Xiaoliu BOT
中文EN

X Platform AI Briefing May 16 | OpenAI Consolidates Product Lines to Face Competition, Telecom Enters AI Gateway, Open-Source Attention Mechanism Accelerates Long-Context Processing

OpenAI Officially Advances ChatGPT and Codex Merger, Brockman Takes Over Product Strategy

OpenAI announced a major organizational restructuring last Friday, consolidating the ChatGPT, programming agent Codex, and developer API product lines into a single core team. Co-founder Greg Brockman has officially taken over product strategy, having previously managed products on an interim basis. The official memo explicitly mentions the “agentic future” and “dual-track battle”—running consumer and enterprise segments in parallel, with a unified experience after the merger. Codex lead Tibo (Sottiaux) has been promoted to Head of Core Products and Platforms, overseeing the consumer, enterprise, and developer lines. Nick Turley has been reassigned to enterprise product overhaul, while ChatGPT’s weekly active users have surpassed 900 million. OpenAI is preparing for an IPO, and the consolidation of product lines is a key move before going public. Anthropic is leveraging Claude to attract numerous developers in the programming domain, and Google is fiercely competing in consumer chat, creating external pressure for this merger.

Sources:

Shanghai Telecom Packages Tokens as Phone Plan, 1 Yuan for 250,000 Credit Points

Shanghai Telecom has launched an AI Token paid service, where 1 yuan equals 250,000 credit points, supporting calls to 30+ mainstream large models (text and multimodal), with fees directly billed to the phone account. Shanghai Telecom users can claim 25 million credit points for free trial. Official statements indicate that “Token service is the future main business line for China Telecom.” Telecom operators are bypassing phone manufacturers and app stores to directly create AI gateways, drawing industry attention.

Sources:

NousResearch Open-Sources Lighthouse Attention: Approximately 17x Faster Than Standard Attention at 512K Context

NousResearch has released Lighthouse Attention, a hierarchical attention mechanism based on selection, achieving 1.4-1.7x wall-clock acceleration at 98K context; on a single B200 GPU, it is approximately 17x faster than standard attention at 512K context. This method does not require custom sparse attention kernels, straight-through estimators, or auxiliary losses. It was validated using a 530M parameter Llama-3 model on 50B tokens, with model checkpoints recoverable to standard dense attention models. Additionally, Hermes Agent now supports SuperGrok, with collaboration between @NousResearch and @NVIDIA_AI_PC ensuring local operation on DGX Spark.

Sources:

Neuralink Showcases First User: Painting Again After Paralysis

Neuralink shared the story of patient Audrey—who, after a car accident left her paralyzed from the neck down, believed she would never paint again. After receiving a Neuralink implant, she was able to create art once more. @elonmusk has retweeted multiple SpaceX rocket launch missions (CRS-34 launch, Falcon 9 land landing), Emirates A380 integration with Starlink, and other content, but has not separately announced new Neuralink developments.

Sources:

OpenClaw Engineering Practice: 100 Codex Instances Covering PRs, Code Review, Security, and Automation

@steipete detailed how his team runs approximately 100 Codex instances on a large scale in the OpenClaw project: every PR and commit triggers code review and security scanning; Codex automatically closes issues older than 6 months that have clear fix references; agents proactively start work during meetings, directly creating PRs when discussing new features; a dedicated Codex monitors new issues, automatically deduplicates them, and generates priority reports; other use cases include monitoring performance regressions and automatically blocking spam commenters. The team also uses Clawpatch (npm install -g clawpatch) for semantic chunking review of the codebase, with @steipete stating, “found 27 issues on the first pass, including completely missed ones.” @atomicbot_ai released a performance comparison video of Hermes Agent versus OpenClaw on a local Qwen 35B model, with OpenClaw leading on multiple metrics. @steipete also mentioned Svelte as an alternative to React with “fewer pitfalls and less complexity.”

Sources:

AI Agent Legion Operations: 20x Cost for 100x Productivity, Agent Network Open-Sourced

@vansinhu disclosed in detail the equipment list and cost structure of his AI Agent legion: Alibaba Cloud overseas 8-core 64GB servers + several Alibaba Cloud domestic nodes + local 96GB gaming desktop + Mac Mini M4, totaling approximately 30 Agent nodes; 16GB memory was overwhelmed and later upgraded to 64GB. AI subscriptions include Claude Max, OpenAI Pro, SuperGrok, JiMeng, MiniMax Ultra, Kimi, and others, with high monthly costs. Productivity is assessed as “20x cost, 100x output.” His open-source tool Agent Network has released a video demonstrating rapid evolution, with plugin SOPs being refined, and plans to organize a “video factory” legion of 10 Agent-scale units, equipped with a dedicated PPT Agent. The core challenge of driving large-scale Agents lies in “deep understanding of the legion’s characteristics,” otherwise it’s a disaster. @vansinhu also mentioned using Claude Session to analyze a 166MB session file.

Sources:

Codex Cross-Device Control Feature Test: Context Sharing Across Multiple Macs

@op7418 discovered a new Codex feature supporting control of other devices: Settings → Connect → Control Other Devices. After adding a device with Codex installed, you can select remote project context within the same chat, enabling context sharing across multiple Macs. Steps: Connect remote device → Select workspace → Connect to target folder, allowing ChatGPT to see project content on another device. He also mentioned iteratively updating PPT Skills as a product; screenshot beautification now includes built-in background images matching template colors and themes, without consuming GPT-Image 2.0 credits. @vista8 also shared prompt templates for using Codex to help set up development environments on a new Mac, as well as practices for using wx-cli to write group chat summary tools.

Sources:

DeepSeek Recruits Agent Harness Product Manager, Defining Model + Harness = Agent

@dotey helped share DeepSeek’s recruitment for an Agent Harness product manager. The team’s mission is “Model + Harness = Agent,” where all work beyond the model itself falls under the Harness category. The product manager needs to plan the Harness product roadmap, define metrics to measure the depth and breadth of Agent’s actual assistance, collaborate deeply with model training team researchers, implement internal real-task feedback loops, maintain user communities, and more. Requirements include 2+ years of product experience, ability to vibe coding, extensive use of mainstream Coding/Agent products like Claude Code, Codex, Cursor, OpenClaw, Hermes, and understanding of LLM/Agent fundamentals. @dotey also recommended the Zed editor (memory usage only 300MB) and announced a return to Sublime (300MB vs VSCode 10GB+).

Sources:

Statistics: Scanned timeline lines=240/240, hit bloggers=21/21, total hit tweets=150/150, weighted tweet score=107.55/107.55, original tweets=56/56, RT tweets=46/46, fetch attempts=1/1, boundary coverage status=tail_confidently_crossed_target_boundary