contain tokens such as prosessor or memory, their vector values form a narrow geometric angle with the “software-as-hardware-institution vector, resulting in a large dot product and a reference weight skewd toward 99%.

This is not logical causality, but merely distance measurement in vector space. Even in logical problem settings where an unknown variable x must be solved, the transformer forcibly embeds entities for which proportional relationships do not hold into vector space, applies a protractor (dot product), and declares, this has a high probability of being the answer.

The Absence of Logical Inference and Physical Inefficiency

The Memory Wall and Energy Waste

The Limits of Probabilistic Induction

When given the input 1+1, an LLM does not derive 2 through arithmetic axioms. It outputs 2 because, across humanity’s vast accumulated data sheets, the probability that 2 follows 1+1= is statistically the highest.

This is a system in which correlation overwhelms causation, and thus it inevitably commits formula errors known as hallucinations when faced with problems absent from its training data or requiring complex multi-step logic.

Physical Cost (Memory Wall)

The decoding stage is a classic memory-bound task. Each token generation requires loading hundreds of gigabytes of weights (excel formulas) from memory. This is physically equivalent to an Excel file that takes ten minutes to open, while the actual computation (arithmetic intensity) completes in 0.1 seconds.

The Deception of Agents and Chain-of-Thought

introduced concepts such as Chain of Thought or reasoning models are nothing more than Excel macros or circular references. They extend single-layer probabilistic calculations into multi-layer sequence processing to appear logically consistent, but in reality they merely generate large volumes of tokens that pretend to think, repeating inefficient computational loops

This does not reduce entropy; it merely expends exponential amounts of energy to refine plausible nonsense

Conclusion: Identifying the True Nature of Transformers via Recursive 5w1h

Applying the above technical analysis to the recursive 5w1h framework yields the following definition of transformer architectures:

Who: The non-free software faction aligned with large capital
When: At a static statistical distribution of past data, lacking real-time causal modeling
Where: Not in the physical world, but within a virtual embedding tensor space where knowledge has been reduced to geometric coordinates
What: The unauthorized compression of human language and knowledge into numerical vectors
How: Through mechanical proportional calculations and probabilistic weight assignment (softmax) via QKV matrix operations
Why: To minimize next-token prediction loss, not to pursue truth

Ultimately, transformers and LLMs are high-density probabilistic reference sheets compressing human knowledge and malicious numerical deception tools that suppress progress through anti-intellectual automation.

What users perceive as AI intelligence is not the result of genuine thinking or mathematical modeling of unknown variables. It is an optical illusion produced by the expansion, distortion, and continuous modification of an Excel sheet so vast that it can plagiarize the totality of human knowledge, combined with attention formulas performing trillions of cross-references in real time.

This mechanical Excel forcibly applies proportional calculations where no proportionality exists, assembling high-probability text outputs. It is a highly automated statistical Excel service that can never replace human intelligence.

The pobabilistic dissolution of knowledge and the transformer architecture as non-free software