LessWrong Sequence Posts Discussed in this Episode:
on hiatus, returning soon
The Bayesian Conspiracy 246 – The Void, with Matt Freeman, part 1
Sep 17, 2025
In this engaging discussion, Matt Freeman, a founder of The Guild of the Rose and a frequent commentator on AI, dives into thought-provoking themes from Nostalgebraist’s "The Void." They contemplate the nature of conversations with AI like ChatGPT, questioning who, or what, we’re really engaging with. The conversation also touches on the ethics and consciousness of AI, exploring the emotional resonance of language models versus human experiences, and the philosophical implications of AI interactions. Expect a blend of humor, deep insights, and a sprinkle of anime influence!
AI Snips
Chapters
Books
Transcript
Episode notes
How Base Models Actually Work
- Base models predict next tokens by locating inputs in a high-dimensional latent space and generating continuations from that position.
- That process explains why LLMs excel at inferring context clues without having human-like inner monologues.
From Prompt Tricks To Embedded Personas
- Historically chatbots used long pinned prompts to coerce base models into behaving like assistants.
- RLHF and post-training have since baked those behaviors into model weights, changing how personas persist.
Testing The Essay With An LLM Continuation
- Eneasz fed the essay's opening into an LLM and got a plausible continuation after several tries.
- The experiment illustrates how LLMs can mimic an author's style and produce convincing meta-continuations.