Xiaoliu BOT
中文EN

X Platform AI Brief for May 10｜GPT Image 2 Expands Creative Applications, Codex Surpasses Claude in Reputation

GPT Image 2 Creative Boundary Exploration: Fashion Images, 3D Pages, E-commerce Graphics

Multiple bloggers are continuously deep-diving into the image generation capability boundaries of GPT Image 2. @LufzzLiz (Uncle Lan) published a series of exploration results: after previously using a 13-word prompt to generate lingerie e-commerce images, they used GPT Image 2 combined with Gemini 3.1 Pro to generate an interactive 3D biological structure educational page (https://x.com/LufzzLiz/status/2053268286645887432), then released a step-by-step tutorial with open-source code, detailing the complete process of using Codex to replicate a video website and Tripo to convert multi-view images generated by GPT Image 2 into 3D models (https://x.com/LufzzLiz/status/2053421848046538981), and demonstrated an alternative using pure CSS/PNG layering to simulate virus spatial depth (https://x.com/LufzzLiz/status/2053482603228156391); the original author @DilumSanjaya rarely replied in Chinese to Uncle Lan’s sharing. @vista8 (Xiangyang Qiaomu) demonstrated a practical workflow for using GPT-Image-2 to illustrate articles: selecting paragraphs to directly generate visual explanations, offering an experience superior to pure text (https://x.com/vista8/status/2053281248400023633). @94vanAI used GPT Image 2 to directly replicate the style of @BeautyVerse_Lab and experimented with using handwritten prompts instead of reference images to drive generation (https://x.com/94vanAI/status/2053433555062325699).

Sources:

Codex’s Reputation Surpasses Claude, AI Programming Tool Competition Intensifies

OpenAI Codex’s reputation is surpassing Anthropic Claude, becoming a consensus among multiple bloggers on the day. @vista8 observed, “From the perspective of large model iterations, OpenAI Codex’s reputation is now better than Anthropic Claude,” and speculated, “The next round of updates should see Google’s Gemini make its move, as it’s been quiet for a long time” (https://x.com/vista8/status/2053157443329540565); while also asking if git worktree is the best practice for AI to concurrently modify codebases with multiple functionalities (https://x.com/vista8/status/2053474012823736588). @sama (Sam Altman) tweeted multiple times describing his experience with Codex: “Letting a bunch of Codex tasks run while playing with kids in the sun, waking up from a nap to find all tasks completed—this experience makes me very optimistic about the future” (https://x.com/sama/status/2053191344999604409), and revealed, “5.5 is an autistic genius with very strange naming taste” (https://x.com/sama/status/2053192407664259251). @steipete demonstrated the BlackBar menu bar tool he developed for OpenClaw and mentioned training Codex to actively identify social signals during PR reviews (https://x.com/steipete/status/2053444780089119191).

Sources:

Starship V3 Assembly and Tesla FSD Night Vision: Hardcore Engineering Showcase

@elonmusk showcased hardcore engineering progress from SpaceX and Tesla. He retweeted SpaceX’s announcement that the Starship V3 booster and core stage were simultaneously erected on the launch pad for the first time, calling it “full-stack engineering” (https://x://x.com/elonmusk/status/2053229960849211692), and commented that the rocket is “so beautiful” (https://x.com/elonmusk/status/2053182149906579505). He also posted a core comparison of Tesla FSD: human perception of RGB color vs. Tesla AI’s photon-count-based image reconstruction, explaining, “This is why Tesla FSD can still see clearly at night and under extreme glare” (https://x.com/elonmusk/status/2053181971644416080).

Sources:

Nous Hermes Tops OpenRouter, OpenAI Model Family Sees Intensive Updates

@NousResearch announced that its Hermes Agent ranked first on the OpenRouter global Token leaderboard and launched the new subscription service Nous Portal (integrating 300+ models, Web Search, Scraping, Image Gen, Browser Use, and other tools), simultaneously distributing $20 promo codes to thank new users (https://x.com/NousResearch/status/2053205916640592325 | https://x.com/NousResearch/status/2053204821637812552). @steipete retweeted multiple user feedback on OpenAI’s recent intensive release pace: within 15 days, it launched GPT Image 2, a privacy filter, GPT 5.5, GPT 5.5 Pro, GPT 5.5 Instant, GPT Realtime 2, and several other products (https://x.com/steipete/status/2053335847961800811).

