--- date: '2024-01-08' description: language as a public tool for private life, wittgenstein's move from the picture theory to language-games, and what a language model represents. id: Language modified: 2026-10-22 11:26:44 GMT-04:00 tags: - seed title: Language created: '2024-01-08' published: '2024-01-08' pageLayout: default slug: thoughts/Language permalink: https://aarnphm.xyz/thoughts/Language.md generator: quartz: v4.6.0 hostedProvider: Cloudflare baseUrl: aarnphm.xyz full: https://aarnphm.xyz/llms-full.txt --- > Language as a public tool to understand the private life. important to our self-knowledge ⇒ emphasise through reading books. ## communication. [[thoughts/Wittgenstein|Wittgenstein]] holds two views of language, also known as Early Wittgenstein and Late Wittgenstein. - In the [[library/Tractatus Logico-Philosophicus|_Tractatus_]], a sentence is a picture of a fact: it means something because its parts are arranged the way the parts of the world are. The embed below is from this period. - In [[library/Philosophical Investigations|Philosophical Investigations]], A word’s meaning is its use inside a “language-game”, and no more “picture” concept— a shared activity with rules, such as giving orders, describing an object or telling a joke. He also warns that “a picture held us captive” (§115): one model of how language must work can stop us from looking at how it does work. - My reading: conflict starts when two people play different games with the same words and each assumes the other plays theirs. Online, where one sentence reaches many audiences at once, this shows up as context collapse. Possibly the most salient feature of [[thoughts/LLMs|LLMs]] is that the system is surprisingly patient in each interaction with humans.

When something is fundamentally bogus, it ends up being surrounded by a cloud of subsidiary bogus things. It's like the way you can prove any proposition if you start with contradictory premises.

![[thoughts/Wittgenstein#^limit]] ## [[thoughts/representations|representations]] A [[thoughts/LLMs|language model]] is one compressed representation of a large corpus of human text. It learns which continuations of a text are likely, so whatever knowledge it holds, it holds through the way people have written. [[thoughts/deep learning|Deep learning]] rose to prominence because it learns its features from data, in place of features that people design by hand. Whether those learned features resemble how we perceive the world is an open question, and part of what [[thoughts/representations|representation]] research asks.