AI is just spreadsheets full of numbers, a tech writer discovers

On a sleepless night, the writer realized he wanted to know what AI actually looked like. After pestering ChatGPT for explanations and searching research papers, he arrived at a deceptively simple conclusion: AI is essentially spreadsheets—huge grids of numbers, typically ranging from -1.0 to +1.0, arranged in a specific order. The next morning, unease returned: he knew what spreadsheets were, but they had never talked to him intelligently before. What made these spreadsheets different? This question launched him on a weeks-long investigation into how large language models actually work. Along the way, he encountered a statement from Anthropic CEO Dario Amodei: "People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work." Yet the writer discovered that understanding is possible—and increasingly, practitioners are learning more about how these systems function. The core revelation is architectural: a modern AI system like ChatGPT is a series of 577 interconnected spreadsheets, or "tensors." GPT-3, released by OpenAI in 2020 and the engine behind ChatGPT's early success, contains exactly 577 of these structures, collectively holding 175 billion numbers. The first spreadsheet is conceptually the simplest: a dictionary. It maps 50,257 words to vectors of 12,288 numbers each. When you type "dog" into ChatGPT, the system retrieves that word's row—a sequence of numbers like 0.06, 0.07, 0.10, 0.12, -0.04 (from the GPT-2 example the author provides). These numbers are not binary code or Unicode; they represent, in numerical form, the semantic and contextual meanings of "dog." The remaining 576 spreadsheets follow a repeating pattern: four square grids (12,288 by 12,288), alternating with wide and tall rectangles. Together, groups of six spreadsheets form what researchers call a "layer." The constant dimension of 12,288 is deliberate—it allows mathematical operations to chain neatly through the system, and it is a multiple of 32, a number favored by the chips that perform AI calculations. In sheer physical terms, the scale is staggering. If printed in 8-point font, GPT-3's first spreadsheet alone would be taller than the Empire State Building. All 577 spreadsheets stitched together as a quilt would cover Central Park five times over. GPT-4, with its rumored two trillion numbers, would require a quilt covering all of Manhattan and the Bronx. The numbers in each cell are immutable once training is complete. OpenAI must spend enormous resources on a formal "training" process to change them—a process that culminates in the release of a new model version. As of the article's writing, the latest is GPT-5.6. The writer's key insight is that despite the "black box" mystique, the structure of AI is knowable and increasingly well-understood. Every conversation flows through the same frozen numbers; only the input changes. The fact that this produces infinitely variable, seemingly intelligent output is what makes the writer's deeper question unavoidable: what exactly do these numbers represent, and how do they work their magic?

The writer's journey reflects a broader gap between how AI systems are publicly described and how they actually function. The phrase "black box" has become a convenient shorthand suggesting that AI's inner workings are unknowable, but the writer demonstrates this is partly a myth—or at least an oversimplification. GPT-3's architecture is well-documented: it is fundamentally a mathematical structure, not magic. The 577 spreadsheets, or tensors, are not randomly assembled; they are the product of a process called "training," after which the numbers are locked in place until the next model is released and trained anew. What makes this explanation valuable is that it shifts the conversation away from jargon ("self-attention," "neural networks") toward a more intuitive framing: numbers in grids. The writer notes that larger models like GPT-4 (rumored to contain two trillion numbers) and GPT-5 are built on the same basic architecture, just vastly scaled up. This insight suggests that understanding AI does not require a PhD—it requires patience, curiosity, and a willingness to ask simple questions that technical professionals might dismiss. The author's observation that we "must" understand AI in order to keep it safe points to a broader argument: demystification is itself a form of accountability.

No comments yet. Be the first to share your thoughts!