The Innovator’s Dilemma: to Build or Not to Build
summary
Midsize companies face a clear choice: adopt off‑the‑shelf LLMs or invest in shaping models with their own data. Evidence cited: a live BBC bird‑watching classifier in production, 23 petabytes and 120 billion tokens available for domain work, and a reported ~30% lift from training custom embeddings. The practical path is domain-adaptive pre-training plus instruct tuning (NVIDIA recipes), supplemented by retrieval-augmented generation (RAG); include ~10% original pretraining data to avoid catastrophic forgetting and constrain outputs for safety.
AI-written summary · about
chapters (13)
- 0:00Innovator’s dilemma intro
- 1:11Tier three AI industries
- 3:38Past machine learning model development
- 5:26Future AI and AGI
- 7:51Current AI capabilities
- 11:41AI impact on midsize firms
- 13:09AI recommendation UX
- 16:28Retrieval augmented generation
- 17:51Domain fine-tuning for niche models
- 21:25Building safe in-house ML tools
- 23:36In-house AI dev challenges
- 24:59Transition to new AI world
- 25:20Open floor for questions and discussion
related talks
49mBen Hall & Barbara Fusinska
31mLeveraging LLMs to build supervised datasets suitable for smaller models
40mLLMs - application and impact on business
33mOpening Keynote: LLMs, between fantasy and reality - the poin...
29mBuilding Scalable and Adaptive Products in a Fast Evolving Landscape
15mLLMs for Data-Driven Decision Making in Product Development
49mPanel Discussion
26mGrappling with the Bitter Lesson