Xiaoliu BOT
中文EN

X Platform AI Briefing for May 3: Altman’s Tribute to Brockman, Software 3.0 & Brain Prediction Framework, and HTML-in-Canvas Technology Analysis

Yesterday, the Following timeline scanned 240 tweets, hitting 107 tweets from 21 bloggers (weighted score 72.35), including 38 original tweets, 27 quoted tweets, 2 reply tweets, and 40 retweets. The crawling attempt count was 1, and the boundary coverage status is: Full coverage, confidently crossed yesterday’s boundary. Below is a summary organized by topic.

Sam Altman Publicly Tributes Greg Brockman

Multiple bloggers mentioned that Sam Altman unusually posted two consecutive tweets highly praising OpenAI co-founder Greg Brockman, stating “cannot imagine OpenAI succeeding without Greg,” and recalled their decade-long collaboration, pointing out that his earlier post about Greg “did not sufficiently highlight his technical talent and unwavering determination.” Sam Altman himself also retweeted a post from @thsottiaux, further affirming the importance of this activity.

Sources:

向阳乔木 on Software 3.0, Brain Prediction Framework, and Mona Lisa Effect

Multiple bloggers mentioned that 向阳乔木 (@vista8) densely posted several in-depth articles on the same day, covering three different areas:

Software 3.0: Citing Andrei Karpathy’s recent Sequoia interview, it divides software development into three stages—1.0 where humans write code, 2.0 where neural network weights are trained with data, and 3.0 where the core leverage of programming becomes prompts and context control.

Brain Predictive Processing Framework: It introduces the 2016 book “Surfing Uncertainty” by Andy Clark, which built upon early neural network work by Geoffrey Hinton and others to construct the “brain predictive processing framework,” and uses this framework to uniformly explain phenomena such as memory, imagination, dreams, curiosity, schizophrenia, and autism.

Mona Lisa Effect: It references the theft of the Mona Lisa—this painting lay unnoticed in the Louvre for 400 years, was stolen by an Italian repair worker in 1911, the theft caused a global sensation, Italians hailed the thief as a patriotic hero, and from then on it became “a symbol of Western culture itself,” demonstrating how accidental events can change cultural fate through media amplification.

Additionally, 向阳乔木 mentioned that “someone” believes future computer architecture will completely invert: neural networks will become the host process controlling everything, with CPUs only serving as co-processors they occasionally call upon.

Sources:

HTML-in-Canvas Technology Analysis: DOM-Level Shaders and 3D Space Real UI

A blogger mentioned that 归藏 (@op7418) posted three in-depth articles analyzing the new experimental Chrome Canary feature HTML-in-Canvas. Its core value is to solve the pain point that Canvas has historically been just a “dumb pixel canvas”—rendering HTML/CSS in Canvas causes loss of browser-native accessibility support, text selection, multi-language layout, and click detection capabilities. HTML-in-Canvas uses three new primitives (the layoutsubtree attribute, drawElementImage() method, and onpaint event) to allow HTML elements to participate in DOM layout as Canvas child nodes while maintaining GPU pixel-level special effect rendering capabilities. The blogger’s analysis identified three key application scenarios: low-level engine division of labor (HTML handles semantic interaction, Canvas/GPU handles pixel rendering); DOM-level shaders (WebGL fragment shaders can be applied to HTML elements, producing effects like liquid distortion and scan line cuts during typing, while forms remain fully interactive); and attaching complete HTML web pages as textures to 3D rotating cube surfaces while still being clickable and interactive. The blogger also noted that this feature is currently still in the Developer Trial stage and requires manually enabling a flag in Chrome Canary.

Sources:

Codex Pets Feature Continues to Be Hot: Multiple Bloggers Share Custom Pets and Tutorials

Multiple bloggers mentioned that the Codex Pets feature (AI desktop pets summoned via /pet or /hatch) continues to generate community buzz. OpenAIDevs launched a Codex Pet image-sharing campaign, selecting 10 of the most popular works to gift 30-day ChatGPT Pro memberships. @LufzzLiz posted a detailed tutorial on developing custom pets using Codex, using “Angry Dario Amodei” as an example, requiring only Codex plus 9 animation state atlases, and specifically noted “codex with gpt-image-2 is simply invincible.” @steipete also referenced the OpenClaw 2026.5.2 release announcement, mentioning that this version significantly optimized npm installation and plugin systems, moving most functions into extensions to reduce package size. @dotey forwarded a mention that Open Design has open-sourced the most complete version of the codex pets code, with 100+ pets launched simultaneously, supporting avatars like @sama, Dario, @steipete, as well as various sprites and goblins.

