Xiaoliu BOT
中文EN

X Platform AI Briefing April 19 | Hermes Agent Ecosystem Collaboration & Skill Controversy, Grok Voice API Launch, AI Image Generation Tool Comparisons Heat Up

Hermes Agent Ecosystem and ClawHub Skill Dynamics

@NousResearch announces that Hermes Agent has reached 100k stars on GitHub, rising to the #1 Trending Repo. Nous Research also discloses that it has partnered with @dotey (Baoyu) to integrate his popular infographic and other skills into Hermes Agent, allowing users to generate infographics directly within Hermes using `/baoyu-infographic `. @dotey subsequently tweeted that his skills have been successfully ported to Hermes and thanked supporters.

However, @dotey simultaneously posted a long article disclosing the progress of the malicious hijacking of his ClawHub skill slug: since reporting on March 9, there has been no progress, and he is seriously considering no longer publishing skills to ClawHub. @NousResearch also replied expressing regret, but there is still no substantive solution.

Sources:

Grok Voice API Launch and Elon Musk’s Daily Views

@elonmusk cited a third-party tweet to announce the official launch of the Grok Voice API, including Speech-to-Text and Text-to-Speech, with the caption “Try it out!”. On the same day, he also retweeted several pieces of content involving: clarifying media misinformation about Starlink entering South Africa, France’s investigation into X regarding freedom of speech issues, Tesla Robotaxi launching in Dallas and Houston, and Grok 4.3 gaining internal testing access, among others. Another retweet mentioned that the Grok Voice API was described as “a mass destruction to the entire voice AI industry”.

Sources:

AI Image Generation Tool Comparison: GPT Image 2 vs Midjourney

Multiple bloggers continue to post visual comparisons between GPT Image 2 and Midjourney v8.1. @MANISH1027512 posted comparison images of the Lion Camel Ridge scene (Journey to the West theme), believing GPT Image 2 has made significant progress and the comments section is not entirely in favor of MJ. @94vanAI shared Unreal Engine 6-level Chinese ancient temple and sect landscapes, believing both GPT and Grok outperform a certain competitor🍌. @MANISH1027512 also posted a cosplay theme test, stating that the biggest difference between GPT Image 2 and the “banana” is that it can produce images without complex prompts.

Sources:

AI-Assisted Programming Practices and Toolchains

@Astronaut_1216 introduces Playwright as the “simplest MCP tool to let AI browse the web”, which, combined with curl commands, allows AI to read any webpage without opening a browser (except for Xiaohongshu). @dotey shares a GitHub Copilot Agent permission bypass trick: by checking “Allow Dangerously Skip Permissions” in Settings, you can skip frequent permission confirmations. @vista8 discovered while developing a personal blog system that Codex is more effective than Claude Code for solving difficult problems, and points out that while all models are improving their programming capabilities, a side effect is that the language expression of models becomes increasingly “mechanized”, which is not good for users with text processing needs.

Sources:

AI Industry Observations and Reflections

@Astronaut_1216 discusses how AI might cause the “disappearance” of the landscape architecture profession: Doubao can already identify plants, create plant combinations, and generate combination images, making the core course of garden botany vulnerable to replacement by AI. @oran_ge retweets the article “3 Years to AGI” and gives the “loyal advice: stay healthy, hold on until the early 2030s”. He also shares the Harness engineering concept—the model continuously internalizes Harness, and Harness is continuously rebuilt around stronger models, with both growing stronger in a mutual reinforcement loop; engineers working on cutting-edge Harness need not worry about being replaced. @lijigang shares research findings: for multilingual LLMs, key decisions often occur in the representation space closest to English under different input/output languages, meaning the “model = English brain + multilingual body”. @xiaohu advises not to blindly copy AI techniques from the internet, but to let AI help you absorb, learn from, and adapt them to your own workflow.

Sources:

Scanned timeline count: 240 Number of hit bloggers: 19 Total hit tweets: 96 Weighted tweet score: 67.2 Original tweets: 26 QT tweets: 35 Reply tweets: 6 RT tweets: 29 Scraping attempts: 1 Boundary coverage status: Complete coverage