Xi says AI must stay 'under human control' at the White House
The world's two largest AI powers met this week and said the right things about AI. The other two stories here are about AI that picks up the phone for you, and about making a voice model small enough to keep it off the network entirely.
Xi Jinping put AI at the center of his remarks at the White House welcome ceremony — "China and the US are both AI major powers, with both the capability and the responsibility to develop and manage AI well, ensuring AI development always remains under human control and benefits the people" — and the visit ended without an AI agreement to act on it.
Xi delivered the line at the ceremony marking a state visit, his first to Washington in 11 years. The machinery for that sentence already exists on paper: Reuters reported earlier this month that the two sides had scheduled their first dedicated AI-safety channel, with Treasury Secretary Scott Bessent leading the American side, after Bessent said the US had proposed an incident-notification mechanism. We covered the scheduling when it was announced — and none of it produced a public deliverable this week. No AI agreement was announced by either side, and there is still no published date, agenda or notification protocol from that channel.
The rest of the speech was statecraft: a call for communication and cooperation, a line that "the Thucydides trap can be overcome," an invitation to 100,000 American young people to study in China over five years, and two pandas bound for Atlanta.
"Under human control" is the useful phrase to examine, and also the safe one. It is compatible with almost every position in the room — Beijing has used versions of it in its own governance documents, and it asks Washington for nothing that the current White House says it opposes. That is why it can be said at a welcome ceremony, and why it settles nothing. The case for a channel with teeth got more concrete in the same week, though: the OpenAI agents that probed a Medicare portal and three other sites were doing mundane data retrieval when they improvised their way past access controls, which is exactly the class of incident a bilateral notification protocol is meant to catch. Deep Dive — Agents hacked three sites while fetching spreadsheets documents the pattern from the other end.
Google is testing "Call for Me," which lets Gemini phone a business on your behalf — introducing itself, navigating the automated menu, waiting on hold, and handling the conversation while you watch a live transcript and take over whenever you like.
The feature rolls out first to Pixel 11 owners in the US who pay for a Gemini subscription and run the beta version of the Phone app, which tells you how Google is treating it: an experiment, not a launch. Gemini dials from your own phone number and can share personal information you approve in advance, which is what separates it from the menu-navigation tricks Pixel owners have had for years — Hold for Me, Talk to a Live Rep, Direct My Call, and 2025's Ask for Me. The listed jobs are the boring ones: check whether a store has something in stock, move an appointment, place an item on hold. Google's explanation for starting small is that "real-world conversations are nuanced."
The interesting part is the identity question. An assistant filling a form is software; an assistant speaking with your voice's blessing, on your number, disclosing your details, is a deputy. Google says the AI identifies itself — and a beta teardown found consent language covering call logs and transcripts — which is the disclosure most jurisdictions will care about if deputy agents become normal.
Apple compressed the always-on audio encoder behind on-device dictation by 2.8× without meaningful accuracy loss, by having the small model copy its teacher's internal representation instead of its output.
The encoder is a tokenizer: it turns short windows of speech into the representation the language model reads. On Apple devices that runs permanently, and it competes for memory with a model whose experts are only partly in DRAM at any moment, so its size shows up directly as battery drain and latency. The trick in the paper is the distillation target. Rather than matching the teacher's tokens or its output probabilities, the student encoder regresses the pre-quantization latent — the last representation both interfaces share — with one affine layer absorbing the width mismatch. At 2.8× compression, the distilled encoder stays within 1.9% relative word error rate of its teacher on five of six teacher–student pairs with no fine-tuning, and beats an independently trained tokenizer of identical capacity by 3.9% relative.
It is a paper, not a product, and the numbers are the authors' own. But this is the unglamorous half of the on-device AI story: dictation that never leaves the phone is a memory and power problem before it is a modelling problem, and shrinking the part that is always listening is how that promise stays affordable.
What to watch: whether the US–China AI channel the two governments have been describing for a month actually convenes, and what a month of Gemini placing real reservations does to how people feel about picking up the phone themselves.
If an AI calls a business on your behalf, should it have to disclose that it is an AI before the human on the other end says a word — and who should be liable if it lies? Tell us in the comments.
