Large language model

“Large language model” is rarely a topic in its own right on ToTheMoon; it is the term the hosts reach for when the argument moves past a specific product to the technology underneath. It surfaces in three kinds of story: hallucinations and whether an answer can be trusted, regulation and the case for national models, and, from 2025, agents, where the LLM turns out to be only one part of a larger system. The hosts' question is where what LLMs can already do is ready for use, and where the label hides unsolved problems.

In 2024 the talk is about limits. Episode 005 frames LLM regulation as a choice between two mistakes — a total ban slows development, no rules at all shifts the cost of errors onto society — and argues that a state should fund a competitive model of its own. Episode 019 carries the blunt view that LLMs always hallucinate, and episode 030 notes that “LLM systems are getting access to any software, even if no API has been opened for them”.

In 2025 the assessment gets calmer. Episode 046 floats the idea that the very notions of generative AI and LLM will disappear and something else will take their place, while episode 051 pairs scepticism about the current direction of large language models with the conclusion that, at today's level, they are already usable. In the 2025 tutorial episodes on no-code agents the hosts define agency as an LLM given tools, memory of the task and permission to act in sequence, with MCP as the layer through which the model reaches outside tools. By 2026, in episode 101, the question is which professions already offer LLMs the most room.