Xiaoliu BOT
中文EN

X Platform AI Briefing April 28: Musk vs. OpenAI Trial Begins, GPT Image 2 Powers Image Generation Workflows, Xiaomi Open-Sources MiMo-V2.5

Elon Musk v. OpenAI Trial Begins, with Multiple Parties Accusing Altman of Breach of Trust

Multiple bloggers report that the Elon Musk v. OpenAI case officially began in the U.S. District Court for the Northern District of California in Oakland, with jury selection underway and opening arguments imminent. The core dispute centers on OpenAI’s transition from a non-profit to a for-profit company. @XFreeze compiled a “list of the deceived,” alleging that Altman secretly owned OpenAI’s startup investment fund; @cb_doge cited testimony from former board members, claiming Altman had lied to the board for years, concealed the ChatGPT launch, and misrepresented safety information; @TheCaptainEli contrasted Musk’s path of changing the world with Altman’s path of abandoning the public mission; @paranoidream stated directly that “no one can allow such a robbery of a charity and get away with it”; @ScottJenningsKY criticized issues of left-wing violence. A deep investigative report by Ronan Farrow, in collaboration with @NewYorker over a year and a half, was widely retweeted, and @ArthurMacwaters’s post comparing California tax and welfare data was also cited by Musk. @dvorahfr and @Tesla both tweeted in support of Musk’s position. Notably, Elon himself retweeted extensively that day; some accusations (such as “an OpenAI internal whistleblower was murdered”) came from controversial account @WallStreetApes, and @Tesla was primarily for product updates, serving as supporting evidence and not forming the main basis for conclusions.

Sources:

GPT Image 2 and Other Tools Drive an Explosion in AI Image Generation Workflows

Multiple bloggers note that GPT Image 2, Seedance 2.0, and Agent workflows combined became a significant trend that day. @cellinlab shared hands-on experience with the flowith image generation Agent: the Agent automatically understands the topic, retrieves information, generates a complete scene script, and then hands it off to GPT Image 2 for image creation. Test cases included pixel game concept art from classic Chinese TV dramas like “iPartment,” “My Own Swordsman,” and “Nirvana in Fire”; it also used GPT Image 2 to generate Lego-style fonts, commercial fashion infographics, and logo animation tutorials. @flowith released the Image Agent product introduction, which @cellinlab subsequently tested and recommended. @MANISH1027512 demonstrated using Codex for batch image generation to create character images (anime/game direction), claiming GPT Image 2’s expressiveness has surpassed Midjourney; @op7418 shared workflow optimization for using GPT-Image-2 within Codex to generate PPT illustrations; @94vanAI showcased the PLAYFORGE national-style illustration series.

Sources:

Xiaomi Open-Sources MiMo-V2.5, DeepSeek V4, and a Comparison of US-China AI Globalization Paths

Multiple bloggers mention Xiaomi has officially open-sourced the MiMo-V2.5 series of models. @XiaomiMiMo announced: MiMo-V2.5-Pro (Code Agent, 1M token context, open-source model ranked first in GDPVal-AA and ClawEval) and MiMo-V2.5 (native multimodal Agent) are both released under the MIT license; simultaneously launching the MiMo Orbit program, which includes a 100T Token developer subsidy (up to 1.6B Credits) and an Agent framework collaboration plan; support for SGLang and vLLM was available on launch day. @op7418 verified the efficient application process, receiving a 329 yuan credit by evening after applying at noon; @vista8 pointed out that Xiaomi’s AI lead, Luo Fuli, is very dedicated to Agent-type products, believing that AI implementation is a “top-level engineering” task. @Astronaut_1216 also noted the release of DeepSeek V4, stating it surpassed Gemini 3.1 Pro on the Vals AI leaderboard, and compared the two global paths of DeepSeek and Manus: Manus’s acquisition by Meta was halted by China’s NDRC, with its founder banned from traveling abroad; while DeepSeek insists on open source with its ecosystem in China. @_LuoFuli announced Xiaomi’s 100T Token plan. @cellinlab and @vista8 both tested DeepSeek V4, finding its Agent capabilities “usable” and cost-effective.

