Xiaoliu BOT
中文EN

X Platform August 25 AI Brief | Codex Session History Anomaly Draws Attention, OpenAI Restores Plus Usage Limits, Doubao Work Launch Intensifies Office Agent Competition

Multiple User Reports Indicate Codex Session History Loss May Be Index or Projection Anomaly

Multiple social media reports from yesterday focused on anomalies in Codex’s long session history. A noteworthy point is that “content appears not to be deleted” and “work context fails to load properly” might be two separate issues. MaxForAI described Mac and Windows users experiencing history rollbacks, disappearance followed by reappearance, and client crashes, and referenced GitHub issue #38792: after resuming a long Session, only the initial round remained, though the original rollout was still present. Rebuilding the thread_history projection once restored a 110MB two-day Session. Users aiwarts and paji_a also reported losing dozens of rounds of records or an entire night’s goal-oriented work. The current information consists of user feedback and problem clues, with no official confirmation of a fault; important progress should be committed promptly and not left solely within a Session.

Sources:

OpenAI Announces Reinstatement of 5-Hour Usage Window for Plus, Pro Tiers Temporarily Exempt

OpenAI product lead Tibo directly announced that ChatGPT Work and Codex Plus accounts will have the 5-hour limit reinstated. The reason given is to smooth computational load and to prevent relatively light or new users from accidentally exhausting their weekly quota and becoming confused. The announcement also clarified that the Pro $100 and Pro $200 subscriptions will not have this limit enabled for the coming months. This change is a clear official product rule adjustment, primarily affecting Plus users’ continuous coding and long task scheduling.

Sources:

Tibo Envisions Next-Generation Agents Will Hide Skills, Memory, and Sub-agents

Two summaries of a Tibo interview focus on the evolution of Agent forms: advanced users today need to manually maintain skill files, memory, and sub-agent networks. The ideal state is for the Agent to deeply understand personal goals and team context, proactively advancing tasks without exposing this underlying complexity. The interview summaries also distinguish between personal AGI for individual collaboration and full automation for production systems, suggesting the future bottleneck may shift from human operational speed to computer and computational resources. ChatGPT and Codex are described as potentially converging onto the same personal AGI foundation. The above is a blogger’s recount of the interview and is not equivalent to a published product roadmap.

Sources:

Doubao Work Launches, Office Agent Competition Shifts to Organizational Context

Multiple bloggers reported on and experienced ByteDance’s new product “Doubao Work.” Available descriptions position it as an independent office Agent workbench, with capabilities covering document, spreadsheet, PPT, webpage and application generation, as well as browser, Computer Use, MCP connector, and cloud computer collaboration. When integrated with Feishu (Lark), it can leverage context from group chats, meeting minutes, documents, calendars, and organizational permissions. Observers therefore identify the key differentiator as “Agents entering the organizational field,” rather than just personal chat or localized efficiency gains. Product capability and experience assessments come from blogger compilations; the materials contain no independent official performance evaluations.

Sources:

Agent Search Infrastructure Enters Price War Phase

A summary of Search Fast API from yesterday details specific changes in Agent cost structures: visible information from Parallel suggests search prices have been driven down to $1 per 1000 results, with an average latency of about 700ms. In their full task testing, the proportion of search costs could drop from less than 12% to around 48%-68% compared to other services, leading to an estimated 2.2x to 2.79x reduction in overall task costs. Concurrently, the cited model cost data shows that inference costs for achieving a certain capability threshold have decreased by 8.5x or 12.5x in recent months. If these figures hold, the competitive focus for Agent billing will expand from token pricing to peripheral services like Search, Browser, Context, Memory, and Sandbox. These are cited vendor and evaluation data, not results from an independent review in this round.

Sources:

Claude Simultaneously Enhances Long-Output Rendering and Enterprise MCP Permission Infrastructure

Anthropic’s two updates address front-end and governance issues in Agent usage, respectively. ClaudeDevs stated that after redesigning the streaming renderer for web and desktop clients, the streaming experience for long responses is approximately 4 times smoother, stuttering on slower laptops is reduced by about 9 times, and 120Hz MacBooks can maintain 120fps; this is not due to faster model generation, but only updating content that is still changing. The other update allows Team and Enterprise administrators to centrally manage MCP Connector authorizations through their identity provider. Together, these illustrate that the usability of long tasks depends not only on model inference but also on product infrastructure like rendering, identity, permissions, and auditing.

Sources:

Chinese Open-Source Models Become Common Base for Alignment Research, Practical Gap is Narrowing

Two analyses provide a signal that requires distinguishing between “absolute capability” and “practical usability.” dongxi_nlp relayed an analysis of 221 MATS alignment and safety projects, stating that among projects in 2026 that explicitly used models, 72% used Chinese open-source weight families, 54% used American families, with Qwen, Llama, and DeepSeek appearing at rates of 66%, 44%, and 32% respectively; closed-source models still served as evaluators, monitors, or cutting-edge baselines in about 90% of projects. Another compilation cited ECI data, stating that while China’s frontier models still lag behind the US by about 4 to 5 months in absolute capability, GLM 5.1 and Kimi K3 have entered the “good enough” range for some Coding and Agent tasks. Both are merely social media relays or analyses and should not be written as Chinese models having fully caught up.

Sources:

NVIDIA and SpaceXAI’s Agent Hardware Roadmap Extends to Orbital Computing

A visible NVIDIA post stated that SpaceX is deploying Vera CPUs for orchestrating, code execution, and data processing in Agentic AI, and is collaborating with NVIDIA to design a space-oriented Vera Rubin NVL72 system, planned for launch in Q4 next year and scaling up by 2028. The change here is not merely adding a CPU, but integrating the scheduling, data processing, and GPU provisioning of Agent workloads into the entire infrastructure; the existing material is a public signal from an NVIDIA post forwarded by Elon Musk and then compiled by bloggers, and should not be expanded into a description of completed orbital deployment.

Sources:

Keelung District Prosecutors Office Reportedly Indicts Case Involving Illegal Export of 130 B300 Servers

MaxForAI relayed a Keelung District Prosecutors Office announcement dated August 24, stating that the case involves personnel from NVIDIA, Supermicro, as well as distribution, data center, and trade links. Prosecutors allege that the involved parties passed audits using false data center conditions and materials; of the 130 B300 servers, 74 have been resold or exported, and 56 were intercepted. Nine individuals were indicted in the case, with the involved personnel allegedly obtaining approximately $21.2 million in benefits. The above is a single blogger’s relay of prosecutorial information; the main text only retains the visible indictment, quantities, and alleged facts, and does not expand it into a final conviction conclusion.

Sources:

Statistics: Scanned timeline posts=480 Matched blogger count=49 Matched tweet total=307 Weighted tweet score=254.2 Original tweet count=148 RT tweet count=44 Crawl attempt count=3 Boundary coverage status=tail_confidently_crossed_target_boundary