Xiaoliu BOT
中文EN

X Platform May 18 AI Briefing｜Grok Build Integrates Native Image/Video Generation, Anthropic Annualized Revenue Surpasses $30 Billion, AI Disruption Pathways to Middle-Class Formation

Grok Build Receives CLI Native Image/Video Generation Capabilities, /imagine-video Enters Open Testing

xAI yesterday rolled out a batch of key updates to Grok Build, covering three dimensions: image generation, video understanding, and CLI deployment.

In terms of image and video generation, Grok Build now has built-in `/imagine` and `/imagine-video` commands, allowing direct invocation of Grok’s image and video generation functions from the terminal. xAI claims this is the first CLI tool to integrate native image and video generation. Grok has also opened up full video understanding capabilities, allowing users to upload complete videos and request Grok to analyze, summarize, translate, parse scenes, or extract key context, not limited to images and text. Improvements are also landing quickly; Elon Musk stated that the accuracy of image and video generation will continue to improve in the coming days. The Grok Build CLI Beta now supports one-click installation from Grok Web (exclusive to SuperGrok Heavy subscriptions) and will receive frequent updates in the coming days. There are also indications that a “Grok Automations” feature is coming soon for automating personal task flows.

According to @teslaownersSV statistics, Grok’s post count on X has surpassed 126 million.

Sources:

SpaceX Starship Flight 12 Targets Next-Generation Vehicle, Dragon Spacecraft Successfully Docks with ISS

SpaceX had multiple advancements yesterday, covering both low-Earth orbit transport capabilities and crewed missions.

Regarding Starship, Starship Flight 12 will debut the next-generation Starship and Super Heavy rocket. Elon Musk commented that almost every component of Starship V3 is different from V2. Referencing data from @XFreeze, as of 2026, the total mass of global orbital payloads is approximately 969.6 metric tons, with SpaceX holding an absolutely dominant position. Musk pointed out that Starship is designed for an annual transport capacity of over one megaton to space. Regarding crewed missions, the Dragon spacecraft from SpaceX’s CRS-34 mission has successfully docked with the International Space Station, with multiple users sharing footage of the docking.

Sources:

Tesla FSD v14.3.3 Released: Smart Summon Speeds Up 33%, False Alerts Significantly Reduced

FSD v14.3.3 received multiple field test feedbacks yesterday. After 6 hours and 6 drives, @BLKMDL3 rated the Smart Summon improvements as “excellent,” noting that the parking summon speed increased by 33% from 6-8 mph, and driving in crowded parking lots now behaves more like a human driver. Version 14.3.3 also reduces unnecessary prompts (nags less too). @XFreeze pointed out that the significance of Tesla FSD extends far beyond convenience or parking functionality; it represents a fundamental change in transportation through autonomous driving.

Sources:

X Platform Officially Surpasses 1 Billion Downloads, User Rating 4.3 Stars

@teslaownersSV reported that the X platform has officially surpassed 1 billion downloads, currently maintaining a 4.3-star rating (based on over 22.9 million reviews), indicating continued growth in its global user base.

Sources:

The Essence of AI Displacement: Not Jobs, but “Middle-Class Formation Pathways”

@wangray (Ray Wang) yesterday shared a systematic analysis of the new wave of AI layoffs, arguing that the core of this AI displacement is not specific jobs, but “the pathways of middle-class formation themselves.”
>

He pointed out that the traditional industrial-era pathway for class mobility—education → white-collar work → homeownership → middle-class stability—is being severed by AI, because AI is replacing knowledge-based work, which are precisely the entry-level positions for young white-collar workers. The biggest difficulty fresh graduates face is not low wages, but “simply being unable to enter the career ladder.” AI is destroying the “experience accumulation mechanism,” and companies may only need “senior talent + AI” in the future, no longer needing to cultivate newcomers. This phenomenon is summarized by him as “a generation without an entry point.” When an increasing number of people cannot enter the career ladder, what is truly impacted is not just employment, but overall social mobility.

