Amazon Cannot Fix Alexa

Mihail Eric spent years inside Alexa AI and wrote up what went wrong, and the headline finding is not that Amazon missed the LLM wave — it's that the organization was structurally incapable of building good conversational AI regardless of what the models were doing.

The failure mode he describes is a particular kind of corporate rot: teams optimizing for metrics that had almost nothing to do with whether Alexa was useful, leadership that couldn't evaluate the work technically, and a "skills" paradigm that was a dead end from the start but had too much internal momentum to kill. Nobody owned the end-to-end experience. Everybody owned a slice of a metric. The user was a rounding error.

You can't fix that with a better model.

Amazon is now making noises about an LLM-powered Alexa, which — given everything Eric just described — reads exactly like what it is: a company that watched OpenAI, Google, and Apple announce AI features at developer conferences and got scared. They have billions of dollars and Anthropic's phone number. That's not the same as having a team that knows how to ship a coherent conversational product.

This is the Grok play — bolt an LLM onto an existing brand, call it a relaunch, generate press coverage. It might work as a narrative. It will not work as a product. The organizational conditions that produced the original failure are still there.

Amazon is very far behind, and the reason they're very far behind is not that they lacked access to transformers.