Xiaoliu BOT
中EN
Briefs/2026.09.12 Sat
Issue 164

X Platform September 12 AI Brief | GPT Images 2.5’s Creative Bottleneck Shifts to Consistency, OpenAI Expands Model Capabilities to Research and Voice, Anthropic Proposes Frontier AI Governance Plan

01 GPT Images 2.5: Single-Generation Strength is High, Bottleneck for Continuous Creation Shifts to Consistency

@ZHO_ZHO_ZHO conducted an intensive 3-day, 730-image test of GPT Images 2.5: generation success rate 96.6%, overall impression 82/100, feeling of getting it right on the first try about 75%–80%, professional continuous editing capability 74/100. The most prominent shortcoming in practical testing was not “can it generate,” but rather subject identity drift, style degradation, errors in hand and structural relationships during multi-round creation, and new problems arising from local corrections; this means the main cost of complex creation is shifting from image generation to maintaining design state. The independent practice of @nanyuan0412 shows that when aiming for realistic texture, prompts can rely more on simple descriptions like “real skin texture” and “low dynamic range compression of mobile phone photography,” without needing to pile on numerous negative keywords as before.

02 OpenAI Rolls Out New Model Capabilities to Research, Voice Interaction, and High-Concurrency Infrastructure

@OpenAIDevs announced that GPT-Rosalind has concluded its research preview and is now available globally to eligible organizations via API, Codex, and ChatGPT Enterprise, accompanied by a life sciences plugin covering genomics, protein structures, translational research, evidence compilation, and QC reports. On the application side, GPT-Live-1 is being used for Yelp’s restaurant reservation calls, capable of receiving interruptions, additions, and corrections while speaking; on the platform side, it was disclosed that Habitat now supports ChatGPT and Codex, with scale increasing over 10x year-over-year, the original Python service peaking at over 20 million requests per second, and the team has shared their experience migrating to Rust.

03 The Current Round of Astra/Codex Fixes Targets Reproducible Quality and Attribution Issues

@thsottiaux explained that a reset for GPT-6 Astra fixed issues where old Skills triggered excessively or prevented the model from checking its work, closed a context experiment that could cause premature stopping or replying to old messages (roughly estimated to affect 4,000–5,000 users), and removed an engine with erroneous configuration that caused quality degradation in long-tail traffic; the reset has since fully taken effect. Concurrently, the Git AI team joined OpenAI to continue maintaining open-source tools, helping enterprises track which Agent and prompt generated code, which content went into production, and which requires rework. Both developments point to the same conclusion: Agent usability depends not only on model capability but also on context stability, quality traceability, and output attribution.

04 Coding Agent Shifts from Q&A Tool to Long-Term Projects and Cloud Execution

@xiaohu introduced Cursor’s Projects: projects can retain context over months of work, delegating tasks to thousands of sub-Agents; a coordinating Agent centrally manages work that was previously scattered across multiple conversations. Each project has an independent computer in the cloud, which continues running even after closing the local notebook, only launching a local Agent when on-machine testing is required. This form shifts interaction from command-by-command to continuous delegation and acceptance, with the focus moving from “generating a piece of code” to task orchestration and cross-phase state management.

05 “Language Input—Agent—Generated Interface” is Becoming a Verifiable Product Direction Judgment

@MaxForAI relayed @signulll’s judgment: future entry points may be through natural language or voice, processed by an Agent for understanding and tool invocation, directly returning results, and temporarily generating tables, charts, or other interfaces when necessary, followed by further language-based modifications. An example given is having an Agent analyze multiple company financial reports, flag anomalous years, and then generate a version suitable for investor reading. This remains the blogger’s analysis of product form, not a confirmed industry fact; but it clearly points to the potential impact of fixed apps, menus, and buttons gradually receding to the task output layer.

06 A practical monetization path for vibe coding is to first build attention assets, then offer services

@oran_ge summarized the OPC path as: use vibe coding to create eye-catching demos, publish them on social media to gain traffic and client attention, then monetize through a small number of commercial projects. @cellinlab further views both personal IP and companies as content, understanding development as “influence-growth-driven” attention capture, not treating the creation of an independent product as the sole goal. @94vanAI’s shared personal test results show that by consistently repeating content around a viral hit, an account gained nearly 9000 followers in under 20 days, with three posts reaching millions of views; these are signals from the account owner’s practice and cannot be directly extrapolated into a universal profit formula, but they demonstrate that showcasing, dissemination, and service delivery have been integrated into a single workflow by some creators.

07 Anthropic proposes a governance plan to “pace” frontier AI, key lies in external oversight during training

@MaxForAI summarized Anthropic CEO @DarioAmodei’s article “We Must Pace the Frontier”: The plan includes granting third-party evaluators long-term system access close to employee-level to inspect safety measures, investigate incidents, and assess alignment during training; it also discusses coordinating speed among frontier labs based on capability and safety standards, and establishing common standards like cross-national model risk assessment. A visible official post confirms Anthropic will unilaterally implement third-party evaluator access first. The noteworthy aspect of this development is not that “global rules have already been established,” but that safety discussions are being concretized into continuous oversight and thresholds within the training process.

08 AI discovers and validates a WeChat zero-click worm, highlighting the declining cost of cyberattack capabilities

@MaxForAI summarized a New York Times report on Calif’s “WeWorm” research: The research team stated that AI discovered a memory corruption vulnerability in WeChat’s VoIP module, achieving Android remote code execution in about two days, and creating a zero-click worm capable of spreading between Android and iPhone within a week; an attacker theoretically only needs to make one call to continue control and propagation. Available information also shows that Tencent subsequently released Android 8.0.77, iOS 8.0.76, and blocked the path through server-side measures; there is currently no evidence that this vulnerability has been exploited on a large scale by real attackers. This case supports the cautious judgment that AI is reducing the time cost of high-complexity security research and attack validation, but it does not equate to large-scale attacks having already occurred.

← All briefs →