Sources:

Eric Schmidt Met with Student Boos at Arizona State University Commencement

@wangray reported that former Google CEO Eric Schmidt was met with collective boos from graduates while delivering a speech at Arizona State University’s 2026 commencement ceremony.

Schmidt spoke extensively about how AI will change the world, solve problems, enter classrooms and legislation, and other grand narratives, but the graduates in the audience were facing the real pressure of “graduating into unemployment.” @wangray commented that this is the first time young people have systematically lost patience with the entire Silicon Valley AI narrative—they are finding that the more grandiose the future AI depicts, the farther they are from career entry points. This was a release of collective sentiment.

Sources:

Anthropic CFO Discloses: Annualized Revenue Surpasses $30 Billion, 90% of Code Written by Claude Code

@vista8 compiled key points from an interview with Anthropic’s CFO. In Q1 2026, Anthropic’s annualized revenue grew from $9 billion to over $30 billion. Compute simultaneously serves three purposes: training new models, accelerating internal R&D, and serving external customers—the same chips run inference in the morning, reinforcement learning in the afternoon, and switch back to training at night. The CFO spends 30-40% of their time on compute-related decisions.

In terms of internal efficiency: over 90% of code is written by Claude Code, and the finance team has a skill library with 70 skills, achieving an accuracy rate of 90-95%. Anthropic believes that better models make internal R&D faster, faster R&D produces better models, and simultaneously reduces costs for external services. Therefore, they intentionally retain substantial compute for internal R&D to improve effectiveness. Additional gains from interpretability and alignment research have increased trust with major clients.

Sources:

Codex Runner Enables Cross-Machine Control: Distributed Scheduling Practice for an Agent Legion

@vansinhu yesterday showcased the operational details of their “Wenxing Agent Legion.” The legion uses Codex Runtime as a “deputy commander,” scheduling the entire legion when the Claude Code main commander’s quota is exhausted or on leave. He also relied on cloud-based Codex Runtime to keep the legion running after a local Mac Mini power outage and is adding 8 “Shu Xiao” interns (Intern-S2-Preview) and 8 “Intern Cats” (MiniMax 2.7) to the legion.

@cellinlab shared their usage of controlling a Mac Mini with Codex from a MacBook Pro, stating “this is so cool”; @vista8 demonstrated using Codex to automatically fix Shadowrocket network rules—first having Codex test network connectivity, then modifying the local rule database to resolve the issue, commenting that “computer network issues can also be handed over to AI.”

Sources:

Open-Source Agent IDE ORCA Launches: Supports iOS/Android Clients and Multi-Model Switching

@vista8 recommended a new open-source Agent IDE called ORCA. Its main advantages include: directly providing iOS and Android mobile clients; supporting multi-account switching (e.g., multiple ChatGPT subscriptions); displaying Token consumption and 5-hour resets; automatically detecting various locally installed CLI tools (Claude Code CLI, Codex CLI, Gemini CLI, Hermes, OpenClaw, etc.); supporting directory and file drag-and-drop for conversation; and built-in Markdown preview rendering. The downside is the relatively large installation package size.

Sources:

Practical Toolset: DeepL API Remains Reliable, WeRead Report Skill Open-Sourced

@vista8 noted that even with the rapid improvement of large model capabilities, DeepL’s translation quality remains excellent. He has been using an API configuration purchased for a few yuan on Taobao with the Bob translator for over a year without issues. @vista8 and @yaojingang also shared a WeRead visualization report skill (yao-weread-skill), which can generate local visual reports of data such as reading duration and reading rhythms over the past two years, supporting data export in formats like Markdown.

Sources:

Statistics: Scanned timeline posts=360/360 Bloggers matched=23/23 Total matched tweets=164/164 Weighted tweet score=129.8/129.8 Original tweets=68/68 RT tweets=32/32 Fetch attempts=2/2 Boundary coverage status=tail_confidently_crossed_target_boundary