Embedded AI Podcast E17: Switching Providers - Insulating Yourself from AI Vendor Lock-in

What happens when your AI force multiplier fails? We share our frustrations with Anthropic's recent service issues and discuss strategies for avoiding vendor lock-in with AI providers. Anthropic has had a rough couple of months: capacity crunches causing regular afternoon outages, model degradation they initially denied despite user evidence, and the Hermes.md billing scandal - where simply having a file with that name triggered unexpected charges with no refunds offered. These issues highlight a growing problem: as we rely more on AI for development, we become vulnerable to provider instability. Ryan explains his approach of stepping back to granular tasks that any model can handle, staying in the loop rather than pursuing full automation. Luca explores the challenges of maintaining flexibility across three layers: models, harnesses (Claude Code, GitHub Copilot, etc.), and authored content (skills, MCP servers, prompts). Each presents lock-in risks. We discuss practical strategies: testing different providers, keeping setups migration-ready without parallel systems, and accepting that models have distinct flavors making switching non-trivial. The conversation covers tokenizers, caching strategies, and why even switching between models from the same provider affects costs and performance. Our advice? Pick one provider and get proficient, but prepare your escape route. The current market is too volatile to assume any provider will remain stable long-term. 0:00 Vendor Lock-In With AI Services 0:33 When Your Force Multiplier Fails 1:50 Anthropic's Capacity Crisis Begins 2:42 Model Degradation and Gaslighting Users 5:28 Hermes File Billing Scandal 8:51 Stepping Back to Granular Tasks 12:52 Models Have Different Flavors 17:02 Staying in the Loop 20:09 Three Layers of Complexity 25:25 Tokenizers and Caching Strategies 29:03 Staying Harness Agnostic 32:23 Trying Different AI Vendors 35:38 Pick One But Stay Ready