2026-05-25 AI Daily | The Diffusion of Programming Power: Codex Becomes a Communication Protocol, AI-Native Workflows Reshape Organizations Link to heading

Dan Shipper’s experiment with a 30-person team confirms that large-scale AI coding by non-technical staff is now a reality, as automation pushes humans toward higher-level judgment and creativity. Meanwhile, OpenAI Codex is evolving from a tool into a “communication protocol” connecting the cloud, local environments, and devices, and Agent goal management is becoming more engineered. Furthermore, Anthropic’s annualized revenue run-rate has surged to $30 billion, with Karpathy quietly joining, intensifying the tension between commercialization and safety philosophies in the industry.

📖 Deep Dive into This Issue’s Watch List Link to heading

Today’s two main themes both point to the same core contradiction: the unmanned future promised by AI is instead giving rise to more intensive human-computer collaboration. The paradox proposed by Dan Shipper in his latest conversation—“the more automation, the busier humans get”—combined with the vivid experiment of all 30 members of his company using AI extensively, is absolutely the most worthwhile content to break down frame-by-frame today. His prediction from a year ago that “non-technical staff will code on a large scale using Claude” has been validated by the daily workflows of editors and operations staff, which directly leads to the second unavoidable topic: the diffusion of programming power. At the same time, drawing from multiple sources today on “AI-native workflows,” we see a clear picture: future organizations won’t eliminate people but will instead push them towards higher-level judgment and creativity, requiring everyone to possess the ability to direct AI. We recommend that all engineering leaders and product managers currently designing team structures add this podcast episode to their must-listen list for the week.

🌐 Breaking AI News on X Link to heading

Topic 1: xAI Completes Training on Massive 1.5 Trillion Parameter Grok V9-Medium Link to heading

• Category: AI · News
• Overview: Trending Time:, Related Posts: 5900
• What it is: xAI has completed the training of its massive 1.5 trillion parameter model, Grok V9-Medium.
• Why it matters: At 1.5 trillion parameters, this is one of the largest language models publicly trained to date. Its successful training validates the feasibility of ultra-large-scale computing clusters and distributed training techniques, potentially pushing the upper limits of model capabilities.
• Discussion overview: The discussion centers on the model’s secrecy, the lack of publicly released performance benchmarks, and comparisons of its capabilities against closed-source models like GPT-4 and Gemini. Concurrently, some question whether such a massive parameter count constitutes over-engineering and raises cost-effectiveness issues.

Topic 2: Huawei Unveils Tau Scaling Law for 1.4nm Chip Density by 2031 Link to heading

• Category: AI · News
• Overview: Trending Time: 10 hours ago, Related Posts: 14000
• What it is: Huawei has proposed the “Tau Scaling Law,” aiming to achieve chip density equivalent to a 1.4nm process by 2031 by focusing on packaging and data rates rather than shrinking transistors.
• Why it matters: This offers a new paradigm for overcoming US chip sanctions and developing independent advanced manufacturing processes. It could reshape the evolutionary path of computing power for AI hardware, reducing the sole reliance on cutting-edge lithography.
• Discussion overview: The discussion focuses on the practical feasibility of this route, whether it can truly bypass export controls, and whether it represents a disruption of traditional Moore’s Law or is a stopgap measure for a specific, constrained environment.

Topic 3: DeepSeek Makes 75% API Discount Permanent on V4-Pro Model Link to heading

• Category: AI · News
• Overview: Trending Time: 21 hours ago, Related Posts: 1600
• What it is: DeepSeek announced a permanent 75% price reduction for its V4-Pro model’s API.
• Why it matters: This move significantly lowers the cost barrier for large model inference, potentially accelerating the development and commercialization of AI applications and intensifying price competition among cloud services and model providers.
• Discussion overview: The discussion focuses on whether this will trigger a new round of API price wars. Developers generally welcome the price cut, but some express concerns about long-term sustainability and the potential impact on service quality. Others see it as a signal that the commercialization of large models in the Chinese market is shifting from ‘pay-as-you-go’ to ‘market penetration at scale.’

