How easy is it to get the data to start training the LLM? Russell Kaplan, a product leader at Scale AI, posits that “language-aligned datasets are the rate limiter for AI progress in many areas.”

to train other types of LLMs (predicting specific software actions, answering healthcare questions, etc.) we need to figure out a way to generate enough relevant training data.

How strong is the data moat you build & accumulate?

Is there a proof of concept that this LLM application is feasible already (e.g. from a larger company)? And how much will it cost?

if you decide to use the API from a large company like OpenAI to build your application & there are no alternatives, you are subject to their pricing power and product SLAs.

Frequently less sophisticated models might be able to achieve the end result, especially if the LLM is not the core product.

An important question for the LLM applications that don’t own the model themselves is the long-term outcome of LLM infrastructure - will it all be commoditized by many providers offering similar models, or will the most cutting-edge company (with the best engineers, hardware, data, compute, & community) become a gatekeeper?