Meta tested Muse calls that humans made, not the AI
Internal Meta posts show the company quietly tested a "human concierge" feature for its Muse agent — real people in call centers placing the phone calls the assistant was supposed to make — then pulled it after employees pushed back. Reuters and 404 Media each saw the internal posts independently. Meta confirmed the test and the rollback, and framed it as feedback-gathering.
The mechanics, per Reuters, which first saw internal company posts on the feature: when Muse enabled calling for half of Meta's employees last week, a "human agent layer" stood behind the AI — trained contractors made the calls and worked them through, with a success rate of 95–98% versus a lower percentage for the AI itself. The backlash was immediate once employees noticed. One warned the company was "one bug away from unnecessary information being leaked to human callers"; another flagged a transcript in which a contractor made a racist remark during a bill-negotiation call. A VP at Meta Superintelligence Labs called launching contractor calls without disclosure "a miss," said the feature was rolled back for now, and said the contractor involved will not work on Meta projects again. A Meta spokesperson said employee feedback was "overwhelmingly positive" and that the feature will ship only "when it's ready and with the proper disclosures."
The tell is the number: Meta's own internal posts put human success at 95–98%, which is an implicit admission of where the AI actually stands. Meta has been here before — a decade ago, media estimates said roughly 70% of Facebook's "M" assistant tasks were done by humans. Muse has 2.5 million downloads per Sensor Tower, so the gap between the agent people think they're using and the pipeline behind it now has real scale. The privacy math is also bad: handing a consumer agent's errands to call-center contractors multiplies the surfaces where personal data can leak, and the disclosure question — did the person on the other end of the line know they weren't talking to the AI? — is the one Meta still hasn't answered.
We've tracked Muse from its sealed-VM launch through its App Store climb and the zero-day found in its launch window. The concierge test suggests the adoption is running ahead of the capability.
What to watch: whether Meta re-enables calling with disclosure, and whether other consumer-agent makers admit to the same human backup.
How would you feel if the "AI" that called the pharmacy for you turned out to be a person in a call center — with your name, account, and errand on a screen?
Sources: Reuters · 404 Media · Cailian Press
Zhipu is now 73% below its June peak. The Hong Kong-listed lab (02513.HK) closed at HK$650 on Tuesday, down 12.4% in its second straight double-digit session — it fell 6.6% Monday after opening higher — and MINIMAX-W slid 5% as Chinese AI model stocks kept sliding. None of the three wires carrying the move gave a cause, but the ZCode data-privacy scandal remains the live overhang, and the stock has now retraced almost everything since its June high of HK$2,410. Call it derating, not panic: turnover was roughly 40% above average, and the slide is day-two continuation of a selloff that began before the scandal broke.
Hugging Face wants your GGUFs inside transformers, not just llama.cpp. The Hugging Face team announced native GGUF support in transformers — the quants load through the regular transformers API, and on Apple Silicon a new packed-inference path reuses ggml kernels to run directly from the quantized weights, hitting roughly llama.cpp speed on the Qwen checkpoints they tested (Qwen3.5-4B Q4_K_M reached 70.4 tok/s versus 71.8 for llama.cpp on an M2 Max). The fine print: the fast path is Apple-Silicon-only and Qwen-only for now, it's on the main branch until the next release, and Hugging Face itself says llama.cpp remains the recommendation when raw local inference speed is the goal. The real value is workflow — quantized checkpoints in the same environment you debug, evaluate, and fine-tune in — and the Reddit thread shows developers already asking whether image tooling gets it for free next.
