In this channel's material, base model means the raw LLM before the tuning passes that turn it into an assistant — before fine-tuning, before RLHF. It is a pure text-continuation engine: a function that predicts the next token given what came before, with no chat persona, no refusal behavior, and no instruction-following instincts baked in [^3e5745ab-279e-5154-9c22-9015dba70a32@1770]. Linus Lee runs his own iMessage chatbot backed by the raw base Llama 2 13B model to show what that means in practice. If you ask a base model a question directly, it does not answer — it may simply keep writing, continuing a blog post around your question, because continuation is all it knows how to do [^3e5745ab-279e-5154-9c22-9015dba70a32@1770]. But with prompting, Lee gets it to talk like “a fairly intelligent and cogent rando on the internet” — and crucially, one with opinions. The canned deflection “I'm an AI, I don't live anywhere” never happens, because no RLHF pass has trained the model to deflect. That personal, unfiltered take is exactly what he wants for a quick vibe [^3e5745ab-279e-5154-9c22-9015dba70a32@1770]. The base-model framing connects to the channel's larger point: everything human-seeming about ChatGPT is packaging on top of a bag of numbers — a statistical model predicting a probability distribution [^3e5745ab-279e-5154-9c22-9015dba70a32@290]. Chat windows, personalities, and read receipts are layers of framing; the base model is the layer underneath, and it preserves a different kind of experience — more like a thought calculator you control than an agent commanding you [^3e5745ab-279e-5154-9c22-9015dba70a32@1241]. - A base model is the raw LLM before fine-tuning or RLHF — a pure text-continuation engine. [^3e5745ab-279e-5154-9c22-9015dba70a32@1770]- Asked a question directly, it keeps writing (a blog post) instead of answering; good prompting makes it talk like an opinionated cogent stranger. [^3e5745ab-279e-5154-9c22-9015dba70a32@1770]- Lee's example: an iMessage chatbot on the raw base Llama 2 13B, chosen because the base model gives personal takes instead of trained deflections. [^3e5745ab-279e-5154-9c22-9015dba70a32@1770]- The assistant behaviors missing from a base model are packaging — framing layers on top of a probability-predicting function. [^3e5745ab-279e-5154-9c22-9015dba70a32@290]