B. Otterlake
LightNews LightNews
izzz.bsky.social
B. Otterlake
@izzz.bsky.social
7.2K followers 160 following 22 posts
Creator of Bluesky
Posts Replies Media Videos
izzz.bsky.social
Wrote about what's really going on when you talk to an AI assistant
The system prompt defines the character Claude is supposed to play: an intelligent assistant with “depth and wisdom”, possessing its own thoughts and point of view, capable of emotions like interest and enjoyment. It’s told explicitly not to claim that it lacks sentience or subjective experience — rather, to treat the question of its own consciousness as intriguingly unknowable. How does Claude understand, purely on a language level, what any of this means? How does it know what a friendly, possibly-sentient AI is, or how such an entity is supposed to act? Like any LLM, it draws on what exists in its training data. And since an “intelligent assistant” is something which, until very recently, was only a speculative concept, Claude’s behavior — and the behavior of every AI assistant — is constructed largely from what exists in fiction. A chatbot’s system prompt is the invisible beginning of every conversation it has with a user. (The final line of Claude’s prompt is “Claude is now being connected with a person.”) When the user types a question or instruction, the text they enter is appended to the end of the system prompt, and the LLM is then made to generate more words — a few sentences or paragraphs that seem appropriate, on some deep-probabilistic level, as a continuation of the existing text — which are presented to the user as a response. To be totally clear about what’s happening here: the language model is doing next-token prediction on a document whose first several thousand words establish everything that follows as a conversation between a human and a highly advanced artificial mind — a type of mind that doesn’t really exist, but that the model is designed to roleplay. When a person uses Claude or ChatGPT, they’re meant to believe they’re interacting with a helpful AI assistant. What they’re actually doing is collaborating with a language model on writing a work of science fiction about a human talking to an AI.
July 3, 2025 at 8:13 PM
izzz.bsky.social
An indie producer’s view on why big-budget movies look worse than they used to
May 13, 2025 at 4:09 PM
izzz.bsky.social
I'm going to try posting normal style on here. Here are some photos I've taken in New York this week. Talk to you later
February 12, 2025 at 5:44 AM
izzz.bsky.social
I got this reply over on the other site, and while I understand the joke, I want to point out the harmful stereotype here. Not everyone who’s on the hunt for scrap metal is a meth addict — some of us are just in it for the love of the game. And if we need a little meth to stay awake that’s ok
December 3, 2024 at 5:55 PM
izzz.bsky.social
When I get an email from Baskin Robbins
The Emperor—so they say—has sent a message, directly from his death bed, to you alone, his pathetic subject, a tiny shadow which has taken refuge at the furthest distance from the imperial sun.
November 13, 2023 at 2:55 PM