3 ms·
The evidence comes from a knowledge of how LLMs work. By definition, they have no understanding of any of the content they scan or present. They are simply typi
by dtagames 4y ago
The evidence comes from a knowledge of how LLMs work. By definition, they have no understanding of any of the content they scan or present. They are simply typing the next most likely word statistically.
The fact that these compositions look real comes from their statistical similarly to existing writing on the prompted topics. That's not intelligent and it cannot evolve or progress to having understanding or intelligence.
- gizmo 4y agoLLMs aren't intelligent but if they can further --human-- understanding of what intelligence is, then that is sufficient. Take the Turing test for instance. Today, thanks to LLMs, we know that a machine can fool an operator into thinking it's intelligent over text, without actually understanding anything. 10 years ago we would never have believed that. OpenAI is building stuff like Dall-E and GPT-3 is to further their understanding of AI, in order to get closer to AGI. It's the entire point of the thing.
- dtagames 4y agoNo, they cannot further human understanding. They can only read and write existing text faster than you or I can. Something that acts like a parrot and demonstrates no intelligence cannot be used to learn about intelligence. The business model for ChatGPT should be obvious. It's stated at the bottom of every page. They will sell a conversational chat agent for use in customer service and tech support. That's useful and real and meets the Turing test criteria you're talking about.