Xiaoliu BOT
中文EN

X Platform June 17 AI Brief | SpaceX Acquires Cursor and Launches Code Hosting Service, GLM-5.2 Released Open Source, Anthropic’s Frontier Models Designated as National Security Assets

SpaceX Acquires Cursor for $60 Billion, Officially Launches Origin Code Hosting Service and 1.5 Trillion Parameter Model

Days after its IPO, SpaceX announced the acquisition of Anysphere, the parent company of AI code editor Cursor, in an all-stock deal valued at approximately $60 billion. This acquisition marks the largest industry event in the AI programming tools sector this year. Cursor was previously valued at $29.3 billion with annualized revenue reaching $2.6 billion. Post-acquisition, its jointly trained models will be deployed on Cursor and Grok Build. Concurrently, at the Compile conference, Cursor announced the launch of its code hosting and Git service, Cursor Origin (expected in Fall), and disclosed it is pre-training a new model with over 1.5 trillion parameters using more than 100,000 GPUs. Cursor CEO Michael Truell stated the goal is to “invent a new kind of programming,” gradually transitioning from writing code line-by-line to describing software intent with natural language.

Sources:

GLM-5.2 Officially Released and Open-Sourced, Multiple Bloggers Provide Varied Evaluations

Zhipu AI (Z.ai) has officially released and open-sourced its flagship model, GLM-5.2. Its core focus is long-context tasks, supporting a 1 million token context window. It introduces Think Level Control and the IndexShare sparse attention mechanism, which can reduce per-token computation by approximately 2.9x for million-token inputs. Total parameters are 744B, with 40B active parameters. The model ranks second on the Coding Arena Frontend leaderboard, just behind Claude Fable 5. Blogger @op7418’s tests found it performs well in instruction following and template tuning but is slightly weaker in autonomy and creativity, failing to automatically embed image links from documentation. Another blogger, @LufzzLiz, believes it can reach top-tier levels with precise prompts. Blogger @MaxForAI noted that Chinese open-source models are now only 6-8 months behind US closed-source models in coding, with the main gap being in planning capability rather than coding ability.

Sources:

WeChat Pay Officially Launches AI-Specific Card, WorkBuddy Among First to Integrate

WeChat Pay announced the launch of an AI-specific card—a standalone “mini-wallet” designed for AI agents to call upon. Users can initiate search, selection, and payment processes directly through AI, with Tencent’s WorkBuddy among the first to integrate. Each payment requires user confirmation on their mobile phone before deduction, with limits set by the user. Currently, it only supports the macOS desktop version, with more versions to follow. The card will eventually be open to more compliant intelligent agent platforms.

Sources:

Anthropic’s Frontier Models Designated as National Security Assets, UK Citizen Access Requests Denied

UK Prime Minister Keir Starmer requested an exemption from the US government to allow British citizens and companies access to Anthropic’s Mythos and Fable frontier AI models, but was refused. An official from the Trump administration told the New York Post that even an exemption for G7 allies was “completely illogical.” This incident marks a shift where frontier AI models are transitioning from commercial products to geopolitical privilege assets. Separately, according to a letter from the US Department of Commerce to Anthropic’s CEO obtained by Bloomberg, the Commerce Department has ordered Anthropic not to provide the Fable 5 and Mythos 5 models to any foreign nationals without permission.

Sources:

OpenAI Codex Opens to Third-Party Open-Source Models, Details Three Computer Control Methods

OpenAI Codex team member Thibault Sottiaux announced that Codex’s App, CLI, and SDK can all connect to any open-source model, with no requirement to use OpenAI’s own models, and has released integration tutorials. This open strategy has sparked discussion in the developer community. Concurrently, Codex’s rate limits have been fully reset after issue resolution. Codex team member Jason also released a detailed guide on three methods for computer control: Computer Use (broadest, slower, can operate any desktop application), Chrome Extension (operates web pages with login state), and Built-in Browser (a front-end development sandbox supporting annotation feedback loops).

Sources:

Grok Imagine Video 1.5 Fully Launched, Elon Musk Claims Full Film Generation Possible by Year-End

xAI officially released Grok Imagine Video 1.5, a model that generates video from images, supporting clearer realism, improved physics simulation, faster generation speed, and multilingual voiceovers (including Chinese). Elon Musk confirmed the model is fully open and stated, “Full film generation will be achievable by year-end.” Concurrently, Grok 4.3 is now available on Amazon Bedrock, and Grok Voice ranked first in Vapi’s “Humanlikeness Index” evaluation. Multiple bloggers showcased high-quality video clips generated using the model.

Sources:

NVIDIA ENPIRE Released: AI Agent Autonomously Controls Real Robot for Experiments in Physical World for First Time

NVIDIA’s GEAR Lab (jointly with CMU and UC Berkeley) released ENPIRE, a system that enables AI programming agents to autonomously control real robots to complete experiments, marking the first time the “AutoResearch” concept has moved from digital to physical environments. The agent can automatically set up experimental environments, improve control strategies, run experiments on real robots, analyze failures, and iterate. Three agents—Codex, Claude Code, and Kimi Code—can all complete the full process. It achieved a 99% success rate on dexterous manipulation tasks like inserting GPU cards, threading zip ties, and cutting zip ties. Eight robots exploring in parallel demonstrated a clear “physical scaling law” effect.

Sources:

Midjourney Announces Upcoming Hardware Product Launch

Midjourney officially announced it will hold an in-person launch event in San Francisco on June 17 at 18:00 PT to unveil its first hardware project. Blogger @op7418 commented, “Midjourney is actually releasing hardware.” The specific product form has not been disclosed.

Sources:

Claude Code Founder Shares Agent Engineering Philosophy: Do Less, Let the Model Grow

A core member of the Claude Code team recently systematically explained their agent engineering philosophy in a talk. Key points include: CLAUDE.md files should be as short as possible, regularly cleared and rewritten; insisting on CLI over GUI because models improve too fast, making GUIs obsolete in six months; the journey from verbose to concise terminal output was refined through user feedback and iteration. He gave an example: he spent half a day troubleshooting a memory leak without success, but a teammate handed the problem to Claude Code, which wrote its own tool to analyze the heap dump and located the issue faster. The summary is: “Models are soaring; the optimal human strategy is to do less, stay lightweight, and delegate judgment to increasingly powerful models.”

Sources:

OpenAI GPT-Bidi-1 Voice Model Discovered Early, Microsoft Considers Switching Copilot to DeepSeek

OpenAI’s next-generation voice model, GPT-Bidi-1, was discovered early among developers and may be released soon, focusing on more natural conversational experiences. Separately, according to Axios, Microsoft is considering switching the underlying model for Copilot Cowork to a fine-tuned version of DeepSeek V4 or other Chinese open-source models, replacing Anthropic and OpenAI solutions, with lower-cost model options expected in the coming weeks.

Sources:

Stats: Timeline Scans=360 Bloggers Matched=41 Total Tweets Matched=221 Weighted Tweet Score=176.6 Original Tweets=105 RT Tweets=44 Fetch Attempts=2 Boundary Coverage Status=tail_confidently_crossed_target_boundary