Topic 4: Anthropic’s Claude Revenue Surges to $30 Billion Run-Rate Amid Philosophical Buzz Link to heading

• Category: AI · News
• Overview: Trending Time:, Related Posts: 334
• What it is: The annualized revenue run-rate for the Claude model from AI startup Anthropic has surged to $30 billion, attracting widespread attention.
• Why it matters: This marks an acceleration in the commercialization of cutting-edge large models, validating that an AI product differentiated by its safety philosophy can also achieve explosive growth, potentially reshaping the AI industry landscape.
• Discussion overview: Current discussions are focused on two aspects: first, an analysis of the enterprise customer penetration rate behind its staggering revenue growth; and second, a debate over whether Anthropic’s long-standing “AI safety philosophy” is a genuine defensive moat or simply a brand marketing strategy.

Topic 5: Andrej Karpathy Joins Anthropic as Member of Technical Staff Link to heading

• Category: AI · Other
• Overview: Trending since: 1 day ago, Related posts: 4,500
• What Happened: Renowned AI researcher and former head of AI at OpenAI and Tesla, Andrej Karpathy, announced he is joining Anthropic as a member of the technical staff.
• Why it matters: Karpathy has significant influence in deep learning, computer vision, and language models. His addition will considerably strengthen Anthropic’s capabilities in AI safety and advanced model research and could impact talent-flow dynamics in the industry.
• Discussion Summary: Discussions are focused on why Karpathy chose Anthropic over returning to OpenAI or starting his own venture. The general consensus is that this highlights the appeal of Anthropic’s safety-focused philosophy. Some also noted his title is “member of technical staff” rather than an executive role, speculating that he might be seeking more freedom for independent research.

Summary of Today’s AI Public Opinion on X Link to heading

Today’s main AI narrative revolves around the tension between the race for scale, technological breakthroughs, and commercialization. The industry consensus is that the commercialization of large models has accelerated significantly. From DeepSeek’s aggressive price cuts to capture market share to Anthropic’s explosive revenue growth, it’s clear that companies are shifting from competing on performance to pursuing large-scale market penetration. Disagreements, however, are centered on implementation paths and long-term value. xAI’s massive-parameter model is being questioned for potential over-engineering and cost-effectiveness issues. It remains undecided whether Huawei’s “Tau Scaling Law” is a paradigm-shifting revolution to overcome chip restrictions or a temporary solution born from a constrained environment. Meanwhile, the debate over whether Anthropic’s safety philosophy is a genuine competitive moat or just brand packaging has intensified with Karpathy joining as a regular technical staff member. Potential risks include intense price wars eroding industry sustainability and an excessive chase for more parameters or alternative techniques leading to resource misallocation and an artificial boom. At the same time, the accelerated consolidation of top talent and capital towards a few entities championing a “safety narrative” could exacerbate the conflict between long-term safety principles and tech-pragmatism under industry-wide commercialization pressure.

💡 Influencer Insights Link to heading

Based on the content of tweets from AI influencers over the past 24 hours, here is the industry report extracted and summarized for you.

AI Daily (2026-05-25) Link to heading

1. Key Tech Trends and Hot Products Link to heading

Today, influencers’ conversations are highly focused on the engineering implementation of AI Agents and their deep integration into workflows. In particular, endpoint Agent products, represented by OpenAI Codex, are evolving from “tools” into “collaborators.”

🤖 Codex: A Full-Scale Invasion, from Browser to OS Link to heading

Codex is the undisputed focus today. Influencers are paying attention not just to its features, but to the work paradigm shift it is triggering.

• Browser Capabilities Become Standard: @zhixianio and @Pluvio9yte both highly praised Codex’s @chrome capability. @Pluvio9yte stated directly that this is the most efficient way to replicate a front-end page, requiring only the command @chrome 帮我复刻 xxx.xx 的页面, with results superior to manually capturing DOM or HAR files.
• Goal-Driven and Long-Task Management: The /goal mode is seen as the key differentiator between a “tool” and an “assistant.” @zhixianio shared a case where they spent several hours building a local information filtering tool with it and pointed out a hidden use for /goal clear. @dotey provided a detailed analysis of management techniques for Goal mode, including using /side to open a Side Chat for monitoring long-task progress, and how to use the panel to pause, edit, and Steer (task intervention), elevating the agent’s task flow management to a project level.
• Becoming a Foundational Communication Protocol: @vista8 discovered that ChatGPT can remotely call a local instance of Codex, allowing them to command their computer via phone to generate and publish music during a fishing break. This implies that Codex is becoming a “communication protocol” connecting cloud models with local computing power and between different devices.

