1、1INNOVATION ENDEAVORSState of Foundation Models,2025|Davis Treybig|June 202512Video presentation here STATE OF FOUNDATION MODELS,2025Edited with Capsule 3TABLE OF CONTENTSSetting the StageModelsUse Cases&ApplicationsBuilding Foundation Model ProductsMarket Structure&Dynamics010203040506Whats Next?07

2、What Were Excited to See Built4TLDR;?Generative AI has gone mainstream 1 in 8 workers worldwide now uses AI every month,with 90%of that growth happening in just the last 6 months.AI-native applications are now well into the billions of annual run rate.?Scaling continues across all dimensions All tec

3、hnical metrics for models continue to improve 10 x year-over-year,including cost,intelligence,context windows,and more.The average duration of human task a model can reliably do is doubling every 7 months.?The economics of foundation models areconfusing OpenAI&Anthropic are showing truly unprecedent

4、ed growth,accelerating at$B+of annual revenue.But,end-to-end training costs for frontier models near$500M,and the typical model become obsolete within 3 weeks of launch thanks to competition&open source convergence.?Just like the smartest humans,the smartest AI will“thinks before it speaks”Reasoning

5、 models trained to think before responding likely represent a new scaling law but training them requires significant advances in post-training,including reinforcement learning&reward models.Post-training may become more important than pre-training.?AI has now infiltrated almost all specialist profes

6、sions From engineers and accountants to designers and lawyers,AI copilots and agents are now tackling high-value tasks in virtually all knowledge worker domains?Agents finally work,but we are early in understanding how to build AI products Agents have finally hit the mainstream,but design patterns&s