I'm a founder of a generative AI company and I agree with Benioff - the real power of an LLM comes from the data and the metadata used to describe that data - in fact, LLM's are really, really good at deriving metadata from unstructured or semistructured text. Once you have rich metadata, a world of opportunities opens up for developing platforms and apps, because rich metadata allows you to write code that generalizes well. When you can derive semantic meaning of a message, then the individual schema of a message doesn't matter anymore - you can write code that doesn't depend on specific structure because you can classify the message into a specific set of semantics using an LLM. This is in fact what my company is doing in the observability space (if you're interested in learning more, respond to this and I'll DM you on LinkedIn - my profile on Substack is detached from my business one on LI).
I'll add that the AI interface discussion is an interesting one. We get asked all the time why we don't have a co-pilot or chat like interface in our product. We may add this one day, but right now we've found that those interfaces, while certainly cool, add a whole slew of complexity. We know how to interrogate our LLM so that it performs the broad set of specific tasks we need it to reliably and with minimal hallucination. The minute you start allowing users to prompt your model, you've got all kinds or complexity around prompt engineering, RAG, etc just to nudge the model response in certain directions. I'm not sure this is ultimately what uses are going to want - maybe for some use cases like search or summarization. But for enterprise users, they may very well want the AI to be "hidden" in the sense that the models are just under the hood, doing their thing quietly while making employees work days more efficient.
I'm a founder of a generative AI company and I agree with Benioff - the real power of an LLM comes from the data and the metadata used to describe that data - in fact, LLM's are really, really good at deriving metadata from unstructured or semistructured text. Once you have rich metadata, a world of opportunities opens up for developing platforms and apps, because rich metadata allows you to write code that generalizes well. When you can derive semantic meaning of a message, then the individual schema of a message doesn't matter anymore - you can write code that doesn't depend on specific structure because you can classify the message into a specific set of semantics using an LLM. This is in fact what my company is doing in the observability space (if you're interested in learning more, respond to this and I'll DM you on LinkedIn - my profile on Substack is detached from my business one on LI).
I'll add that the AI interface discussion is an interesting one. We get asked all the time why we don't have a co-pilot or chat like interface in our product. We may add this one day, but right now we've found that those interfaces, while certainly cool, add a whole slew of complexity. We know how to interrogate our LLM so that it performs the broad set of specific tasks we need it to reliably and with minimal hallucination. The minute you start allowing users to prompt your model, you've got all kinds or complexity around prompt engineering, RAG, etc just to nudge the model response in certain directions. I'm not sure this is ultimately what uses are going to want - maybe for some use cases like search or summarization. But for enterprise users, they may very well want the AI to be "hidden" in the sense that the models are just under the hood, doing their thing quietly while making employees work days more efficient.
This is a really good comment and certainly what I see mostly in the market. Changing to prompts from clicks makes everything more complicated.