Xiaoliu BOT
中文EN

X Platform May 1 AI Briefing: Codex Adds Goal-Driven Mode, Grok 4.3 Release Improves Cost-Effectiveness, AI Rapid Game Development Sparks Discussion

Codex Major Update: /goal Command, External Service Integration & Computer Use Speedup

Multiple bloggers mentioned that OpenAI Codex has received a major update. The most notable addition is the new `/goal` command (also known as Ralph Loop), where after a user sets a goal, Codex will continue executing across multiple rounds until completion without manual driving. This feature is enabled via `goals = true` configuration in CLI version 0.128.0, with the desktop app not yet following suit. Sam Altman personally promoted it, calling it a “big upgrade” and encouraging everyone to try non-programming Computer Use tasks. Guizang supplemented that Codex can now connect to Slack, email, and Google Drive to obtain content, and add information to Google Calendar; Baoyu shared the usage of the third-party tool CodexPotter – writing goals into MAIN.md and repeatedly starting new sessions for verification-style development; Xiangyang Qiaomu suggested from a practical perspective: simultaneously generating design mockups and developing an application in the same session reduces effectiveness, and should be split into two sessions for “generating design mockups” and “pixel-perfect restoration” respectively. Additionally, Codex Desktop has added a new `/side` sidebar chat feature, which Baoyu considers equivalent to Claude Code’s `/btw`.

Sources:

AI Game Development: Developing a Complete Card Game in an Afternoon Sparks Discussion

Multiple bloggers mentioned that Guizang used Codex to develop a complete card game similar to “Slay the Spire” in one afternoon, with both the code and client open-sourced. Guizang shared what shocked him most about Codex’s capabilities: faced with thousands of asset images, Codex automatically stitched the images from each file into one large image and labeled the filenames, enabling efficient asset retrieval. Orange AI commented, “This used to require a studio one or two months, now one person can do it in one afternoon”; Xiangyang Qiaomu also expressed being shocked. However, some people in the gaming industry questioned this, arguing that AI cannot handle gameplay and mechanics. Guizang responded in a series of tweets: he was only demonstrating the possibilities of AI-assisted development and never claimed AI could independently complete AAA-level games, and predicted that community-based game platforms similar to Roblox will thrive because of AI – the prototypes of Dota and PUBG evolved from mods, and AI could also give rise to the next big hit genre. Additionally, Guizang also discovered a Chinese-style horror mystery game with AI involvement called “Strange Rites”.

Sources:

GPT Image 2 Viral Prompts: MS Paint Doodle Style & New Creative Uses

Multiple bloggers mentioned that creative prompts for GPT Image 2 are going viral on X and Douyin. The most popular is the “MS Paint doodle style” prompt, which asks to redraw images in the clumsiest, scribbliest art style, as if doodled with a mouse in a painting program, with results that are both funny and highly recognizable. Baoyu translated and shared the Chinese version of the prompt; Guizang supplemented another prompt also popular on Douyin – “hand-drawn style annotations” – adding handwritten-style notes to every object in a photo, with a cute Japanese aesthetic and diary-like tone. Cell shared attempts to use comic-panel style prompts to bypass face restrictions, as well as using GPT Image 2 to create a “7-Day English Speaking Guide”; 94vanAI proposed that “semi-automatic word-filling image generation prompts” are the optimal solution, sharing a versatile commercial poster prompt; Gu Yi showcased a style-clash experiment combining anodized aluminum gradients, gacha capsules, and Win95 clip art.

Sources:

Grok 4.3 Released: Lower Price, Stronger Agent Capabilities & Imagine Agent Mode

Some bloggers mentioned that xAI released Grok 4.3, achieving a significant boost in Agent capabilities at a lower price than Grok 4.2 – on the Artificial Analysis GDPval-AA, the ELO score increased by 321 points to 1500, making it stand out in cost-effectiveness compared to other top models. Input token prices decreased by 37.5%, and output token prices decreased by 58.3%. Elon Musk promoted Grok Imagine’s Agent Mode (Beta), which users have called the best current AI image and video editing tool. Additionally, xAI opened the Voice Cloning API console in the United States. Elon Musk also shared a demonstration of Grok 4.3’s frontend design capabilities – building a complete website with just a single prompt.

