Cognition announced it has raised over $1 billion at a $26 billion valuation, led by Lux Capital, General Catalyst, and 8VC. Enterprise usage has grown more than 10x year-to-date, with annualized revenue reaching $492 million. Devin’s positioning has shifted from an “AI software engineer demo” to a cloud-based agent infrastructure for enterprise software development, with clients including Citibank, Mercedes-Benz, Goldman Sachs, Dell, the U.S. Army, and the U.S. Navy. Internally, 89% of Cognition’s code commits come from Devin, with the remainder primarily from the Windsurf local agent. The company is also working on model evaluation, routing, and its own SWE-1.6, betting that software engineering will become an agent-driven production system.<\/p>\n
Sources:<\/p>\n
Sam Altman announced that the OpenAI Foundation is investing $250 million, with directions including measuring economic changes, supporting worker transitions, and establishing long-term economic security mechanisms. The foundation had previously committed at least $1 billion over the next year, covering life sciences, employment impact, AI resilience, and community projects. This time, it explicitly aims to track where the value created by AI flows\u2014wages, corporate profits, consumers, government tax bases, or capital owners\u2014and to study sovereign wealth funds, AI dividend distribution, and other schemes. This is the first time a major AI company has systematically put the issue of AI economic institutions on the table.<\/p>\n
Sources:<\/p>\n
X (Twitter) has fully rolled out its auto-translation feature, automatically translating all tweets based on the user’s language settings. Multiple bloggers confirmed the feature is active, with mixed reviews: @op7418 thinks it works well and is the only international content platform to achieve full auto-translation; @imwsl90 and @cellinlab said they prefer reading the original and that the AI-translated Chinese reads “very strange,” sharing how to disable it. @xiaohu remarked, “Babel has fallen.” This feature will significantly change the timeline experience for non-English users, but has also sparked discussions about translation quality and reading preferences.<\/p>\n
Sources:<\/p>\n
OpenAI announced that on June 2, it will deprecate GPT-5.2 and GPT-5.3-Codex in Codex, and GPT-5.5 will become the default frontier model for free users. OpenAI’s Tibo (@thsottiaux) said this is to simplify Codex compute cluster management, hinting that new models may be coming soon. Multiple bloggers interpreted this as “making room for new models.”<\/p>\n
Sources:<\/p>\n
@vista8 introduced the Claude Code security guidance plugin, which has 160,000 installations. The plugin uses a pre-tool hook to automatically intercept Write, Edit, and MultiEdit operations, covering six risk types: GitHub Actions command injection, unsafe Node.js child_process.exec() calls, eval()\/new Function(), front-end XSS (including dangerouslySetInnerHTML and innerHTML), Python pickle deserialization, and os.system command injection. Warnings are session-level, with each issue only alerted once. Installation: type \/plugins in Claude Code and search for security-guidance.<\/p>\n
Sources:<\/p>\n
@vista8 summarized Anthropic’s latest Computer Use best practices document. Key recommendations: For the Claude 4.6 series, default to 1280\u00d7720 resolution; for Opus 4.7, recommend 1080p; when building API requests, text instructions must be placed before images; each screenshot consumes approximately 1000\u20131800 tokens. Three counterintuitive conclusions: sending cropped screenshots does not improve accuracy, and drawing coordinate grids is useless; enabling Low thinking saves tokens compared to not enabling it (because it makes fewer mistakes); Max Thinking is completely unnecessary\u2014medium offers the best cost-performance ratio, being half the price of Max\u2014because visual operations are perceptual and mechanical tasks, not logical tasks.<\/p>\n
Sources:<\/p>\n
A new version of the MCP protocol will be released on July 28, with three key new features: servers can deliver HTML interfaces for user interaction, long-running tasks have formal management mechanisms, and authorization and security are enhanced. Meanwhile, OpenAI has launched the Secure MCP Tunnel feature, allowing ChatGPT and Codex to call enterprise intranet MCP servers: install a tunnel-client on the intranet that actively pulls an encrypted HTTPS channel to OpenAI, requests travel through the tunnel to the intranet MCP server and return the same way, solving the problem of “how to securely connect internal systems that enterprises are unwilling to expose to LLMs.”<\/p>\n
Sources:<\/p>\n
Multiple bloggers simultaneously signaled a period of reflection in the AI industry. @oran_ge summarized four points: model+harness is the product, not the model itself eating everything; full automation is self-deception\u2014valuable things require human involvement; slow and meticulous work is a new luxury in the AI era; AI’s ROI is sometimes lower than humans. @vista8 reposted and commented on an in-depth interview, proposing the judgment that “the stronger AI gets, the busier people become”\u2014Every company doubled its employees over the past year, each Agent needs a human caretaker, the CLI era is over and GUI is the main battlefield, SaaS won’t die and Agents will bring more users. @FuSheng_0306 analyzed the organizational change signal behind Karpathy joining Anthropic with only an MTS (Member of Technical Staff) title: Anthropic has 3,000 people, annualized revenue near $30 billion, nearly $10 million per person, and “hive mind” and “campfire model” replace traditional pyramid hierarchy.<\/p>\n
Sources:<\/p>\n
@op7418 open-sourced a Xiaohongshu (Little Red Book) image-text layout Skill, providing 2 themes, 28 layouts, 9 color schemes, and adaptation for 8 major Xiaohongshu content categories. Core design logic: when there are beautiful images, highlight them; when there are no images, automatically match high-quality stock; when image quality is average, beautify it; and proactively avoid using AI image generation to prevent content from being flagged as AI-generated. After a few days of exposure, the Skill has received commercial licensing requests. @op7418 stated that the PPT Skills and Xiaohongshu Skill have proven their commercial value.<\/p>\n
Sources:<\/p>\n
Xiaomi MiMo officially announced the first 13 ecosystem partners joining the MiMo Orbit Program, including gitlawb and others. Previously, the MiMo-V2.5 hundred-trillion Token creator incentive program officially concluded on May 28, distributing 100 trillion Token credits for free to developers within 30 days. Concurrent API price adjustments mainly significantly reduced cache input costs. Some developers reported that their monthly Token quota surged from 700 million to 38 billion. Although the incentive program has ended, the price reduction strategy remains, and Xiaomi has quickly accumulated a group of AI builder users.<\/p>\n
Sources:<\/p>\n
Elon Musk revealed that SpaceX has largely completed development of its internal AI training stack V1.0, written in C, precisely mapped to 220,000 GB300 GPUs (equipped with 800G network cards), heavily using pipeline parallelism to get as close to bare-metal performance as possible. Compared to JAX for large-scale training scenarios, the potential speed improvement exceeds an order of magnitude.<\/p>\n
Sources:<\/p>\n