2. New model releases differ in the degree to which they cause substitution from existing models or expand the market.
🚨 New working paper 🚨
Demand for LLMs: Descriptive Evidence on
Substitution, Market Expansion, and Multi-Homing
A key question for the business of AI is the extent to which LLMs are differentiated from each other. I use data from OpenRouter to take a first look.
andreyfradkin.com/assets/deman...
A new episode of Justified Posteriors just dropped. We discuss the recent papers of Chad Jones about existential risk and AI safety investment. Check it out!
open.substack.com/pub/empiricr...
Great point from @andreyfradkin.bsky.social: as a movement’s policies become more incoherent and harmful, the quality of “experts” who are willing to publicly defend it decreases precipitously.
Through the use of three case studies of model releases (Sonnet 3.7, Gemini 2.0 Flash, Gemini 2.5 Pro), I document three stylized facts:
1. New models are adopted quickly, with increased demand stabilizing within a few weeks.
3. There is substantial multi-homing, with users of the same app employing a mix of models.
This is still very preliminary work, so comments are welcome!