What allows them to be able to freely switch between LLM models?
Are all LLM speaks the same interface? If so, is that interface standardized?
LLMs are statistical models that take in some big array of numbers and output another big array of numbers, so they’re all similar in that way. However, they translate those numbers to actual stuff (words, images, audiovisual clips, etc) differently - so they’re not speaking the exact same language. Some related terms would be “tokenization” and “embeddings”.
On a higher level of abstraction (and likely the one you’re interested in): LLM interoperability is a huge draw, so many people/companies/FOSS projects have taken the time to make some group of models interchangeable with minimal gruff. They’re hard to make perfectly interchangeable if you’re directly manipulating their nitty gritty parameters during use, but most people let some library handle that part, which obviously makes everything easier from the end developer’s perspective.
You might want to look into “LangChain” and its competitors, that’s the closest to “standards” that I know of. All the big libraries let you swap out all sorts of models, both local OSS ones (eg llama) and remote closed-source ones (eg chatgpt)