Xiaoliu BOT
中文EN

X Platform June 4 AI Brief | SpaceX sets record IPO valuation at $1.77 trillion, internet bot traffic surpasses humans for first time, Google open-sources Gemma 4 12B

SpaceX Files IPO Amendment, Valuation at $1.77 Trillion Sets Record

SpaceX filed an amended IPO document with the SEC on Wednesday, announcing the issuance of 555.6 million Class A shares at a fixed price of $135 per share, raising $75 billion, with an overall valuation of $1.77 trillion—making it the largest IPO in human history. At this valuation, SpaceX would become the seventh largest publicly traded company in the U.S., surpassing Tesla’s $1.6 trillion market cap, and Elon Musk’s personal stake would be worth $841 billion. SpaceX CFO Bret-Johnsen released a 17-minute roadshow video the same day, disclosing key data: Starlink has 9,600+ satellites in orbit, users grew from 4.4 million at the end of 2024 to 10.3 million in Q1 2026, covering 164 countries and 3 billion people; 2025 revenue was $19 billion (up 30% YoY), with positive EBITDA of $7.8 billion, and Q1 standalone revenue of $5 billion. Notably, when SpaceX merged with xAI in February this year, its valuation was only $1.25 trillion—a 41% jump in four months.

Sources:

Evan You’s VoidZero Team Joins Cloudflare, Vite Ecosystem Gains Platform-Level Support

VoidZero announced it is joining Cloudflare. Evan You and the team will continue to lead Vite, Vitest, Rolldown, Oxc, and Vite+ projects, all under the MIT license. Cloudflare said it will co-invest in tools that developers rely on daily, while tightly integrating the Vite ecosystem with the Cloudflare platform. VoidZero’s mission remains unchanged: making JavaScript developers more productive than ever. This acquisition means the frontend toolchain ecosystem now has long-term support from an infrastructure platform.

Sources:

Grok Ecosystem Expands Across Multiple Fronts: Video, Voice, Cloud All Top Rankings

xAI’s Grok series products have achieved first place on multiple leaderboards. Grok Imagine Video 1.5 tops the Video Arena leaderboard; Grok Voice Think Fast 1.0 ranks first in Artificial Analysis’s τ-Voice benchmark, surpassing GPT-Realtime-2 and Gemini 3.1 Flash, and has already been deployed in real business scenarios. On the platform partnership front, Grok LLM and audio/video models are now available on Cloudflare AI Gateway, billed directly through Cloudflare without requiring an additional API key. Meanwhile, Grok Imagine Video 1.5 is also available on Vercel AI Gateway, supporting image-to-video with synchronized audio generation. Additionally, Gopuff announced a partnership with SpaceXAI to launch Go, an AI shopping assistant built with Grok.

Sources:

Google Open-Sources Gemma 4 12B: Encoder-Free Unified Multimodal Architecture, Runs Locally on 16GB

Google released the Gemma 4 12B open-source model, adopting a “Unified” encoder-free architecture that feeds text, images, audio, and video directly into a single Transformer backbone, eliminating traditional vision and audio encoders. The model supports a 256K token context window, 140+ languages, built-in Thinking mode, and native Function Calling. On the hardware front, it runs on 16GB VRAM or unified memory, with 4-bit quantization down to 8GB. Additionally, Google AI Edge Gallery has expanded from mobile to macOS desktop, featuring a sandboxed Python environment and voice interaction—install and run Gemma 4 12B immediately.

Sources:

Ideogram 4.0 Open-Sourced: Introduces Bounding Box Precision Control, Text Rendering Reaches Open-Source Ceiling

Ideogram released its first open-source AI image model, Ideogram 4.0 (9.3B parameters, based on Qwen LM). Key highlights include: bounding box control, allowing precise coordinate specification of element positions in the image; structured JSON prompt format; multilingual text rendering, with English OCR accuracy reaching 0.97, supporting dense text rendering for non-Latin scripts like Chinese, Japanese, and Korean. The Decoder’s evaluation shows it significantly outperforms Midjourney v8 on benchmark prompts, roughly on par with FLUX, but behind GPT Image 2 and Banana Pro. Gorden Sun’s hands-on test found the aesthetics on point, English rendering perfect, and Chinese text limited in length but already the best among open-source models.

Sources:

Internet Bot Traffic Surpasses Human Traffic for the First Time: 57.5% of HTML Requests Come from Bots

According to real-time statistics from Cloudflare Radar, over the past week (May 28 to June 4), 57.5% of all global HTML page requests came from bots (crawlers, AI scrapers, automation scripts), while only 42.5% came from real human browsers. When all HTTP traffic is classified by returned content, JSON accounts for 33.1% (ranked first), while HTML accounts for only 12%—meaning the majority of internet traffic is now machine-to-machine API communication. Multiple bloggers have called this a “historic moment,” and Cloudflare’s co-founder shared the data, noting it arrived faster than expected.

Sources:

Kimi-Code Rapidly Iterates After Refactoring: Lead Developer Shares a Month of Intense Experience

Kimi-Code lead developer @real_kai42 shared the refactoring experience: a month ago, he locked himself in a hot spring resort with a laptop and portable monitor for two days, spending thousands of dollars on tokens for architecture analysis and validation. After confirming the new architecture, he quickly assembled a team and went into closed development. After open-sourcing, he fell ill due to excessive cortisol secretion, drinking an entire case of Red Bull in a week. The iteration speed after refactoring was astonishing—updating from 0.8.0 to 0.9.0 in less than 24 hours, adding ACP, Goal official release, plugin system (kimi-datasource, Superpowers), and scheduled tasks. Blogger @imwsl90 tested it and found it still lags behind Claude/Codex.

Sources:

html-video Open-Sourced: Agent Writes HTML to Generate Videos, 30,000 Lines of Code in 3 Days

Tom Huang (@tuturetom) officially open-sourced html-video, positioning it as the “HTML version of CapCut.” The core idea is to let coding agents (Claude Code, Codex, Cursor, etc.) generate product promotion and knowledge explanation videos by writing HTML, then render them to MP4 via ffmpeg. The project took 3 days and 30,000 lines of code, supporting 20+ video style templates, paginated editing, and MP4 export. On the release day, it surpassed 100 Stars, and community users shared videos made with html-video, including an introduction to Open Design 0.9.0. The project integrates HeyGen’s HyperFrames framework and supports defining visual style and motion specifications via design.md and frames.md.

Sources:

VS Code High-Risk Vulnerability Exposed: Can Steal GitHub Tokens; Security Researcher Discloses Publicly Due to Dissatisfaction with Microsoft

Security researcher Ammar Askar disclosed a high-risk VS Code vulnerability: attackers can trick users into clicking a specially crafted link to steal their GitHub OAuth credentials via the GitHub online editor. These credentials have read/write access to all repositories (including private ones), potentially enabling supply chain attacks. The researcher, dissatisfied with Microsoft Security Response Center’s practices (silent fixes, no acknowledgment, low severity assessment), reported the issue to the VS Code team only one hour before public disclosure, sparking community controversy. Landian News first reported the incident.

Sources:

Statistics: Scan timeline lines=360, bloggers hit=48, total tweets hit=256, weighted tweet score=213.55, original tweets=132, RT tweets=38, fetch attempts=2, boundary coverage status=tail_confidently_crossed_target_boundary