⚙️ Agent Engineering and Ecosystem Building Link to heading

The Agent ecosystem is evolving from isolated tools to an interconnected network.

• Multi-Agent Deployment and Communication Bridges: @dotey provided an in-depth analysis and recommendation for feishu-claude-code-bridge, a project that allows users to command Claude Code directly from within Feishu, enabling two-way “IM Message ↔ Agent CLI” operations. @ruanyf also pointed out that Feishu’s open-source CLI toolkit is the most feature-complete for Agent integration, with over 10,000 stars, signaling that office automation is becoming a key scenario for Agent implementation. @zhixianio also mentioned that Telegram supports bots managing other bots, which facilitates multi-agent deployment.
• Revolution in Development Paradigms: @Pluvio9yte, addressing the pain point of “not knowing what you’ve written after vibe-coding,” recommends the opensec development paradigm. It generates detailed process documentation for Claude Code, Codex, Cursor, etc., shifting AI programming from “freestyle improvisation” to “traceable engineering.”

💻 Fortifying the Compute Foundation for On-Device Models Link to heading

• Hardware Platforms Are Ready: @zhixianio noted that AMD has released the Ryzen AI Halo mini-PC, specifically designed for running LLMs locally. Combined with his previous attention to the Qwen on-device model and Mac Studio, this indicates the industry is actively building up the compute foundation for an explosion in on-device models. This also resonates with his judgment that Anthropic might implement KYC due to high demand, leading to the approach of “big tech tyranny,” making on-device models a crucial counterbalancing force.

2. Noteworthy Unique Perspectives and Industry Foresight Link to heading

1. “Markdown is the language of AI, HTML is the language of humans” — @Pluvio9yte This is a highly insightful perspective. It precisely defines the division of labor for information formats in current human-computer interaction: AI excels at generating structured Markdown, while the final visual presentation and interactive experience (HTML) are designed for humans. This offers a new direction for AI content output standards: having AI output both .md and .html formats may be the best practice for balancing efficiency and user experience.

2. "‘Content Assetization’ in the Age of Agents" — @AI_Jasonyu He imported all his content from the past three years into Obsidian and fed it to Claude for “style distillation.” His practice highlights that every piece of content you publish today is training data for your future personal AI. This is not just a personal branding technique, but a forward-looking “digital asset” accumulation strategy—making AI an extension of your own style, not a replacement. @Pluvio9yte corroborated this from the opposite angle, observing that reports written by interns who heavily use AI have a strong “AI flavor.” This raises a new warning: AI is not just learning from humans, it is also assimilating them, making it a new challenge to maintain a unique style.

3. “The Best Practice for Browser Agents is ‘Verifiable Research’” — @Pluvio9yte He provided clear boundaries for the application of browser agents: their most suitable initial use is not placing orders, but conducting web research with source verification. He proposed a five-element output standard: “source, time, field, conclusion, blank,” and emphasized that “research without sources is essentially just chat-based generation.” This offers a highly practical methodology for using agents reliably.

4. Industry Wake-Up Calls and Reflections

   • The Murky Waters and Backlash of AI Content: @Pluvio9yte exposed the chaotic commercialization of proxy review websites that “want to have their cake and eat it too,” warning that industry transparency is eroding. Meanwhile, @gefei55 revealed the phenomenon of AI tool sites using low-quality traffic from sources like “Youtube Downloaders” to inflate their numbers and deceive investors, sounding an alarm for capital and data professionals.
   • The Disappearance of Software Moats: The point previously made by @ruanyf that “the moat of code no longer exists; testing is the new moat” still resonates today. When AI can replicate Next.js for $1,100, the key to preventing replication is no longer the code itself, but the test cases that define the product’s behavior and quality.

