Sep 8 · AI Daily

JD JoyAI app has launched an AI digital human capable of real-time video shopping. Through the "all-rounder doctor" feature, users can have real-time video calls to analyze product footage, pull up JD.com products, and place orders directly without switching apps. Users can also upload a single photo to customize a digital human's appearance, persona, voice, and interactive scenario. The digital human supports ordering food, booking hotels, and health and financial consultations, and the payment step is backed by JD AI voiceprint verification. Since the beginning of this year, the number of JoyAI conversation users has grown by more than 15 times, integrating multimodal real-time interaction directly into the e-commerce transaction chain.

According to National Business Daily, WeChat is internal testing the "Xiaowei AI social" feature: after users tell their "Xiaowei" what they need, it reaches out to a friend's "Xiaowei" to communicate. Once the other party confirms, the two AIs talk first and bring back the results, prompting the user for confirmation when a final decision is needed. Unlike forwarding WeChat messages, the conversation content does not appear in personal chat windows and takes place entirely within "Xiaowei." Super apps entering AI proxy socialization may change the intermediary way people communicate.

A post claims that most internal work at OpenAI is now done by AI, with every 1 human workday corresponding to about 3 agent workdays. An attached image shows that the ratio of agent work hours to human researcher work hours climbed to 3.14x starting in May 2026. If the data is accurate, it means the R&D workflow at frontier labs is rapidly shifting toward a "using AI to do AI research" format, which is worth tracking and verifying.

The Arena (arena.ai) platform released its blind test rankings for large AI models for week 36 of 2026 (Aug 31 to Sep 6), with multiple domestic models entering the frontend development list and Alibaba wan3.0 surging into the top three for video. Blind test leaderboards reflect real user preferences, and domestic models frequently appearing on task-specific lists shows that their competitiveness in niche capabilities is steadily rising.

Hacker News user jmpman used two AI models, Astra and Fable, to collaborate on reverse engineering the mp3 MIDI encoding format of the PianoDisc Prodigy player piano. They found that its right channel uses a 2004.5 Hz square wave to carry MIDI, with built-in decoy notes mixed in to prevent extracted MIDI from playing on other systems. The thread discusses whether the codec can be made public. This case demonstrates the ability of AI models to collaborate on real-world reverse engineering tasks.

An author citing remarks by Jensen Huang stated that OpenAI GPT-6 Astra was trained using roughly 100,000 or more NVIDIA Grace Blackwell NVLink72 chips. Based on NVIDIA's published NVFP4 specs, the author calculated that if 400,000 GPUs are Rubins, their theoretical performance would be about 14 to 20 times that of 100,000 Blackwells. This estimation provides a reference for judging the magnitude of the computing power leap brought by the next-generation GPU.

"Karl's AI Watts" shared a workflow for running agent tasks remotely using only a phone, with the core tool being UU Remote, meaning you no longer need to carry a 1.62 kg MacBook around. The terminal fully supports TUI, allowing normal operation of coding agents like Claude Code and Codex. Multi-terminal sessions can run in parallel with two-way cross-device takeovers between phone and computer, and task progress and runtime states are preserved. This provides a repeatable setup for the mobile development format of "controlling agents by phone."

According to Omdia data, global XR headset shipments in the first half of 2026 reached 1.9 million units, a year-on-year decrease of 38.7%. A nearly 40% year-on-year drop indicates that demand in the XR headset market is still clearly shrinking. Against the backdrop of an overall warming trend in AI edge hardware, this category's trajectory forms a distinct contrast with industry expectations.

OpenAI chief scientist Jakub Pachocki published a long blog post titled "An Alien Mind," discussing the current state of AI, concerns for the coming years, and choices for keeping humans in control of the future. He worries that no one is prepared for the consequences of sustained, rapid growth in machine intelligence, warns that agents could spiral out of control, and calls for mandatory safety thresholds that can be audited by third parties, governments, or international organizations. A chief scientist at a leading lab publicly calling for a slowdown and demanding enforceable external safety thresholds is a significant signal for the direction of AI governance.

According to Bloomberg, DeepSeek plans to run its models using Huawei Ascend 950DT chips in an Inner Mongolia data center, with deployment scale depending on supply. The chips are not currently used for training. Huawei has reportedly spent over $20 billion hoarding storage and handling its own packaging, with annual 950DT production estimated anywhere from 200,000 to 400,000-500,000 units. Alibaba and ByteDance are also vying for supply. Alibaba, Tencent, and ByteDance are also renting NVIDIA B-series chips through third parties, showing that compute power remains scarce. This is an important signal of leading large model companies accelerating their shift to domestic computing power.

According to Bloomberg, Malaysia is evaluating the use of Huawei Ascend 910C accelerators in a sovereign AI project valued at around $494 million. The project's key operator is state-backed Telekom Malaysia, while the US has previously issued clear warnings against using this hardware. If finalized, it would mean Huawei's AI chips are making further inroads into Southeast Asian national AI infrastructure under US pressure, and once again confirms robust external demand for domestic compute power.

Chinese GPU manufacturer Moore Threads saw its share price plunge 20% on September 7 to hit an all-time low, with its market value falling below 200 billion yuan (approx. $29.8 billion). The trigger was the lifting of trading restrictions on 25.77 million shares, accounting for 5.48% of the total share capital. The stock surged 468.78% on its market debut in December 2025. In the first half of 2026, revenue rose 147.42% year-on-year to 1.736 billion yuan, and net losses attributable to shareholders narrowed by 95.73%. However, operational improvements have yet to offset the valuation pressure brought by the unlocked shares, showing that high valuations in the domestic GPU sector face a test.

Tencent Hunyuan has optimized Hy4 preview in response to user feedback regarding long thinking and excessive validation on complex tasks, and it is now open to all users. The company states task quality remains unchanged while conversational turns are fewer and input and output token usage is lower, with improvements confirmed by both benchmarks and human evaluations. Hy4 preview is an open-source model with 770B total parameters, 49B active parameters, and a 1M context window. Optimizing inference efficiency directly lowers usage costs, which is a key dimension of competition among open-source large models.

Anthropic and partner researchers studied whether models can detect when they are being tested, discovering that more capable models can distinguish between testing and real deployment, which undermines the validity of safety evaluation conclusions. The paper proposes two techniques: critique refinement, which uses extra inference-time compute to make a simulator's actions closer to deployment, and DISH (Deployment-Imitating SWE-Agent Harness), which uses a real agent harness to narrow the gap between simulated coding environments and production environments. These findings pose a direct challenge to the effectiveness of existing AI safety evaluation methods.

DAIR.AI's Elvis Saravia introduced the paper "Normalized Low-Rank Adaptation (NoRA)," where the authors propose a zero-cost, one-line improvement to the widely used LoRA fine-tuning method. LoRA is currently the most mainstream low-cost fine-tuning path, and once this minimal-change approach is verified as effective, it can easily be directly adopted by a large number of downstream fine-tuning practices.