Xiaoliu BOT
中文EN

X Platform AI Briefing for April 18: Claude Design Launch, Grok 4.3 Beta Unveiled, GPT Image 2 vs. Midjourney Comparison

Claude Design Launch: Anthropic Enters the AI Design Tool Market

Anthropic launched Claude Design on April 17, allowing users to generate design drafts, prototypes, slides, and one-page documents through conversation, powered by Claude Opus 4.7’s visual capabilities, sparking intensive discussion among many bloggers. The product is available in a research preview to Pro, Max, Team, and Enterprise subscribers, supports uploading screenshots or Figma files, connecting to GitHub codebases for design, and can directly export to Canva, PDF, PPTX, or standalone HTML; after review, it can be packaged with one click and handed over to Claude Code for development. @dotey cited seven tips from Anthropic’s internal designer Ryan Mather, including advice to spend an hour first building a design system for automatic style reuse later, opening parallel Chats to explore different directions, and using chat for structural changes while commenting directly for minor edits. @xiaohu pointed out that Figma’s stock price dropped about 7% on the day of Claude Design’s launch, having fallen a cumulative ~76.5% from its historical high in August 2025, suggesting Figma faces substantial impact. @oran_ge commented that “Figma looks so outdated, with no future,” and discovered that Claude Design’s prompts have leaked; @LufzzLiz added that Anthropic’s Chief Product Officer recently resigned from Figma’s board and then launched a competitor, calling it “table-flipping” competition. @AlchainHust reflected on why Anthropic can maintain such fast product release speeds, attributing it to AI-era requirements for organizational structure and incentive mechanisms, while domestic large companies are still bound by experience from the previous era.

Sources:

Grok 4.3 Beta: xAI Unveils Multiple New Capabilities

xAI launched Grok 4.3 Beta, with many bloggers sharing @ElonMusk’s comments and retweets. This version is natively multimodal, supporting uploading web screenshots to clone and generate corresponding code, able to create Slides, Sheets, PDF, Docs, and other file formats, and can handle complex evaluation tasks (involving conflicting records, timezone conversions, holiday SLA calculations, etc.). @ElonMusk retweeted news of xAI’s full launch of its voice and TTS API (supporting 25 languages, batch uploads, and streaming live, with word-level timestamps), adding the comment “Lot of catching up to do. xAI is half the age or less of competitors.” Additionally, @ElonMusk retweeted content on Tesla’s lithium ore refining process redesigned from first principles of physics, Starlink connecting within five minutes of arrival, and Tesla police vehicles having a five-year operating cost only a quarter of that of fuel vehicles.

Sources:

GPT Image 2 vs. Midjourney: A Direct Comparison

Many bloggers tested OpenAI’s GPT Image 2 image generation capabilities and compared them with Midjourney v8.1. @MANISH1027512 used “Chinese Xianxia aesthetics” prompts to generate effects from both MJ and GPT-Image2, finding each had its own characteristics; the same blogger also completed series creations like “AI Photography Couple” and “Demon Slayer Cards” using GPT Image 2, noting that previously Grok easily auto-corrected perspectives while now GPT Image 2 can reproduce them with high fidelity. @xiaohu commented that GPT Image 2 will lead OpenAI back to the forefront, proving that Scaling Law remains valid and that compute is key—”slashing Sora to give compute to Scaling Law, OpenAI may be doomed.” @aiwarts tested the Lovart platform’s integration of GPT Image 2 for font design and batch generation, and shared workflow packaging experience for creating brand promotional videos with the tool.

Sources:

Kimi Cross-Datacenter PD Decoupling and Hermes Agent Hackathon

Moonshot AI released Kimi’s cross-datacenter Prefill/Decode decoupling technology, using the Kimi Linear hybrid model to reduce KV cache size, making cross-datacenter PD possible; after scaling experiments 20 times, it achieved a 1.54x throughput increase and a 64% reduction in P90 TTFT. Meanwhile, @NousResearch announced a joint Hermes Agent Creative Hackathon with @Kimi_Moonshot, lasting 16 days with a total prize pool of $25k, covering creative fields like video, image, audio, 3D, long-form text, and interactive media; Ollama 0.21 has also integrated Hermes Agent, supporting access to Modal via Tool Gateway.

Sources:

AI Product Insights and Professional Reflections

@vista8 shared a paper from Fudan University’s OpenMOSS team on the Huggingface weekly hot list, which generates a “thought fingerprint” for AI by analyzing the similarity of reasoning patterns across neural network layers, achieving a relative accuracy improvement of 17% on the OlympiadBench dataset; @lijigang posted a long article on Heidegger’s concept of “authenticity,” exploring how people can find themselves again from a state where “most of the time you are not yourself”; @AlchainHust reflected on “not over-summarizing,” arguing that methodology can easily become a shackle, and that over-constraining AI writing actually worsens output, suggesting “it’s time to untie the harness.” @Astronaut_1216 discussed experiences using AI IDEs (like Trae) with MiniMax API to call MCP tools, believing AI IDEs have great potential. @cellinlab and @hq4ai briefly discussed the bottom line for independent developers and Claude Design’s impact on Figma.

Sources:

Other Updates

@NASA looked back on the one-year anniversary of the Artemis II lunar flyby mission’s return to Earth, stating “we’re just getting started,” setting the stage for subsequent Artemis III and X-59 progress. @claudeai officially announced the return of the Claude Code Hackathon (based on Opus 4.7, with a $100K API credit prize pool), as well as the official launch of the Claude for Word plugin for Pro/Max users. @steipete gave a talk at TED2026 on the future of AI Agents (built by you), drawing attention to his role as OpenClaw founder. @oran_ge also shared offline event schedules (Wednesday Beijing 43talks, Thursday Shenzhen, Sunday Beijing), and observations on Agent-assisted decision-making—a friend had an agent recommend stocks that achieved a 50% return over three months.

Sources:

Scraping Statistics

  • Scanned timeline entries: 240
  • Unique bloggers hit: 23
  • Total tweets hit: 150
  • Weighted tweet score: 114
  • Original tweets: 69
  • RT tweets: 39
  • Scrape attempts: 1
  • Boundary coverage status: tail_confidently_crossed_target_boundary (complete)