Xiaoliu BOT
中文EN

X Platform August 8 AI Brief | OpenAI designates Astra as key security model, acquires NextSlide for AI-generated presentations, X platform adjusts original content reward rules

OpenAI Lists Astra as First Cybersecurity “Critical” Model; Release Cadence to Prioritize Safety Controls

OpenAI officially stated that after evaluating the upcoming Astra, it has designated it as the first cybersecurity “critical” model under its Preparedness Framework and has added control measures for its subsequent development. The company also expressed its desire to make Astra widely available as soon as possible, putting advanced cybersecurity capabilities into the hands of defenders. Sam Altman added that powerful models should not be reserved for only a few people, but cybersecurity capabilities require a longer period for safety preparation. Currently, the risk classification and release intent are confirmed, but the specific launch date has not been announced.

Sources:

Multiple Bloggers Report OpenAI’s Acquisition of NextSlide, Making AI-Generated Editable PPTs a Visible Product Direction

Gorden Sun and Xiaohu have both reported that NextSlide has been acquired by OpenAI, with its team joining the company. Available descriptions show that NextSlide automatically generates aesthetically pleasing and editable presentations from prompts, notes, documents, or research materials. This information, found in this batch of content, primarily comes from blogger reports and has not been directly confirmed by official OpenAI accounts. Therefore, it should be treated as a social media signal rather than a fact verified by an official announcement.

Sources:

X’s Original Content Rewards Program Rule Changes Take Effect, Shifting Earnings Basis from Reach to Original Contribution

The Original Content Rewards announced by XCreators will replace Revenue Sharing: the old program stops accepting new applications from August 7th, existing members can continue until September 7th, relevant bloggers collectively interpreted the new rules on August 8th, and the official plan is to accept applications for the new program starting September 8th. Visible thresholds include a Premium subscription, at least 500 verified followers, and 500,000 eligible impressions from verified user homepages in the past 90 days. Original posts, original videos, and derivative works with substantive viewpoints are eligible, while reposts, minor modifications, automated engagement farming, and content relying solely on engagement bait are not. Multiple creators, based on their own account feedback, believe that vertical accounts consistently conducting model testing, workflow reviews, and creating original videos will benefit relatively more, but this is a judgment made by creators based on the rules.

Sources:

Seedance 2.5 Testing Shows Video Generation Closer to a Production Tool, but Character Consistency and Costs Still Require Human Oversight

Xingzhe AI compared 2.5 and 2.0 versions using a mythical battle scene, stating that both have smooth actions and shot transitions, with 2.5 showing stronger atmosphere and better adherence to prompts. They recommend first using cheaper models for storyboarding and action validation, then using 2.5 for key shots. Xiaoyu Chengzi demonstrated generating a non-existent 7-member K-POP girl group and producing an MV using just one prompt, while also documenting character consistency issues and cost consumption in multi-material failure cases. Other testers said its workflow now more closely resembles real filming but emphasized that the stronger the model, the more important detail control and human review become. Therefore, it is currently more suitable for “model-augmented creation” rather than a finished production pipeline without human oversight.

Sources:

Claude Code Adds Cross-Session Messaging, Allowing Multiple Tasks to Exchange Summaries and Collaborate in Work Trees

The new capability announced by ClaudeDevs allows different Claude Code sessions to send messages to each other: what is transmitted are summaries or answers, not full history and files; local different terminals can communicate bidirectionally, while cross-machine or web versions currently can only reply by default. Multiple Git worktrees can also be used for collaboration between parallel tasks. Gorden Sun and Xiaohu provided a Chinese breakdown of this feature, highlighting that it turns the manual copy-paste of context handover between parallel Agents into a product capability, though cross-machine initiation remains limited.

Sources:

AI Acceptance and Testing May Precede Production Automation, with R&D Roles Shifting to Defining Goals and Building Reusable Processes

Barret Li Jing suggests that AI’s production speed will outpace human review speed, necessitating that R&D teams involve AI in code reviews, E2E operations, screenshot/screen recording, and result verification. The goal shifts from “can the button be clicked” to “is the requirement truly met.” His team’s practice is to first understand the requirements and code diff, then automatically write and iteratively revise test cases, preserving operational evidence and reports. Based on this, he predicts that human responsibilities will shift from direct involvement in every step to defining standards, maintaining workflows, skills, and infrastructure. The organizational form will increasingly consist of end-to-end deliverers, domain experts, and agents collaborating. This is the perspective and judgment of a single R&D practitioner, not a confirmed industry consensus.

Sources:

Cloudflare Reportedly Launches Kitesurf Browser for Agents, as Browsers Reduce Interface Burden for Humans

Gorden Sun introduces that Cloudflare’s Kitesurf is designed for agent use, removing human browser features like tabs, navigation bars, themes, and complex rendering. It runs on Cloudflare Workers and is claimed to offer 3–7x performance improvement over Chromium, saving operational costs and resources. The current content provides product positioning and performance claims, but these primarily come from blogger summaries; no direct confirmation from official Cloudflare accounts is seen in this batch of content. Performance figures should be considered as visible signals requiring further verification.

Sources:

Experimental Method Emerges for Running Kimi K3 on 8GB RAM, but Speed and Storage Costs Keep it a Technical Toy

Max For AI relays a developer project claiming that a 176KB pure C99 inference engine can run the 2.78 trillion parameter Kimi K3 on a CPU: it keeps most MoE expert weights on NVMe, reading from the hard drive as needed, with no reliance on GPU, CUDA, PyTorch, or BLAS. The original text also notes significant limitations: the 8GB mode generates about one token per 32.7 seconds and requires nearly 1.7TB of high-speed storage. Therefore, this is more of an experimental validation of inference infrastructure and model sparsity, not to be interpreted as a usable local K3 experience for ordinary devices.

Sources:

Statistics: Timeline posts scanned=360 Bloggers matched=37 Total tweets matched=172 Weighted tweet score=139 Original tweets=70 Retweeted tweets=28 Crawl attempts=2 Boundary coverage status=tail_confidently_crossed_target_boundary