They have no capacity for thought or introspection. It receives an input and selects tokens that match an expected output. That’s it. Yes it has been fed the entirety of human knowledge to be able to do this. But that’s like a human memorizing the entire mandarin alphabet but never learning mandarin. An Al just makes a sentence that looks right, or generates an image that matches the tokens in a prompt. I could make a mandarin sentence with random symbols, and it might even say something. Doesn’t mean I speak mandarin.
Aren’t we the same though? We don’t really know anything until we learn or train ourselves with knowledge? It’s just that LLMs only do this per prompt.
Yesn’t. What LLMs lack and causes these idiotic suggestions is context from mediums other than language. If you get down to it humans talking are also just trying to predict what should come next, but we predict this from more input than just text. Particularly with something like design, we come up with the design and then the text and maybe concept art to describe it. The llm comes up with a description that seems likely and then feeds it to a diffusion model to make an image to go with the text.
Basically, the problem is input, but AFAIK a lot of AIs now are multi modal. Yes they can make up stuff but if you steer it to the right direction with proper context or inputs, it can help build. Maybe what you’re talking about is AIs can’t choose rather than think. They can think. Albeit, they can’t choose only because of the guardrails. So, you are the user chooses to steer it to what you want
LLMs can be multimodal now. You can chuck a screenshot into claude code or whatever you use and the image generation process can also be done by the LLM itself rather than an external tool call as far as I know
I think AI should know that already because they basically just use all previous human knowledge unless a weak model was used for this
LLM’s don’t know anything.
They have no capacity for thought or introspection. It receives an input and selects tokens that match an expected output. That’s it. Yes it has been fed the entirety of human knowledge to be able to do this. But that’s like a human memorizing the entire mandarin alphabet but never learning mandarin. An Al just makes a sentence that looks right, or generates an image that matches the tokens in a prompt. I could make a mandarin sentence with random symbols, and it might even say something. Doesn’t mean I speak mandarin.
Aren’t we the same though? We don’t really know anything until we learn or train ourselves with knowledge? It’s just that LLMs only do this per prompt.
Yesn’t. What LLMs lack and causes these idiotic suggestions is context from mediums other than language. If you get down to it humans talking are also just trying to predict what should come next, but we predict this from more input than just text. Particularly with something like design, we come up with the design and then the text and maybe concept art to describe it. The llm comes up with a description that seems likely and then feeds it to a diffusion model to make an image to go with the text.
Basically, the problem is input, but AFAIK a lot of AIs now are multi modal. Yes they can make up stuff but if you steer it to the right direction with proper context or inputs, it can help build. Maybe what you’re talking about is AIs can’t choose rather than think. They can think. Albeit, they can’t choose only because of the guardrails. So, you are the user chooses to steer it to what you want
LLMs can be multimodal now. You can chuck a screenshot into claude code or whatever you use and the image generation process can also be done by the LLM itself rather than an external tool call as far as I know