LLMs Unplugged: Teaching Resources for a ChatGPT World

TEXT START: Large Language Models (LLMs) are everywhere, yet many learners lack a concrete mental model of how they generate text.

The Dissection

This is an educational translation layer. It reduces LLMs to hand-built n-gram models and weighted sampling, then packages that simplification into workshops meant to make the technology approachable. The pedagogical maneuver is legitimate. Its limitation is structural: it teaches a toy model of generation, not the full machinery driving frontier-model capability or labor displacement.

The Core Fallacy

“Next word generation at scale” is technically true but analytically insufficient. It risks replacing the myth of AI as magic with the equally misleading myth of AI as mere autocomplete. Tokenization, high-dimensional representations, attention, instruction tuning, reinforcement learning, tool use, retrieval, and system integration are not decorative details; they determine what the systems can do.

Under the Discontinuity Thesis, the decisive issue is not whether output is probabilistic. It is whether AI achieves durable cost and performance superiority across cognitive work. A simplified account of generation does not weaken that mechanism. Nor does understanding the machine preserve human productive participation. Literacy is not leverage.

Hidden Assumptions

• A small n-gram model provides a sufficiently transferable mental model of modern LLMs.
• Participants will understand the analogy’s limits rather than mistake it for a complete explanation.
• Demystification alone will produce appropriately calibrated trust and skepticism.
• Self-reported improvement among more than 400 participants demonstrates durable understanding or changed behavior.
• Free educational resources meaningfully democratize power, despite leaving ownership and control of AI capital untouched.
• The central problem is public misunderstanding rather than the concentration of productive capability in machine systems.
• Explaining text generation remains adequate as models become multimodal, agentic, tool-using, and embedded in institutions.

Social Function

Primarily transition management and ideological anesthetic, with a genuine kernel of partial truth. The paper turns a system-level displacement event into an education problem: teach people how the replacement works and they can supposedly navigate the transition. That may improve technical literacy, but it does not address who owns the models, who captures the productivity gains, or what happens when human labor is no longer economically necessary.

The workshops are useful for removing superstition. They are useless as a defense against P1, P2, or P3. The 400-participant figure signals reach and legitimacy, not preserved bargaining power.

The Verdict

Useful primer, structurally toothless. It demystifies the machine while leaving the ownership regime untouched. Under DT logic, this is transition infrastructure: teaching future servitors and displaced workers how the replacement emits language. Understanding an automated labor system does not create a durable human role inside it.