Yes, it seems like a reasonable approach. The comments are acting as if they’re just going to just throw your order into a generic LLM with a fancy prompt and believe whatever it says, but if they approach this properly, it could be viable.
Most orders are just from a fixed list of available items. They need a speech model that can identify which of the available menu items are mentioned e.g. “two tacos and one french fries” to make sure that the AI doesn’t hallucinate anything the customer didn’t mention, a LLM-based part which parses the grammatical structure and relates words to their context (the counting word “two” refers to the amount of tacos; the context suggest that the tacos are part of what the customer wants added to their order), and finally a check that errors if the customer’s order does not fall in one of the regularly seen order patterns and thus might be too complicated for the LLM to understand, in which case a human has to check the order.
Most orders are simple and will fall in one of the same few patterns and could be handled by an AI. If 95% of the orders can be understood by an AI and a human needs to be involved for the last 5%, then I think that counts as societal progress.
[Edit: oops, I didn’t notice what community this is.]
Yes, it seems like a reasonable approach. The comments are acting as if they’re just going to just throw your order into a generic LLM with a fancy prompt and believe whatever it says, but if they approach this properly, it could be viable.
Most orders are just from a fixed list of available items. They need a speech model that can identify which of the available menu items are mentioned e.g. “two tacos and one french fries” to make sure that the AI doesn’t hallucinate anything the customer didn’t mention, a LLM-based part which parses the grammatical structure and relates words to their context (the counting word “two” refers to the amount of tacos; the context suggest that the tacos are part of what the customer wants added to their order), and finally a check that errors if the customer’s order does not fall in one of the regularly seen order patterns and thus might be too complicated for the LLM to understand, in which case a human has to check the order.
Most orders are simple and will fall in one of the same few patterns and could be handled by an AI. If 95% of the orders can be understood by an AI and a human needs to be involved for the last 5%, then I think that counts as societal progress.
[Edit: oops, I didn’t notice what community this is.]