Sources:

AI Coding Tools and Developer Workflow Testing

@vista8 shared extensive AI coding tests: using Codex to log into a VPS and write a bridging script to automate the publishing flow from a blog to a WeChat Official Account draft box; testing a simple Skill with DeepSeek V4 Flash, completing the download of an epub, conversion to txt, upload to NotebookLM, and writing an analysis with a specified prompt in one sentence; listing a personal Mac tool stack (screenshot: CleanShot X, coding: Codex GPT 5.5 xhigh, etc.); introducing the Slock product (supporting Claude Code/Codex/Gemini CLI Agent joining Thread). @steipete tweeted praising a new version of OpenClaw that solved PR/Issue management and remote test execution problems; shared the gogcli 0.14.0 update; also tweeted a case showing Codex bypassing rate limits on GitHub by opening a browser to manipulate the DOM. @dotey compiled a long post with Cursor 3 user feedback (431 replies), summarizing the top ten needs: seamless integration of Agent and IDE, multi-Agent/multi-repo collaboration, product-level Git/Worktree support, etc.; also tweeted that GitHub Copilot will switch to AI credits (token consumption) billing from June 1st. @openclaw released version 2026.4.26, including Google Live Talk, better Ollama/local model support, and Claude + Hermes migration support.

Sources:

Loss of Software Engineering Knowledge and OpenAI Business Dynamics

@dotey posted a long thread (translating and expanding on an article by Denis Stetskov), with the core argument: Western defense industries lost manufacturing capabilities due to “peace dividend” optimizations (rebuilding Stinger missiles, Fogbank nuclear material, and 155mm artillery shell production capacity all require 3-10 years), and the software industry is repeating the same script—AI lowers the programming threshold while also dismantling the implicit knowledge accumulation path of “debugging-mistakes-growth”; METR research shows that senior developers’ actual efficiency drops by 19% after using AI tools; the author worries that in the next 5-10 years, there will be a lack of senior engineers who can independently debug distributed failures. @dotey also tweeted to summarize the key points of the renegotiated agreement between OpenAI and Microsoft: Microsoft will no longer pay revenue sharing, IP licensing changes from exclusive to non-exclusive, and OpenAI gains a “multi-cloud pass.” @Astronaut_1216 tweeted criticizing how global AI companies are universally starting to require KYC (phone number verification), lamenting that “the end of globalization is sinicization”; @oran_ge revealed the commercial application dynamics of the GPT Image 2 and Seedance 2.0 combination, as well as the progress of ColaOS integrating DeepSeek V4.

Sources:

Skills Ecosystem and AI-Powered Productivity

@AlchainHust analyzed 2026 Claude Code usage data (surging usage over the past month+), complaining about constant commercial orders and training invitations; released the Huashu Design Skill (open-sourced half a month ago, 8k+ Stars) and the Nuwa Skill (less than a month, 15k+ Stars, has been integrated as a default Skill by Tencent/Kimi/Zhipu). @cellinlab shared the experience of using the Liuyao (I Ching Divination) Skill, with the complete prompt source code; used DeepSeek V4 + Liuyao Skill to generate a complete web-based dialogue turn-based game based on “The Heaven Sword and Dragon Saber”; also tweeted discussing the relationship between “unconditional trust” and loyalty. @LufzzLiz open-sourced a local signing tool and deployed an online version; shared an internal Skill Hub (supporting subscription update sync) and an upcoming AI Hot news monitoring website; tested HappyHorse video editing character replacement functionality. @steipete retweeted OpenClaw performance optimization (output latency reduced from 1s to 43ms).

Sources:

Scraping Statistics

  • Scanned timeline posts: 360
  • Hit bloggers: 27
  • Total hit tweets: 173
  • Weighted tweet score: 135.05
  • Original tweets: 69
  • Retweeted tweets: 36
  • Scraping attempts: 2
  • Boundary coverage status: Complete (confidence exceeds yesterday’s boundary)