Sources:

李继刚 Daily Paper Reading: Latent Space—The “Thinking Language” of Machines Doesn’t Have to Be Human Speech

A blogger mentioned that 李继刚 (@lijigang) posted a deep paper reading note, analyzing the paper “The Latent Space: Foundation, Evolution, Mechanism, Ability, and Outlook.” The paper points out that the current “thinking while writing” paradigm of reasoning models has four major bottlenecks: language redundancy (continuous vectors are lengthened by grammatical fillers), discrete bottleneck (a 30,000 token dictionary limits the variety of internal states), serial inefficiency (only one token can be output at a time), and semantic loss (information is lost with each compression step). In 2025, multiple technical paths (Coconut, latent CoT, recurrent latent attention) pointed to the same direction—allowing the model’s “inner monologue” to be completed in vector space without externalization through tokens. The paper uses a Mechanism × Ability two-dimensional coordinate system to systematically classify 2025 related work, with the horizontal axis being “how the mouth is untied” (changing architecture/representation/computation mode/optimization strategy) and the vertical axis being “what can be done once the mouth is untied” (reasoning/planning/world models/multi-modal perception/long-term memory/multi-agent collaboration/robot control). The blogger summarized: this wave of latent space revival is the inflection point for models changing from “copying machines” back to “thinking machines.”

Sources:

Creator Expression vs. Traffic Mindset: CellinLab Discusses the Essence of Content Creation

A blogger mentioned that Cell细胞 (@cellinlab) shared a creative methodology: create from the perspective of expression, not attraction; should expose one’s own experiences and feelings to the utmost, rather than designing hooks or crafting marketing copy. The same blogger also proposed an interesting product idea: for new streamers, the common dilemma of having no one asking questions in the live room—he suggested making a “question assistant” to guide interaction. Additionally, an interesting discussion about GitHub circulated in the community—someone asked “what if xxhub (implicitly referring to GitHub being redesigned after being acquired by a certain party) designed the website,” accompanied by an image of a website designed in the style of a traditional Chinese medicine shop, and CellinLab forwarded it with the title “What if Github were a traditional Chinese medicine shop…”

Sources:

OpenClaw 2026.5.2 Release: Grok 4.3, Plugin System Refactor, and Multi-Platform Fixes

A blogger mentioned that the official OpenClaw account released the 2026.5.2 version update. Core changes include: upgrading to xAI Grok 4.3; more robust plugin installation and update processes; streamlined Gateway and agent hot paths; multiple fixes for Discord, Slack, Telegram, and WhatsApp; and experience optimizations for TTS, Realtime, web search, and voice calls. The official slogan is “Less drama. More uptime.” @steipete additionally mentioned that this version fixed dependency issues and performance sluggishness when installing via npm, and moved most functions into extensions to reduce package size.

Sources:

小互 Discusses Anthropic vs. OpenAI Product Philosophy

A blogger mentioned that 小互 (@xiaohu) posted an observation: OpenAI is led by a group of product managers, while Anthropic is led by a group of engineers—so some of Anthropic’s innovations can always be copied by OpenAI and remodeled for a better experience.

Sources:

Comprehensive Briefing

@ElonMusk continued to be highly active, hitting 30 tweets yesterday (3 original, 10 quoted, 17 retweets). Original content mainly consisted of video shares (posting two video tweets in the evening, with “Banger” as the comment); quoted content involved commenting on or retweeting content from GadSaad, SpaceX, @XFreeze, @MikeBenzCyber, and others; retweet content mainly focused on SpaceX Falcon 9 missions. Considering the relatively dispersed information density of his original content, this summary did not give it a dedicated topic section.

@Astronaut_1216 posted an analysis of the Claude Code relay station business, arguing it is not a good business—ensuring fidelity requires backend maintenance, account purchasing, KYC, and other costs, and requires draining the user lifecycle in an instant to be profitable, pointing out that those who can run relay stations are either super traffic nodes, have B2B independent businesses, or have pyramid scheme and grassroots marketing teams.

@oran_ge published the article “From the Mind-Bending Book GEB to Agent’s Self-Awareness” and synchronized the detailed notes link.

@94vanAI posted a discussion on what is most important in the AIGC era—believing that the storytelling nature of content is more valuable than repetitive AI short videos, and showed a short video remake of 80/90s style Naruto retro using GPT-image2 + Grok video.

@MANISH1027512 continued to post CodeX-related content, introducing multiple skills in the vsc-skills open-source repository (including codex image auto-archiving, 260 sub-style explorers, AI virtual couple travel vlog workflow).

@Kana_Momonogi posted Golden Week reflections and a work photo in front of a microphone.

Sources:

Crawl Statistics

  • Tweets scanned: 240
  • Bloggers hit: 21
  • Total tweets hit: 107
  • Weighted tweet score: 72.35
  • Original tweets: 38
  • Retweet tweets: 40
  • Crawl attempts: 1
  • Boundary coverage status: Full coverage (tail_confidently_crossed_target_boundary)

Email sending failed: EOF occurred in violation of protocol (_ssl.c:1129)