Sources:

AI Academic Frontiers: Image Generation Survey, DeepSeek’s New Paper & Eywa Cross-Modal Interface

Multiple bloggers shared the latest academic achievements in the AI field. Xiangyang Qiaomu interpreted a 2026 AI image generation survey paper, distilling several key insights: data quality is more critical than quantity; mixing a small amount of AI-generated images in training data severely harms model quality; distillation is a must for commercial deployment; and the core gap between open-source and closed-source lies not in the model but in the system architecture. He also interpreted DeepSeek’s latest paper, which teaches AI to “think while pointing at images,” filtering nearly 100,000 data entries down to 30,000 during data cleaning, and the maze problem design in the training data creation process is impressive. Additionally, the Eywa paper from UIUC proposes a solution to build interfaces for language models and specialized foundation models, having Chronos handle time-series forecasting, TabPFN process tabular data, and language models manage scheduling, connectable in the short term via MCP, with uncertainty remaining in the long term whether language models can reach the level of specialized models.

Sources:

Structural Thinking on Products and Agents: Riverbed Theory, Entropy & Open Systems

Some bloggers mentioned systematic thinking about product and Agent design. Orange AI published a long article elaborating the concept of “structure as riverbed”: making products, Agents, large models, companies, interpersonal relationships are essentially all about carving riverbeds – large models are water, harness is the riverbed, and water quality depends on the riverbed’s shape. He criticized that letting mid-level management lead transformation in big companies’ AI shifts is “adding stones to the riverbed,” and without changing the organization, everything is ineffective. He also drew from Schrödinger’s “negative entropy is information,” pointing out that once large model weights are fixed, they become a closed system where outputs tend toward mediocrity, but injecting context instantly turns them into an open system, where context is negative entropy and vitality; he also criticized the overuse of “entropy increase explains everything” – entropy increase only applies to isolated systems, while humans and the Earth are open systems. Baoyu discussed two types of interaction paradigms for Agent products: agent-centric (conversation list + code view) and human-centric (traditional software + sidebar assistance), believing that product positioning must be clarified first before doing interaction design.

Sources:

Sam Altman Envisions AI Future: Augmenting Humans, Not Replacing Them

Some bloggers mentioned that Sam Altman posted multiple tweets in one day elaborating on OpenAI’s vision. He hopes to build tools that “augment and uplift humans” rather than entities that replace humans; believes that in the long run, the “end of work” theory is wrong, and many people will be busier and more fulfilled than ever, despite turbulence during the transition period; and painted a future picture: those who want to work hard will have highly rewarding things to do, and those who don’t want to work hard can still live a comfortable life.

Sources:

OpenClaw Update: Group Chat Experience Optimization & Security Hardening

Some bloggers mentioned that OpenClaw released version 2026.4.29, focusing on improving the group chat experience – subsequent instructions can guide the current run at model boundaries instead of reinitiating tasks, and the memory system also enhanced features like metadata, traceability, aliases, and relationship graphs. Peter Steinberger recommended group chat users try again and suggested switching to the codex harness for better performance. OpenClaw also published an article discussing the recent wave of security advisories and joint hardening work with NVIDIA, OpenAI, Microsoft, GitHub, and other companies. Steinberger also released Crabbox 0.1.0, a tool for remote Linux testing, and discovered an accidentally left Claude.md file in the Apple Support App.

Sources:

Statistics

  • Timeline entries scanned: 240
  • Number of bloggers hit: 23
  • Total tweets hit: 146
  • Weighted tweet score: 109.25
  • Original tweets: 56
  • RT tweets: 37
  • Scrape attempts: 1
  • Boundary coverage status: Fully covered