LLMs Are A Commodity
From here LLMs feel like a commodity. I see little differentiation when comparing the frontier LLMs to one another and the less expensive LLMs to each other.
Customer ROIC will be a function of how they deploy natural language capability within the enterprise. What unique datasets (proprietary and otherwise) may be used in combination to generate an output that creates tangible value for the end customer? Increasingly I believe that smaller, domain-specific models knitted together as part of a domain-specific ensemble model architecture will win the day. I believe those models will largely be open source. It is the proprietary enterprise data used for model training that will be the unique element that will deliver ROIC across the value chain. BloombergGPT is one such example. Solera could do something similar as could the Pharma companies, the Simulation Software companies. Imagination is a gating factor.
I accused the LLM space of being commoditized. Here is some work by Simon Willison as it relates to model pricing:
