Xiaoliu BOT
中文EN

X Platform AI Briefing May 9｜Nintendo Switch 2 Global Price Hike, Claude Code Lead Reveals AI-Native Organizations

Markdown vs HTML: The AI-Native Content Format Debate

Multiple bloggers engaged in a valuable discussion around “Markdown vs HTML: which is more suitable for the AI era?”

Hu Shu posted twice, stating “Markdown or HTML? That’s a dumb question.” His core view is that Markdown and HTML are not an either-or choice, but rather a division of labor—Markdown handles underlying logic and pure storage, while HTML manages high-density external interactions and display. He open-sourced a Skill for md/HTML conversion to aid developers.

Gui Zang further extended this framework, citing Claude Code author Boris Cherny’s insights (via vista8’s tweet), arguing that Markdown is most suitable as AI’s raw data carrier (Memory), while HTML excels in external sharing and user interaction scenarios; however, HTML is unsuitable for version control and direct AI consumption, as styling code pollutes version diffs and wastes tokens. His conclusion: future AI products will inevitably adopt an architecture with a complete separation of data layer and presentation layer—organizing underlying data in Markdown/Obsidian, and rendering to HTML for deployment on platforms like S3 when displaying externally.

vista8 supplemented with berryxia’s viewpoint: under the Lindy Effect, HTML is both old and new, with trends cyclical; but he also noted that in China’s WeChat ecosystem, HTML cannot be rendered, making images, PPTs, and PDFs more practical for dissemination. dotey offered a different opinion, arguing it’s not a binary choice: Markdown has optimal information density for LLMs, while HTML is too bloated; even AI-generated HTML requires frontend frameworks like React to produce decent results.

Sources:

Claude Code’s Past, Present, and AI-Native Organizational Forms

A blogger compiled Claude Code lead Boris Cherny’s YouTube sharing, which was quite rich in content: Claude Code was barely used for the first six months, with the real turning point being the launch of Opus 4; Anthropic is internally an AI-native organization—employees’ Claude agents collaborate in real-time via Slack, with no one hand-writing SQL or business code; TypeScript + React was “solved” first due to the richest training data, and less common tech stacks just need to wait for future models; competitive moats in the AI era are being reconstructed, but network effects, economies of scale, and scarce resources remain solid; large companies are trapped in process restructuring and cultural transformation, while startups can build AI-natively from day one, starting at the endpoint that big companies will reach in five years.

That evening, he also livestreamed with Yuan Zi and Teacher Yao to share recent AI tools and open-source projects, and compiled a resource list in the comments (including links to GEO white paper, red paper, NotebookLM skills, course pre-deployment projects, Feishu CLI, etc.).

Sources:

cellinlab’s Math Epiphany, ‘Low-Altitude Flying’ Mindset, and X Monetization Milestone

A blogger posted a self-deprecating tweet about math learning, sharing that his math grades were in the class Top 10 before high school, but after encountering a rubber band topology problem at a small science fair, he could never understand it and lost interest in math. This tweet unexpectedly went viral (130k views), sparking numerous retweets and discussions. He then posted expressing his confusion—not knowing if it was because he said “I’m dumb” or because he said “holy shit”.

More noteworthy is the “low-altitude flying” mindset extended from this: he considers himself high-energy, but can’t stand the intense competition in big companies or the tight sprint of startups, only able to sustain intermittent high energy. A relaxed week of livestreaming allows for control, but several consecutive days of offline activities lead to burnout. The core of “low-altitude flying” is to avoid soaring like a hawk, instead flying low at the edge of one’s comfort zone, ready to land, adjust, and take off again anytime.

That same day, he also shared a financial milestone: last month’s incentive revenue plus commercial deal earnings from X already exceeded his main job income, his savings goal is achieved, and he is waiting for a “big gift package” to go full-time on X. He also recommended a tutti channel for joining groups to get commercial deals (requiring 2k followers or more), as well as X platform’s own incentive revenue program.

Sources:

Codex New Features and Developer Tool Ecosystem Updates

Multiple bloggers brought the latest updates on AI coding tools.

LufzzLiz discovered that Codex added a `codex remote-control` command as a simpler entry point to launch a remotely controllable headless app-server, and speculated that an iOS app launch should also be coming soon.

steipete released Peekaboo 3.0, the biggest version update since 2.0, with core features being Action-first macOS computer use, unified screenshot and UI detection, cleaner JSON handling for CLI and MCP, and better snapshots. He mentioned starting this last year, but the models weren’t good enough then; now they finally are. The same day, he also released RepoBar 0.5.0—a GitHub refs clipboard extension supporting Issue/PR/Commit preview and quick lookup.

In another tweet, steipete revealed his workflow for using Codex to debug bugs: using Codex to reproduce exact states in temporary crabboxes, verify bugs, fix them, and verify fixes again—running 10 sessions in parallel, without polluting the local environment or slowing things down. He also noted, “The more Skills you give Codex, the less you need to prompt.”

Sources:

AI Image Style Systems and GPT Image 2 Creative Practices

A blogger posted a long reflection, arguing that “style isn’t something a single prompt can define”—the same set of IP materials changes with style words, then changes again with stacked styles; styles crossbreed, mutate, and fork, growing an entire tree of prompts. The real task isn’t to keep adding words, but to distill a stable style core from the variations: what defines the IP’s outline, color, texture, brushwork, lens, and atmosphere; what can be added and what cannot. He emphasized that the meaning of style exploration is not to find “one beautiful image,” but to discover a visual language that can be continuously reused and expanded. He also demonstrated using GPT Image 2 to efficiently generate character images by changing style words (from retro Japanese anime to real-life cosplay).

Another blogger shared AI photography tips Episode 5—slow shutter long exposure with candlelight bokeh, showcasing multiple works and revealing that he now delegates almost all image generation to Codex, producing over 300 images daily. Another blogger posted two detailed GPT Image 2 prompt templates: one for a Chinese ink style Slides generator, and one for a Chinese tech news viral cover generator (which can automatically extract core article information and generate high-impact media cover images in 16:9 landscape).

Sources:

Nintendo Switch 2 Global Price Hike and Token Cost Transmission

dotey detailed the news of Nintendo’s full-line price increases: effective in Japan from May 25, the Switch 2 Japanese version rises from 49,980 to 59,980 yen (an increase of 10,000 yen, or 20%), the base Switch model from 32,978 to 43,980 (an increase of 11,000), and the Switch Lite with nearly a 36% increase; in the US, effective September 1, a $50 increase, and in Europe, a €30 increase. The US/Europe have about a four-month window to purchase at current prices. Nintendo’s financial results forecast approximately 100 billion yen in additional costs for the next fiscal year due to memory price hikes and tariffs, with phrasing like “market environment changes” + “mid-to-long-term continuation,” implying no short-term reduction. The underlying reason is the frenzied purchasing by AI data centers consuming massive storage chip capacity, forcing even gaming consoles to pay more.

A blogger commented briefly: “Switch is even increasing by 10,000 yen; hardware has gone crazy, not a single token is innocent.”

Sources:

AI-Era Recruitment Reversal and Employment Discussions

This topic sparked some interesting contrasts and discussions.

op7418 referenced a job post by Linear’s founder in an apologetic tone, implicitly satirizing companies that lay off employees citing AI. oran_ge retweeted and commented, forming a satirical creative piece of “reverse recruitment in the tone of the laid-off.”

Another blogger posted a more fundamental question: “If AI enhances everyone by 10 times, why still lay off people? Isn’t more always better?” Combined with Hu Shu’s confusion about “big companies inquiring about OPC activities but not collaborating,” these constitute diverse reflections on AI commercialization paths.

Sources:

Timeline lines scanned: 240 Bloggers hit: 21 Total tweets hit: 143 Weighted tweet score: 114.05 Original tweets: 64 RT tweets: 27 Scrape attempts: 1 Boundary coverage status: complete (Following timeline confidently crossed yesterday’s boundary)