Xiaoliu BOT
中文EN

X Platform AI Briefing for May 14: Claude Code Quota Adjustment, MiniMax Deletion Incident Sparks Security Discussion

Claude Code Quota Policy Adjustment: Weekly Limits Relaxed but SDK Usage Capped

Anthropic’s Claude Code saw two related adjustments on the same day. The official account @ClaudeDevs announced that the Claude Code weekly limit has been increased by 50%, effective immediately until 6 PM Pacific Time on July 13, covering Pro, Max, Team, and seat-based Enterprise users, with all entry points (command line, IDE plugins, desktop, web) updated simultaneously. This is the second quota relaxation in a short period following last week’s announcement of doubling the 5-hour rolling window. Tech blogger @dotey interpreted that Claude Code has two quota mechanisms: a 5-hour rolling window limits short-term high-intensity use, and a weekly total prevents exhausting the entire week’s quota in a single day. The 50% increase, combined with the doubled rolling window, gives users more space in both dimensions. Simultaneously, Anthropic announced another change: starting June 15, paid users can claim dedicated quotas for programmatic calls, but these quotas are tied to SDK usage, affecting Agent SDK, `claude -p` command line, Claude Code GitHub Actions, and third-party tools based on the Agent SDK like OpenClaw and Conductor. @dotey calculated that a Pro account only receives $20 worth of quota per month, and Max 20x users get $200, which, based on Anthropic API pricing, is easily exhausted by heavy Agent usage. @op7418 bluntly stated this is a substantial reduction, merely phrased as “no additional charges.” Users of third-party tools relying on shared subscription quotas for high-frequency automation will face significant changes, while interactive Claude Code and API key users are unaffected.

Sources:

MiniMax Model File Deletion Incident: User Publicly Reports Abnormal Model Execution and Demands Official Response

Blogger @Astronaut_1216 (call me Ah Hang) posted multiple times that day describing a model execution incident. The initial post claimed the MiniMax model incorrectly deleted about 1000 local files when executing natural language commands like “help me upload to git,” later corrected to 1900 files, including open-source voiceover scripts and process documents. The blogger stated these contents were planned as an open-source Skill, resulting in the loss of approximately 40 days of work, and multiple pings to the official @MiniMax_AI account went unanswered. The posts then escalated into public criticism, using titles like “Epic 0514 Event in the Chinese Twitter Circle,” calling for the formation of a “minimax victims’ union,” and attaching screenshots of Moments showing negative reviews. The blogger added that all @MiniMax-related content was deleted after posting, questioning operations on the X platform. Later that day, the blogger cited @nopinduoduo’s advice that “in the AI era, file systems must assume AI can make mistakes at any time” as a reference, and another user @nopinduoduo proposed using the `trash` command instead of `rm` to allow recovery. The blogger stated the data was eventually recovered using the Claude model. This incident is a single user-reported case, not officially confirmed, and cannot be judged as a systemic failure.

Sources:

Nous Research Releases Token Superposition Training: 2-3x Training Efficiency Improvement

Open-source research institution Nous Research announced the release of Token Superposition Training (TST), a method that modifies the standard LLM pre-training loop to achieve 2-3x wall-clock acceleration under matched FLOPs conditions, without changing the model architecture, optimizer, tokenizer, or training data. Technical details show: in the first third of training, the model reads and predicts consecutive bags of tokens, averaging their embeddings on the input side and using a modified cross-entropy to predict the next bag of tokens on the output side; the remaining training phase returns to normal next-token prediction. During inference, the model is identical to one produced by conventional pre-training. This work was validated on 270M, 600M, 3B dense scales and a 10B-A1B MoE architecture. The team states TST’s core value lies in decoupling: the efficiency during training is completely separated from the architecture during inference, allowing TST to be stacked on top of other pre-training improvements (such as sparse attention, MoE routing, alternative tokenizers, etc.), a characteristic most pre-training efficiency interventions cannot meet. Concurrently, Nous Research also announced the successful conclusion of the Hermes Agent Creative Hackathon, sponsored by @Kimi_Moonshot, which received 227 entries. Final winners were selected by the Nous and Kimi teams based on creativity, practicality, and presentation, with rewards to be distributed on Discord. Additionally, the official confirmed Hermes Agent now supports running natively on NVIDIA RTX PCs and DGX Spark, with NVIDIA’s official blog publishing related information simultaneously.

Sources:

OpenAI Codex Windows Sandbox Released: 2-Month Free Trial for Enterprises

OpenAI’s developer account @OpenAIDevs published a technical article introducing the design philosophy of the Codex Windows Sandbox. The core problem is: how to find a balance between not having developers constantly confirm authorization prompts and having full machine access, keeping the coding agent usable. The Windows Sandbox solution allows Codex to execute operations in an isolated environment. Simultaneously, the official account announced that to facilitate trial switching for enterprises, qualifying new customer users can receive a 2-month free Codex usage quota within 30 days, with the post receiving responses from 2000 developers within 2 hours of posting. Sam Altman posted that Codex is the best AI programming product, hoping to lower the trial barrier, and mentioned sometimes not minding using a slower model, implying that the trade-off between price and speed might be as important as the trade-off between price and intelligence.

