Post #2412737

2025-07-16 20:55 UTC

@jadedtwin@corteximplant.com @WhiteCatTamer@mastodon.online @Eatsbluecrayon@rollenspiel.social > that's cause it's not an AI, but a really fancy autocorrect. That is a little annoying Go to DeepSeek 's web page, https://chat.deepseek.com/, turn on the "deep think" toggle, enter a query and watch it's internal process laid out explicitly The LLMs and Transformers at the core are implementing a "stochastic parrot" (but what am I. What are you?) but the end result is ( machine) intelligence. That is the best term for it.

Replies (2)

@starsider@valenciapa.ws 2025-07-16 21:30

@worik@mastodon.social @jadedtwin@corteximplant.com @WhiteCatTamer@mastodon.online @Eatsbluecrayon@rollenspiel.social LLMs are not "fancy autocorrect" but they ARE fancy autocomplete. Most of their training is just any kind of text that makes sense, and only on the final stages of training you feed it conversations between "user" and "assistant", to specialize it in following instructions. The illusion of a personality disappears as soon as you're allowed to request the LLM to keep completing the text: after autocompleting the answer of the "assistant" it starts autocompleting the message of the "user". LLMs are impersonation machines, it's just that in most cases they're made to impersonate the "assistant" side. When I have access to the raw text autocompletion of a LLM I have a bit of fun seeing what it autocompletes as the user, or how it autocompletes a conversation that is not a user-assistant conversation. For me it demystifies this magic thinking that people have about LLMs. The "thinking" that some LLMs do is just an extension of a technique called "chain of thought", to make it have more information in the context and to be able to resolve contradictions. It doesn't need to be a "thought" in the same language or even in natural language at all, it works just as well if it appears as random symbols to us. It's just that deepseek trained it to be readable. It's not real thinking. It works better than without it, certainly, but real thinking would involve much more than just generating a bunch of stuff to improve the quality of next generations. Actual thinking involves abstract non-verbal thought as well as being able to learn from experience (even from just one single experience). The only way LLMs learn nowadays is by "training" on a lot of data so it eventually recognizes patterns, and you can't just train it on one conversation to make it learn, it doesn't work like that.

Open ##2412738
@ariaflame@masto.ai 2025-07-17 02:56

@worik@mastodon.social @jadedtwin@corteximplant.com @WhiteCatTamer@mastodon.online @Eatsbluecrayon@rollenspiel.social humans confabulate too.

Open ##2412748