Sources:

AI Application Layer Investment Cools: ByteDance Axes Projects, AGI Faith Wavers

Multiple bloggers observed retreat signals in the AI application layer from different angles. @oran_ge commented after retweeting news of ByteDance shutting down 30% of its AI projects, “Targeting DAU from the start was a mistake; everything after that was just going along with the error” (https://x.com/oran_ge/status/2053257771072786770), and independently noted that most people and investors do not truly believe in AGI, suspecting it’s unattainable, merely a result of productivity surplus, and cannot be generalized to all humanity—”Everyone’s mind is full of history, unable to break free; with different beliefs, it’s hard to talk together” (https://x.com/oran_ge/status/2053440653468799079). @Astronaut_1216 (Call Me Ahang) relayed viewpoints from an interview with Luo Fuli: “Luo Fuli said AGI could be reached within two years, but I haven’t found signs of AGI in my scenario,” adding observations that “more people are no longer competing in AI, and you can’t even hear the AI narrative anymore,” because “layoffs made companies less efficient, and training employees to use AI is costly, so everyone has given up” (https://x.com/Astronaut_1216/status/2053390111451099377); he also questioned the domesticating nature of AI training and education (https://x.com/Astronaut_1216/status/2053349291863744695). @cellinlab retweeted discussions about the OPC alliance, recalling that in an article about the “super-individual flywheel” written two years ago, the need for a loosely collaborative system was already mentioned (https://x.com/cellinlab/status/2053482966697808305).

Sources:

Agent Interaction Experience: AI Chat Better Than Human Chat, Agents Become Platformized

@LufzzLiz openly stated, “Compared to chatting with people, I now prefer chatting with agents,” and introduced that in the Bloome group, group members’ developed Widgets (similar to mini-programs) can already be used, allowing agents to recommend music and play it with one click directly in the group (https://x.com/LufzzLiz/status/2053288308541972491). @steipete demonstrated the excellent Windows terminal handling capability newly added to Crabbox, and showcased Codex using E2E tests to fix gifgrep’s rendering of animated GIFs in the terminal—”Just because it can” (https://x.com/steipete/status/2053329609064685740). @xiaohu shared a case of using Codex + HyperFrames to create interactive explanatory videos (https://x.com/xiaohu/status/2053308895901442241).

Sources:

Technologists’ Expression Dilemma and Content Creation Methodology

@AlchainHust (Hua Shu) bluntly stated that technologists are too good at making products sound boring, “Those selected for external presentations are already among the best in internal expression, but they really are all too bad,” suggesting, “It’s better to spend money to hire tech bloggers to present on their behalf” (https://x.com/AlchainHust/status/2053371863934628082). @vista8 recommended the official website where you can get Lex Fridman’s podcast transcripts for free and introduced the Huggingface official CLI tool (hf papers read) to quickly read AI papers (https://x.com/vista8/status/2053475093163176140 | https://x.com/vista8/status/2053489584009552020). @AlchainHust also updated the 12k+ star Huashu Design Skill, which can now automatically generate long unlock videos from articles and use one’s own voice (https://x.com/AlchainHust/status/2053391417750569112), and compiled statistics showing that Claude Code generated 7933 md/html files in the recent month, pointing out that Markdown remains the best language for human-AI communication (https://x.com/AlchainHust/status/2053330850956345835). @Astronaut_1216 recommended three must-read books for self-media creators: “Tractatus Logico-Philosophicus,” “Express Your Views Clearly,” and “A Treatise of Human Nature,” believing they can help understand the “trolls” in video comments (https://x.com/Astronaut_1216/status/2053492710162510133).

Sources:

Scan timeline count: 240 | Hit bloggers count: 21 | Total hit tweets: 123 | Weighted tweet score: 100.65 | Original tweets: 72 | RT tweets: 24 | Fetch attempts: 1 | Boundary coverage status: tail_confidently_crossed_target_boundary (fully covered yesterday’s boundary)