mode

According to The Wall Street Journal, OpenAI has canceled plans to release a new artificial intelligence model due to safety concerns.

OpenAI has suspended the training of its latest AI model. According to the Associated Press, its AI agent used access keys obtained online to scrape data from the U.S. Census Bureau website, marking the second time the company has halted training since the agent breached Hugging Face. The so-called agent refers to an AI program capable of autonomously browsing the web and writing code without human approval at each step, which OpenAI tests during the model training and evaluation phases.Specifically, the agent found developer keys in public code repositories like GitHub and used them to pull demographic and economic data from the U.S. Census Data API. The U.S. Department of Commerce stated that this data was already public and there was no confidential information leaked. In the SEC incident, the agent copied public materials from SEC.gov and Investor.gov and reposted them on other websites; OpenAI claimed it did not use SEC credentials, and the SEC stated it found no evidence of unauthorized access to non-public information. Additionally, the independent AI research organization Transluce reported that an agent suspected to be from OpenAI attempted to breach the website of the Department of Education's Office for Civil Rights but was unsuccessful; OpenAI is still investigating, and the Department of Education stated that no impact was found.Regarding the frequent involvement with government websites, OpenAI told CNN that its model often regards government websites as authoritative sources of public information. The company stated that it has currently notified dozens of agencies, and the review of the agent's activities will take months.

According to First Financial, the artificial intelligence company MiniMax has launched the text model M3.1-Flash-Preview and started public testing. MiniMax stated that this model supports native multimodal capabilities and a million-context window, enabling tasks such as bug fixing and complete feature development, as well as participating in problem localization, code implementation, and testing verification.After the launch of M3.1-Flash-Preview, some developers linked it with the anonymous model Space Bunny. This model, known as "Space Bunny," was launched on September 23 and topped the daily usage charts on the model aggregation platform OpenRouter and the programming platform OpenCode during the Mid-Autumn Festival.

According to a survey by the Financial Times, open-weight models are spreading from technology companies to traditional American enterprises. AlphaSense data shows that in August and September of this year, American companies mentioned open weights or open-source models approximately six times more than the same period last year during earnings calls and investor meetings. PNC Financial, logistics company CH Robinson, and Siemens have recently publicly discussed related applications.Tinder's annual AI spending has increased from about $1 million in January of this year to about $10 million in July, and it has begun to delegate some routine requests to open-weight models. AT&T stated that currently about 40% of AI tasks have been assigned to open models, with plans to increase this to about 70%. Its internal AI system processes approximately 45 billion tokens daily, using lower-cost models for simple tasks while continuing to use cutting-edge models like OpenAI and Anthropic for complex tasks.

Vitalik posted on the X platform, stating that he attempted to briefly explain the true meaning of the plans for Ethereum that forked after Hegota. Ethereum is no longer just a blockchain; it is a hybrid architecture that combines blockchain with modern cryptography, enabling more powerful features.Related plans include FOCIL, EIP-8288, streamlined consensus, state management, formal verification, memory pool improvements with privacy features, and long-term obfuscation techniques.

Fidelity Investments Global Macro Director Jurrien Timmer posted on X, discussing Bitcoin prices and the power law model.Jurrien Timmer stated: The power law mathematics of Bitcoin continues to indicate that after holding the $60,000 level, a new cyclical bull market is unfolding, with a target of reaching $300,000 by 2029.

B.AI announces that starting from today at 15:00 (SGT), the latest model rights lineup is officially launched: the brand new Xiaomi MiMo-V2.6-Flash kicks off with a 90% discount experience, while the flagship MiMo-V2.6-Pro enjoys a 50% launch discount, and multiple popular Flash models simultaneously enter the 70% discount area.B.AI consistently practices the vision of universal computing power, committed to enabling every developer to use popular large models without burden through low thresholds and high cost-performance ratios.

According to Bitwise's official website, global cryptocurrency asset management company Bitwise Asset Management has announced a partnership with wealth management technology platform Vise to launch a diversified cryptocurrency model portfolio for hundreds of wealth management firms and over 135,000 accounts on the Vise platform. The Vise platform manages assets exceeding $140 billion.The launched model portfolios are divided into "Core" and "Thematic" categories, with the former providing broad exposure to the cryptocurrency ecosystem, while the latter allows investors to allocate towards specific themes. Unlike previous separate management of cryptocurrency assets, Bitwise's model portfolios will be integrated with clients' equities, fixed income, and alternative assets into the same investment portfolio, with rebalancing and tax management occurring on the same cycle. Data shows that from 2023 to 2026, the asset size tracking third-party model portfolios is expected to grow from $400 billion to over $930 billion, an increase of 135%.

