Universal AI: Execute Your Models Where Your Data (And Users) Are
summary
Claim: run ML next to users and data so pipelines stay fast, cheap and compliant. Platform Bakalao (Expanso) runs on WASM, Docker, Wom and arbitrary binaries, supports disconnected networks with declarative GitOps rollouts, unique job hashes and audit trails, and handles heterogeneous devices. Demo shows 301 nodes coordinated, city use case with 15,000 cameras producing 4.4 petabits/day if centralized, and customer calculators claiming ~87% stack cost reduction and 90%+ processing / 93% bandwidth savings.
AI-written summary · about
chapters (13)
- 0:00Universal AI and Bakaya intro
- 1:02Global data explosion challenges
- 2:54Centralizing data slow costly insecure
- 5:40Data pipeline as global network
- 7:47Bakaya flexible data processing
- 11:04ML as data application
- 13:10Distributed infrastructure for ML
- 16:21Runtime challenges deploying ML
- 18:27Live demo Bakaya global jobs
- 25:56ML near users benefits
- 28:22Cost savings and use cases
- 30:11Summary infrastructure vision
- 30:33Q&A hardware and consensus
related talks
36mApplied AI on the edge
25mBuilding AI Workflows: From Local Experiments To Serving Users
43mLLMs in the Wild - Streaming, RAG, and Real Time GenAI at Scale
32mUniversal Basic Compute for Actually Open AI
26mEngineering Excellence: Platforms and Architectures for Real-Time AI Analytics
15mScaling AI-Driven Financial Services: Ethical ML at Scale
55mLarge Scale Distributed AI on the JVM
31mRunning AI workloads with WASM - is it production ready? [Panel]