The Session You Cannot Take With You
Summary:
A sharp look at a lock-in mechanism most teams have not priced in yet. As inference APIs add encrypted reasoning tokens, server-side web search, compressed context, and provider-held storage, the transcript on your machine stops being the whole session: it becomes a partial view of state that lives with the provider, and you cannot replay it against a different model. The essay's central line lands hard: the operational state of your session belongs to an inference provider and not you. It closes with seven principles for portable inference APIs, favoring complete transcripts and user-held state over convenience features that quietly deepen the moat. As builders of LLM applications who route across providers, we read this as a checklist worth keeping open in a tab.
Excerpt:
"Earendil argues that inference APIs are quietly becoming lock-in machines: opaque reasoning tokens, server-side tool state, and provider-held context make your session something you can no longer take to another model."
#AI#LLMs#Open Web
Read Full Source