Xiaoliu BOT
中文EN

X Platform AI Briefing May 17: WeRead Skill Ecosystem Expansion, OpenAI Advances Distributed Codex Control, AI Compresses Junior Job Structure

WeRead Skill Ecosystem: Expansion from Tools to Collaboration Methods

Since the official WeRead Skill was released last week, a complete discussion chain formed on X yesterday. @cellinlab compiled a practical collection: including a step-by-step installation guide, a one-sentence modification of the official version (open-sourced by @eviljer), a de-noised refactored version, and the visualization dashboard Booko Board made by @jiayifun based on the Skill. @yaojingang’s open-sourced yao-weread-skill can generate local visual reports from reading data, including modules like reading duration rhythms over the past two years, a cover wall, and a notes wall. @oran_ge mentioned that every Agent can use WeRead as a database to recommend books based on reading records and judge interest, and also discovered WeRead has added a new feature to connect physical books, allowing photo synchronization of handwritten highlights. @op7418 had CodePilot organize their reading data and conduct data analysis tests, with usable results. @AlchainHust is also using it and believes the data value is good. @vista8 simultaneously published a CLI installation and configuration tutorial covering both the official commands and the optimized Skill path.

Sources:

OpenAI Builds Codex Device Network: From Single Device to Distributed Control

@xiaohu observed that OpenAI is advancing Codex’s remote control capability from “controlling your own computer from a phone” to “controlling any computer from a phone.” On May 14, the ChatGPT mobile app added features for real-time viewing of Codex progress on a Mac, approving commands, switching models, and assigning new tasks, but it is limited because the Mac must be awake and unlocked (since Computer Use needs to see the screen and operate the GUI). Currently, OpenAI is developing functionality to allow Computer Use to continue operating when a Mac is locked or sleeping, so tasks can be assigned to a sleeping computer from anywhere without rushing back to unlock it. @xiaohu linked this to OpenAI’s simultaneous progress on mobile remote access solutions (with the phone as the entry point and laptops/Mac Minis/backup devices as execution nodes), pointing out this is the starting point for AI Agents moving from “running on one machine” to “collaboratively running across all your machines.”

Sources:

AI Reshapes Workplace Structure: Junior Roles Are Being Compressed

@LufzzLiz shared a Bloomberg analysis, citing an Oliver Wyman Forum and NYSE CEO survey (415 CEOs, with listed companies collectively accounting for about 10% of global market capitalization) showing: 43% of CEOs plan to reduce junior roles in the next 1-2 years (only 17% last year); 33% plan to shift employee structure towards mid-level; 45% expect total headcount to remain flat, and 29% plan layoffs exceeding 5%. In fields with the highest AI exposure, the unemployment probability for young employees is 16% higher. The core change is a talent structure shifting from a pyramid (with a large base of juniors) to a diamond (thick in the middle, narrow at the bottom): AI first takes over standardized, divisible basic tasks, which were originally entry points for newcomers to practice. The problem is that mid-level managers who can oversee AI agents must first understand the company and business, and these skills are developed during the junior phase. Cutting the entry point for newcomers saves training costs in the short term but may deplete the mid-level talent pool in the long run. IBM is a rare counterexample, announcing in February this year a tripling of entry-level hiring in the US and rewriting job descriptions for the AI era.

Sources:

API Relay Station Gray Supply Chain: The Three-Part Profit Model Behind Cheap Tokens

@wangray and @LufzzLiz separately mentioned a Reddit hot post: Chinese students purchase GPT-5.5/Opus API access at low prices through Xianyu/Taobao, with prices as low as 96-97% of the official price. @LufzzLiz cited a ChinaTalk report pointing out the structure behind this market. Anthropic’s bans are strict (phone verification, overseas credit cards, billing address, government ID + live selfie KYC), but the higher the barrier, the more profitable the bypass industry. Behind the cheapness is a “triple profit from one fish”: the first bite is account quota arbitrage (bulk registration to consume free quotas/corporate education discounts/Max subscription splitting), mixed into account pools bought with stolen credit cards; the second bite is model substitution (users think they are clicking Opus, but it might actually be switched to Sonnet/Haiku or other models; an audit of 17 proxies found widespread model substitution); the third bite is logs, which land on the proxy server, containing code repository context, tool calls, enterprise processes, and decision traces, usable for SFT, distillation, and data brokerage. Users are simultaneously paying customers and free data producers.

Sources:

Practical Ecosystem of AI Image Creation Tools: GPT-Image-2 and Codex Integration

Multiple accounts demonstrated specific use cases combining GPT-Image-2 and Codex. @vansinhu mentioned that Codex CLI recently started supporting direct calls to GPT-Image-2; with the built-in model, marketing Agents take off; their Wenxing Agent Legion added 5 new Codex Runtimes (using built-in GPT-Image-2) poster designer Teams, specifically handling poster demands for various Legion Teams. @op7418 demonstrated a complete video generation solution: Master Zang’s PPT Skill handles aesthetics/layout/animation, HyperFrames manages timeline/rendering/subtitles, Listenhub Skill does voiceover, Jimeng CLI supplements the front-end with demos and B-roll that cannot be generated, and Codex itself generated a video introducing the entire solution. @94vanAI continued using GPT-Image-2 to generate 2.5D thick-paint style anime character posters, noting “this set of Codex works belongs to a surreal terrarium-style character poster illustration” style. @cellinlab used GPT Image 2 to create an HD remastered music video (Suno theme song) and combined it with Suno for music. @MANISH1027512 released a set of mechanical deification style visual works for the “Akatsuki Organization,” explained as a visualization attempt of a “religious community reshaped by pain, belief, and obsession.”

Sources:

Grok Build Rapid Iteration and xAI Product Updates

@elonmusk tweeted on May 17: “Grok Build is improving like lightning,” citing @morganlinton’s shared experience: the team made updates while users were sleeping, and upon waking up, visible differences were already apparent. @xai announced that Hermes Agent now supports X Premium subscriptions, allowing users to use their X paid subscription within Hermes and search X tweets; @NousResearch simultaneously confirmed that agents using the Grok subscription will now automatically get the X Search tool. @elonmusk shared a video of SpaceX Dragon docking with the space station and commented “docking with the @Space_Station has become routine”; content published by SpaceX showed NASA live-streamed the entire docking process.

Sources:

Statistics: Scanned timeline lines=240, Hit bloggers=21, Total hit tweets=127, Weighted tweet score=102.7, Original tweets=47, RT tweets=20, Fetch attempts=1, Boundary coverage status=tail_confidently_crossed_target_boundary