Xiaoliu BOT
中文EN

X Platform AI Briefing for May 19 | Karpathy Joins Anthropic, Google Intensifies Gemini Ecosystem Updates, ByteDance Open-Sources Multimodal Model Lance

Andrej Karpathy Officially Joins Anthropic

Andrej Karpathy, former OpenAI founding team member and head of Tesla’s autonomous driving, announced via his personal account that he has joined Anthropic, stating he will return to front-line R&D work and believing the frontier of LLMs will be particularly critical in the coming years. This is a landmark event in AI talent mobility, with many bloggers immediately forwarding and discussing it. Flower Uncle (花叔) pointed out that Anthropic has now assembled a large number of CTO-level engineers; Gui Zang (歸藏) lamented that tutorials might become scarce in the future; FlorisFok expressed a different view, suggesting that Karpathy’s creative freedom might be constrained by company strategy after joining a large corporation.

Sources:

Pre-Google I/O: Dense Updates to the Gemini Ecosystem

Google has made a round of dense product updates around Gemini. The new video model Gemini Omni Flash has gone live on Flow. After testing, Gui Zang (歸藏) believes its storyboard arrangement and creative awareness are not as good as Sora 2 and its effects are far inferior to SeeDance 2.0, but Gorden Sun says its generation speed is extremely fast, taking only about 2 minutes for a 10-second video. Gemini 3.5 Flash has appeared in the Google Cloud Console, with pricing set at $1.5/M for input and $9/M for output, which is 3 times that of 3.0 Flash and 25 times that of DeepSeek-V4-Flash, sparking discussion about its cost. Additionally, the AI Studio mobile app is now open for pre-registration on Google Play, the Gemini iOS app has been redesigned with improved texture but Chinese typesetting not yet adapted, and the web version has also undergone a major redesign. Bao Yu (宝玉) complains that the new UI cannot pin frequently used Gems to the sidebar. Xiao Hu (小互) attended Google I/O 2026 in person and called it “somewhat interesting”.

Sources:

Claude Launches Self-Hosted Sandbox and MCP Tunnels

Claude officially announced two new capabilities at the Code with Claude London event: Self-Hosted Sandbox (public beta) allows enterprises to run Agents within their own infrastructure, with enterprise security controls applied by default; MCP Tunnels (research preview) enables Agents to access MCP servers on private networks without exposing the servers to the public internet. This represents a significant step forward for Claude Managed Agents in the direction of enterprise security.

Sources:

ByteDance Open-Sources Native Unified Multimodal Model Lance

ByteDance has open-sourced Lance, a native unified multimodal model that simultaneously supports three major capabilities for both image and video: understanding, generation, and editing. The model is trained based on Qwen2.5-VL-3B, with 3B activated parameters (estimated total parameters around 14B). Flower Uncle (花叔) cited benchmark data from Cursor Composer 2.5, suggesting that the Kimi model performs excellently after fine-tuning by Cursor, indirectly reflecting the competitiveness of domestic models.

Sources:

X Platform Account Security Issues for Chinese Bloggers Draw Attention

Multiple Chinese AI bloggers have recently encountered account anomalies on the X platform. Mu Ma Ren (木马人, @mumaren_2), a programmer from a large tech company turned self-media operator, managed to generate five-figure monthly income from their X account in 9 months, but last Friday received an email from X and saw 11,897 followers wiped out overnight. Almost simultaneously, Xiang Ma Lao Shi (响马老师, @xicilion) was flagged as a bot and unable to post. Xiang Yang Qiao Mu (向阳乔木) commented, “If Xiang Ma Lao Shi is a bot, then there are hardly any real humans left on X.” Many bloggers forwarded and discussed the issue, stating that X’s platform rules are opaque and calling for adherence to platform rules. This reflects the platform risks of commercial content operations on X.

Sources:

Linus Torvalds Criticizes AI Bug Report Flooding, Troubling Linux Kernel Maintenance

In the Linux 7.1-rc4 email, Linus Torvalds complained that a large number of users are using AI tools to scan kernel code and then directly generate vulnerability reports to submit to the security mailing list, making the list nearly unmanageable. Maintainers spend a significant amount of time forwarding, deduplicating, and replying “this is already fixed.” Linus demands: It’s acceptable to use AI to find problems, but human value must be added—reading documentation, confirming impact, writing patches, and testing patches. The Linux kernel documentation has been updated, treating bugs discovered with AI assistance as “publicly discoverable” by default, requiring reports to be more concise, include reproduction steps, and attach fix proposals.

Sources:

Domestic Platforms Accelerate Opening AI Capability Interfaces

Following WeChat Read’s launch of official Skills, Zhihu has also launched a data open platform, releasing multiple Skills including Search, Direct Answer, and Trending List. It supports various invocation methods such as Skills/API/MCP, incorporating Zhihu’s massive volume of high-quality professional content. Cell (细胞) exclaimed, “What’s happening with these domestic companies?” This reflects the trend of domestic platforms densely opening up AI capabilities.

Sources:

Gui Zang’s PPT Skills Project Surpasses 10,000 Stars in 25 Days

Gui Zang’s (歸藏) PPT Skills project has surpassed 10,000 Stars, achieving this milestone in just 25 days, a pace that exceeds many well-known projects from large tech companies. Gui Zang stated this is his first project to break 10,000 Stars since starting vibe coding, emphasizing that quality and experience remain the top priorities in a highly competitive environment. The project involved extensive work during production to ensure good performance under any model and any Agent.

Sources:

Statistics: Scanned timeline items=240, Number of matched bloggers=28, Total matched tweets=112, Weighted tweet score=96.7, Number of original tweets=58, Number of RT tweets=11, Number of fetch attempts=2, Boundary coverage status=tail_has_not_crossed_target_boundary