5. Philosophical and Business Reflections

   • @lijigang, analyzing from a cybernetics perspective, states that “constraints” are key to understanding corporate and individual behavior, as they carve out a predictable space for action. This is not only an interpretation of AI logic but also a methodology for observing the business world.
   • The business logic behind the $150/month unlimited AI package shared by @gefei55 reveals that “selling certainty (eliminating quota anxiety)” is itself a high-profit business model. The premium customers pay for “peace of mind” far exceeds the cost of their actual usage.

3. Recommended Tools and Resources Link to heading

• Open Source Projects and Tools

  • feishu-claude-code-bridge / lark-channel-bridge: Developed by @zarazhangrui, strongly recommended by @dotey. It allows you to direct Claude Code in Feishu (Lark) just like chatting with a colleague, with two-way operations for files and documents.
  • opensec: A development paradigm tool recommended by @Pluvio9yte. It supports Claude Code, Codex, Cursor, etc., providing detailed process documentation for AI-generated code and turning “Vibe Coding” into “Documented Engineering.”
  • CloakBrowser: A new open-source browser automation solution discovered by @Pluvio9yte. It’s based on source-level modifications to Chromium and claims to be able to stably pass security checks like Cloudflare, representing a new direction for stealth browsers.

• Capcap: An open-source Mac screenshot tool shared by @jaywcjlove and forwarded by @dotey, which integrates window snapping, scrolling screenshots, and annotation.

• YY远程 (UU Remote): A free remote desktop solution reviewed in detail and recommended by @Pluvio9yte, who specifically noted its excellent support for keyboard and terminal control on the iPad, making it ideal for connecting to a primary machine for troubleshooting or development from a lightweight device anytime, anywhere.

• Industry Resources & Best Practices

  • PDF Report on OpenAI’s Internal Use of Codex: @Pluvio9yte shared this internal document, revealing how its security, infrastructure, and frontend teams use Codex daily for codebase understanding, refactoring, and test generation. It serves as a firsthand resource for understanding the internal practices at a leading AI company.
  • Translation of “DeepSeek’s $10 Trillion Grand Strategy”: An article shared and translated by @dotey, offering a perspective on DeepSeek’s macro strategy.
  • Reuters “AI and News” Report: Cited by @vista8, data from this report shows that only 12% of the public is comfortable with purely AI-generated news, whereas the “human-led, AI-assisted” model has a 43% acceptance rate, pointing the way forward for the content creation industry.

📚 Appendix: Today’s Watch List Source Updates Link to heading

Timeframe: Last 3 days; covers 16 sources; 1 update in total

Lenny’s Podcast (A_full) Link to heading

• The AI paradox: More automation, more humans, more work | Dan Shipper
  • Published: 2026-05-24 20:31 Beijing Time
  • Abstract: - Dan Shipper is the co-founder and CEO of Every, a media and software company that has become a living laboratory for the future of work.
    • His company has about 30 employees, and everyone is an AI early adopter; from editors to operations staff, they use artificial intelligence to do most of their work, giving everyone a unique perspective on where the world is heading.
    • A year ago on this show, Dan predicted that people would rely on Claude Code for non-technical work, which proved to be remarkably prescient.
    • Today he’s back with another set of calls: the SaaS apocalypse is dumb, CLIs are over, the forward-deployed engineer is the most valuable new hire, and to keep your job, the only thing you need to do is master the models.
    • The future of work will happen inside Codex or Claude Code.
  • Key Points (EN):
    • Dan Shipper is the co-founder and CEO of Every, a media and software company that’s become a living laboratory for the future of work
    • Everyone at his company of about 30 people is an AI early adopter; from editors to ops people, they use AI to do much of their work, giving Every a unique lens…
    • A year ago on this show, Dan predicted that people were sleeping on Claude Code for nontechnical work, which proved to be remarkably prescient
    • Today he’s back with another set of calls: the SaaS apocalypse is dumb, CLIs are over, the forward deployed engineer is the most valuable new hire, and the only…