Show HN: Building an LLM from Scratch in TypeScript

URL SCAN: Show HN: Building an LLM from Scratch in TypeScript
FIRST LINE: No black boxes

The Dissection

This is a pedagogical decomposition of an LLM: dataset, dataloader, attention, transformer blocks, GPT, and training loop, implemented in Node.js and TypeScript. Its real product is not a competitive model. It is the feeling of technical demystification—turning an industrial-scale system into runnable chapters small enough for one person to inspect.

The Core Fallacy

The implied fallacy is that understanding the machinery confers strategic power over it. It does not. Reimplementing attention and a training loop teaches mechanism; it does not provide frontier-scale compute, proprietary data, distribution, capital, or control over AI infrastructure.

“No black boxes” is also theatrically absolute. The project necessarily rests on black boxes below its chosen layer: hardware, kernels, compilers, numerical libraries, operating systems, data provenance, and the emergent behavior of optimization. It removes selected abstractions for educational purposes, not opacity from the system.

Under the Discontinuity Thesis, this distinction is fatal. Cognitive automation is driven by scalable performance and cost, not by how many individuals understand transformer internals. A runnable TypeScript implementation is a laboratory specimen, not ownership of the factory.

Hidden Assumptions

• That conceptual literacy will remain economically scarce after AI systems can explain, generate, and debug the same concepts on demand.
• That building a small model transfers meaningfully to competing with industrial models trained on vastly greater compute and data.
• That the relevant bottleneck is understanding algorithms rather than owning compute, energy, data, deployment channels, and capital.
• That human curiosity and craftsmanship will translate into productive participation even as AI makes implementation itself cheaper.
• That TypeScript accessibility creates durable leverage rather than merely widening access to a tutorial whose subject is being automated.
• That the current educational market rewards knowledge acquisition more than demonstrated control of AI capital or indispensability to its owners.

Social Function

Primarily partial truth, prestige signaling, and ideological anesthetic.

The partial truth is real: inspecting the components can produce durable technical understanding, and that understanding may help a person become a Servitor in deployment, evaluation, infrastructure, or maintenance. The prestige signal is the identity of being someone who understands “how it really works.” The anesthetic is the suggestion that individual mastery of internals is a sufficient response to a system in which productive leverage is concentrating around compute, capital, and ownership.

The guide is useful as a map. It becomes copium when mistaken for a moat.

The Verdict

This project is not obsolete as education; it is obsolete as a presumed route to economic sovereignty. It teaches the anatomy of the machine while leaving the decisive assets—compute, energy, data, distribution, and ownership—outside the learner’s hands. Its surviving value is narrow but genuine: verification, debugging, research literacy, and transition intermediation. Anyone expecting a career moat from “building an LLM from scratch” is studying the engine room of a ship they do not own, while the owners are replacing the crew.