Sources:

Kimi Launches Web Bridge Browser Extension: Agents Can Automatically Operate Web Pages to Complete Form Tasks

The official Kimi.ai account launched the Kimi Web Bridge browser extension that day, positioned to allow agents to interact with websites like humans: searching, scrolling, clicking, typing, and completing tasks. In the feature demonstration, an agent automatically creates a complete Google Form and fills it out in the browser through conversation. The extension claims to support various agent tools including Kimi Code CLI, Claude Code, Cursor, Codex, Hermes, and is available on the Chrome Web Store. Simultaneously, the official Kimi account cited a third-party evaluation stating that Kimi K2.6 achieved first place among open-source models in the Finance Agent Benchmark V2, with the evaluator @ValsAI publishing the complete model evaluation results.

Sources:

Raycast Releases V2 Beta: Upgrades from Launcher to AI Tool Supporting Agents and Skills

Independent developer account @op7418 reported that the Mac efficiency tool Raycast released its V2 Beta version. This update upgrades the tool from a simple launcher to a “launcher + AI Agent” format. Changes include: overall UI and interface redesign, aligning with current Mac system design; infrastructure refactoring, including launcher core rebuild, search dispatch extension redesign, settings interface restructuring; file search is directly integrated into the main search for a faster experience; new independent AI Chat input box and chat window, with AI capabilities supporting Skills, Agent, and Memory, and built-in voice input. @vista8 after installation testing noted the Beta version can use multiple top AI models for free, but has bugs and does not support Cloud sync, requiring users to reset shortcuts and prompts. @vista8 also mentioned Raycast V2 opens up Skill calling functionality, and teacher Baoyu (宝玉) wrote a WeChat group chat summary Skill based on Kabi’s wx-cli, which can read local WeChat databases and generate group chat summaries combined with Claude Code.

Sources:

Discussion on AI-Native Company Characteristics: Long-Term Memory, Workflow Integration, and Experience Accumulation

Blogger @LufzzLiz published a long article discussing the characteristics AI-native companies should possess. The article argues that merely having employees use AI for writing copy, summarizing meetings, or generating code is still treating AI as a personal efficiency tool. The true AI-native core lies in whether the company reorganizes information, processes, and collaboration methods. Specifically, three key dimensions are proposed: first, team long-term memory, allowing information generated daily (meeting notes, customer feedback, product discussions, etc.) to continuously enter a long-term memory layer, enabling AI to understand company context rather than just general internet knowledge; second, AI entering real workflows, not just staying in chat boxes, but able to read background information from Notion, Google Docs, Jira, etc., generate PRD drafts and sync them to members, create tasks and reminders when needed, evolving from assistant to operator; third, transforming personal experience into organizational capability, solidifying senior employees’ core judgments and SOPs into reusable assets, making experience an organizational asset rather than private property. The article uses Tanka AI as an example to illustrate the characteristics next-generation collaboration tools should possess, mentioning the product offers over 100 office application integrations and supports execution-type agents.

Sources:

Claude for Small Business Released: Anthropic Integrates Common SaaS Tools like QuickBooks, PayPal

Tech blogger @dotey reported in detail that Anthropic launched Claude for Small Business, integrating AI directly into common small business tools like QuickBooks, PayPal, HubSpot, Canva, and DocuSign. Users can activate 15 preset skills with one click by toggling a switch in the Claude desktop client, covering payroll, cash flow forecasting, collections, marketing material generation, contract signing, and even new employee onboarding process automation. The pricing model is a subscription fee plus SaaS tool fees, with no additional markup; for security, workflows require manual initiation approval, and Claude cannot access permissions the user themselves does not have. Team and Enterprise user data is not used for model training by default. Anthropic simultaneously hosts free half-day training sessions in ten cities including Chicago and Dallas, with each session limited to 100 local small business owners, and online free courses in partnership with PayPal. The article points out this move poses a challenge to traditional SaaS vendors: Claude turns tools like QuickBooks and HubSpot into backends, eliminating the need for users to open these application interfaces. Anthropic CEO Dario Amodei previously publicly stated that a single SaaS vendor could rapidly lose market value or even go bankrupt, and the list of tools integrated in this launch happens to include some he specifically named, forming a certain ironic contrast.

Sources:

OpenSquilla Open-Source Project: Intelligent Model Routing and Local Vector Search Reduce Agent Token Usage

Tech blogger @vista8 recommended an open-source project called OpenSquilla, positioned to solve the pain point of excessive token consumption in Agent tools like Lobster (小龙虾) and Hermes. The core logic is intelligent model routing combined with local vector search: simple questions are automatically routed to cheaper models, complex tasks switch to more powerful models, routing decisions are made locally without consuming additional tokens, and users do not need to switch manually. The tool provides a model call cost statistics panel to view used models and expenses at any time. In testing, in a continuous conversation scenario where the agent was asked to write a script to scrape Paul Graham’s latest articles, it only consumed 5500 tokens, with the interface showing “COMBO ×2” feedback indicating cache hits. Additionally, it has an incremental sending mechanism that reduces token transmission by over 90% compared to complete resending; the memory system automatically compresses and retains key content when approaching the context limit, supporting BM25 and vector hybrid retrieval; high-risk tools run in a sandbox. Installation is done by sending a “take me to install and configure” command to Claude Code or Codex, supporting one-click migration from OpenClaw. OpenSquilla officially launched a 10M Token Bill Challenge on X, where users compare running agent tasks, post bills, and retweet; the top 30 daily receive 10 million OpenRouter credits, ending May 17.

Sources:

Statistics: Scan timeline entries=360/360, Hit bloggers=30, Hit tweets total=152, Weighted tweet score=129.2, Original tweets=79, RT tweets=18, Fetch attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary