Scott Johnson

It’s now been 9-10 months since LLMs really entered the larger consciousness of Silicon Valley. We went through the cycle of euphoria and are now catching ourselves for breath. At this juncture, there are many who believe that LLMs are interesting but ultimately are an extension of current mechanisms and products. Their key fallacy is to box LLMs into our understanding of existing technologies. The most mistaken analogy is to classify it as a new information retrieval (read: search engine) system. After all, it responds to queries with answers! This mistaken comparison brings forth comparisons like "indexing", "retrieval", "ranking". The more expansive (and poorly understood) view is to recognize that LLMs are actually reasoning engines. The abilities here are not deeply understood or appreciated yet. We are starting to see developers figure out what is possible by experimenting by throwing a lot of mud at the wall and seeing what sticks. This part isn’t always pretty but a necessary aspect of turning the technology over and over in our hands till we understand what it can do. This is particularly the case with technology like LLMs that demonstrate wild emergent properties. Agents, Planning, Reasoning etc. are all in their infancy but it is abundantly clear that we are still scratching the surface of what we can build with these new capabilities. The more time I spend looking at all this — the more I am deeply convinced that we are just at the beginning of a fundamentally new way of computing. I don’t claim to know where exactly the value will accrue etc. etc. but it’s crazy to me that anyone would believe that this isn’t the best time to build in a long long time.