The Day in AI

• OpenAI launched ChatGPT Wallet on 13 November, letting its agents complete purchases for US ChatGPT Pro subscribers after a Face ID or fingerprint approval.
• ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 after about 3.5 years of development, and both projects now follow semantic versioning.
• Google Cloud open-sourced TPU Sync, a C++ library for moving KV caches and weights, and reports over 90% lower KV transfer latency on large MoE models.
• Shi Yunfeng, who ran ByteDance's AI chip and DPU teams, has resigned, according to LatePost, after the first chip fell short of expectations.
• Meta opened Muse outbound phone calls to US businesses to all paying US subscribers on 12 November, after a September test with early adopters.

Until now ChatGPT agents could recommend products and fill carts, and the user still had to pay. With Wallet, a Pro subscriber saves an eligible US-issued Visa, Mastercard or American Express card. Each order is charged with a payment token, so the agent and the merchant never see the card number, and every purchase needs Face ID or a fingerprint. OpenAI says spending limits and pre-approved purchases by category or time period will come later. It also published a Commerce Partner Playbook telling retailers how to take agent orders.

On the open-source side, ggml-org tagged llama.cpp v1.0.0 and ggml v1.0.0, the first major versions of the local inference library and the tensor library under it. The change is a promise about stability. From now on, anything that breaks the public C headers of libllama or libggml gets a major version bump. Minor versions carry new features, backend updates and model support, and patch versions carry bug fixes.

Google Cloud released TPU Sync, an open-source C++ library with Python wheels for PyTorch and JAX. It treats TPU high-bandwidth memory, host DRAM, shared storage and the data center network as one memory hierarchy, so a KV cache (the stored attention state of a conversation) can be moved or tiered without custom plumbing. It plugs into vLLM and SGLang as a KV connector and as the transfer backend when prefill and decode run on separate machines. Google reports over 90% lower KV transfer latency on MoE models of 235B to 480B parameters, measured on TPU v7x.

• Shi Yunfeng, who joined ByteDance in 2018 and led its AI chip and DPU teams, has resigned, according to LatePost. The TPU-style accelerator reached tape-out this year but missed performance targets, and a successor chip is in development.
• Meta made Muse's outbound calls to US businesses such as restaurants and the DMV available to all paying US subscribers on 12 November, with plans from $20 per month.
• Edge0 open-sourced Edge0-145B and Edge0-120B on 12 November and says the larger model runs on a laptop with 29GB peak memory. The company reports 64.6 on DeepSWE 1.1, and its speed figures are only partly confirmed.
• ElevenLabs released ElevenAI in preview on 11 November, an app for meeting notes and dictation in over 90 languages that can hand follow-ups to ElevenAgent.
• Google and Samsung opened preorders for the first Android XR glasses on 12 November, without a display, from $395 in Warby Parker frames and shipping in December.
• OpenAI has disbanded its standalone robotics team and moved the staff into research and Applied, according to Fortune.
• SpaceXSI laid off about 60 Grok Imagine staff in the past week to focus on coding tools, according to The Information.
• Fractile, a British AI chip startup, has reportedly raised $720 million at a $7 billion valuation.