Post #2412740

Replies (1)

• @starsider@valenciapa.ws 2025-07-17 10:54

  @worik@mastodon.social @jadedtwin@corteximplant.com @WhiteCatTamer@mastodon.online @Eatsbluecrayon@rollenspiel.social Thought refers to the mental processes involved in cognition, reasoning, imagination, memory, and planning. It includes both the conscious and subconscious mental activities that allow individuals to interpret, evaluate, and respond to their environment and experiences. LLMs have a very narrow and limited version of this: They don't have imagination, instead they "think" of something and deduce that something else is above or behind or inside, etc. Some multimodal models have something resembling imagination. It doesn't have subconscious activity or inner abstract thought, although recurrent depth models (latent reasoning) kind of resembles abstract thought. Memory is a hack: LLMs don't have recollection of previous conversations. Instead what some systems do to give it "memory" is to store chunked conversations in a vector database, and inject these chunks in the context when some vector seems relevant (when two embeddings have a short distance). LLMs don't have environment and experience. They're fixed in a point of time given by their training and fine-tuning, but only after receiving a staggering amount of "environments" and "experiences" in text form. LLMs are one piece of the puzzle to allow machines to think like a human, but they can't really think to learn, and currently they're extremely limited. By conversing with a LLM you cannot teach it to, for instance, elaborate a mathematical proof. You can instead feed it a lot of mathematical proofs and it becomes better at making them or checking them, but they still fail much more than a human (it's been tried with a fine tune of R1). Because it doesn't come from its experiences. It doesn't come from them realizing their mistakes in one conversation to learn them in another conversation. If they can't learn from experience, in my opinion that's not true thought. It's only part of it. If it was not obvious by now, I'm really interested in how LLMs work and how to make thinking machines that can become individuals. But the current crop of LLMs ain't it. Also OpenAI and other corporations waste way too much energy and spam our servers, for goals that do not align with mine at all. I very much prefer to play with small LMs that run in my computer, without sending my private data to them. That's another issue, they have staggering amounts of private data, and even if your terms of service promise that they won't be used for training (and assuming they keep the promise), your data is still very useful (for example for evaluation and validation of training batches) so they will keep it and they could still be leaked or sold in the future.

  Open ##2412742