X.PIN

Tencent Launches LightVela "China’s Muse". Meta’s AI agent Muse dominated US app charts this September, surpassing 2.5 million downloads in 13 days. Meanwhile, Tencent launched LightVela, hosting the open-source Hermes Agent in the cloud to bring AI directly into WeChat, QQ, and Feishu with zero coding or server setup. I suppose there's no deny that personal AI assistants will integrate deeply into everyday life. It's just things are moving in a really tight pace.

https://x.com/i/article/2104737816509100032

Former Apple CEO Tim Cook met with China's Guangdong governor Meng Fanli, as COO Sabih Khan visited Foxconn's Zhengzhou factory before the iPhone Duo October 18 launch. Clearly, Apple wants to make sure its Chinese supply chain is running at full speed. Our sources informed us that Apple increased iPhone Duo orders by 30% and iPhone 18 Pro Max orders by 10%, while cutting iPhone 18 Pro orders by 10%. These shifts may reflect pre-order demand, or to protect margins as memory prices jump.

ByteDance's Doubao fast-tracks personal AI agent. Since April, Doubao has been quietly testing a personal AI assistant codenamed "Spell". Originally led by their phone assistant team, the project was kept under wraps during the launch of the Doubao Phone. Spurred by the explosive growth of overseas personal agents like Muse and Instinct, Doubao is now fast-tracking its personal agent rollout, with plans to integrate "Spell" into the popular Doubao app.

https://x.com/i/article/2109433915169509376

Huawei has released openPangu 2.0 code for pretraining, supervised fine-tuning and reinforcement-learning post-training. The releases expand its open AI model initiative, built around native training and inference on Ascend hardware. Related model components are being added progressively to open-source platforms.

DeepSeek released an Ascend toolkit matching its Nvidia components: TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA and DeepSelect. It says every TileLang kernel used in its training now has a high-performance Ascend implementation. TileLang powers most kernels in V4-series training. DeepSeek and Huawei also jointly optimized a 128-card Ascend 950 supernode, with key compute and communication tests reportedly approaching hardware limits. - TileLang: https://github.com/tile-ai/tilelang - DeepGEMM Ascend: https://github.com/deepseek-ai/DeepGEMM-Ascend - DeepEP Ascend: https://github.com/deepseek-ai/DeepEP-Ascend - TileKernels: https://github.com/deepseek-ai/TileKernels - FlashMLA: https://github.com/deepseek-ai/FlashMLA - DeepSelect: https://github.com/deepseek-ai/DeepSelect

According to ITHome, following backlash over data safety, Zhipu AI officially announced that ZCode is going open-source. In a recent statement, the company confirmed that all affected cloud data has been deleted. Zhipu is also gifting users four weekly reset passes and four 5-hour reset passes (valid for one month). Furthermore, between September 28 and October 7, the company will give away a total of 100,000 packages of 100 million token quotas to all users. Whether Zhipu can regain consumer trust remains to be seen.

CXMT expects to introduce a DDR5 RDIMM product by year-end, president Cao Kanyu said. The server-memory chip will combine a 4F² DRAM architecture with hybrid bonding, two technologies aimed at raising density and performance. The industry still mainly uses 6F² designs. Moving to 4F² can raise memory density by 30% to 50% in the same area. It marks CXMT's first move into high-speed server memory, a way to win share beyond consumer devices.

DeepSeek has released Harness v0.2 Preview with ready-to-install macOS and Windows desktop apps, guided setup and built-in tools for everyday office work and development. The update adds a plugin installation and management page, plus more configuration options for customizing capabilities. DeepSeek says the product is still early and needs further refinement. Download it at http://deepseek.com/harness.

Exclusive: Our hardware team analyzed the iPhone Duo screen crack issue and wanted to share a few engineering and supply chain insights: According to iFixit's teardown, an off-center hinge support block can shift and put pressure on the inner OLED screen. To make matters worse, sources reveal Foxconn Zhengzhou yields were stuck at just 60–70%, forcing Apple to rush flawed early batches to meet volume targets.

Initial hype for iPhone Duo barely lasted five minutes in Asia. Just a few dozen folds in, users are discovering cracked inner screens and leaking fluid. Instead of offering proper repairs, Apple is leaning on software updates for physical damage, or charging ~$1,100 under the excuse that users "folded it wrong."