AI innovation begins with trust
summary
AI monetization shifts from massive training to widespread inference: training costs ~10^25–10^26 flops while a typical inference is ~10^14–10^15 flops, a 10^11 training‑to‑inference ratio that means edge usage can pay back training quickly. ARM reports SME2 4–5x inference boost for matrix ops, growing demand for Confidential Compute at the edge to protect model IP and data, and new POE2 lightweight compartmentalization plus virtualized memory tags and FAT prototypes that aim to cut process‑switch overhead while keeping models and pipelines sealed.
AI-written summary · about
chapters (8)
- 0:00AI innovation begins with trust
- 3:07AI investment cycle
- 8:23Training vs inference costs
- 12:00Privacy in edge AI
- 15:35Security in AI systems
- 21:22Confidential compute for AI
- 24:33ARM security tech future
- 33:38Conclusion
related talks
29mEnhancing AI with Edge Computing: A Leap Towards Affordable Performance
04m4 real-world AI deployments at enterprise scale
20mFireworks AI: Go from Hype to High - Value AI
38mAI’s Evolving Landscape: Opportunities Strategy Economics and Sustainable Advantage
21mAhead of the Plasticity Curve: AI, Biology and Technology
34mThe Future of Computing: AI technologies in the exascale era
14mReliable Agentic AI, the Open Way
18mDriving Innovation with AI