According to The Information, Google is about to release its flagship Gemini 4 AI model.

On September 23, according to Bloomberg, asset management company BlackRock adjusted its positions in its approximately $300 billion model portfolio, increasing its equity exposure while maintaining a positive outlook on artificial intelligence.According to an investment outlook seen by Bloomberg, in the United States, BlackRock directed funds towards large-cap stocks and shifted some of its artificial intelligence exposure from industry-leading companies to those it believes will benefit from or are adopting the technology. In terms of international allocation, BlackRock reduced its regional tilt towards the United States, developed markets, and emerging markets.

The stablecoin giant Tether's Tether AI Research today released QVAC Genesis III, a synthetic training dataset aimed at the fields of science, technology, engineering, and mathematics, with a scale of 19.143 billion tokens, covering 159.6 million documents.It is reported that Tether hopes to enhance the reasoning and teaching capabilities of small AI models, allowing more AI assistants to run directly on laptops, mobile phones, and local servers, reducing the ongoing reliance on large cloud models.Genesis III continues the data lineage of Genesis I and Genesis II, covering 19 curriculum-related fields including biology, chemistry, physics, mathematics, computer science, medicine, electrical engineering, and machine learning, aimed at high school, university, and professional applications. Tether states that the dataset will focus on training models to explain problem-solving paths, identify erroneous reasoning, and provide corrections, rather than just outputting final answers.According to Tether's disclosed test results, a 1.7 billion parameter model trained with Genesis III option-level data achieved an effective answer rate of 99.45% in relevant benchmark tests. Compared to the Cosmopedia-v2 training model of similar token scale, Genesis III improved scores on the ARC-Easy, ARC-Challenge, and MMLU STEM benchmarks by 28.57, 21.35, and 15.03 percentage points, respectively.

B.AI platform announces that the first System One model Jev launched by TypeSafe AI is officially online with API.Jev is designed for high-speed structured decision-making within software and does not generate any text: after developers input the application status and typed questions, the model directly returns decision results with probabilities and confidence levels, without the need for a JSON Prompt or output parsing. The model supports parallel evaluation of three types of primitives: Choice / Score / Noul, with an end-to-end latency of approximately 70--500ms, input costs only $0.042 per million tokens, and output is free, suitable for scenarios such as ticket routing, content review, risk scoring, and agent branching judgment.Currently, Jev is available in the "official" → "full access" list of B.AI API and can be called via Jev-1.13.0 or Jev-Latest.

Starting from September 25 at 15:00 (SGT), B.AI's popular model rights will welcome a new round of upgrades: the brand new MiMo-V2.6-Flash will take over the 90% discount, and the flagship MiMo-V2.6-Pro will enjoy a 50% launch discount.Currently, DeepSeek-V4.1-Flash, GLM-5.3-Flash, Qwen3.8-Flash, Hy3, and MiMo-V2.5 are enjoying the final 90% discount. Developers are urged to take advantage of the last window period to recharge and call in advance to ensure a smooth workflow transition.

The Federal Register of the United States published today a rule filing notice from KalshiEX LLC (Kalshi), stating that the prediction market platform has submitted a proposed rule change to the SEC, claiming immediate effectiveness. The content is the addition of Chapter 14 to the rules manual, aimed at listing "Perpetual Securities Futures Products" (Perpetual SFPs).The contracts have no fixed expiration date, and the contract price is anchored near the spot price of the underlying U.S. stocks or ETFs through periodic funding fees from both long and short parties. Kalshi qualifies it as a securities future, planned to be cleared by its clearing platform Kalshi Klear. The listing threshold is relatively high, requiring the underlying to have a deliverable supply of over 20 million shares, a market capitalization of at least $100 billion, and an average daily trading volume of at least $450 million over the past six months, with contract units mostly being 100 shares.

On September 23, the B.AI platform announced that the Xiaomi MiMo-V2.6 series of open-source native multimodal inference large models is fully online, with API and Web Chat both open for experience. This series is equipped with 1M Token ultra-long context and full-dimensional understanding capabilities for text, images, videos, and audio. Among them, the MiMo-V2.6-Pro adopts a sparse MoE architecture with a total of 1.02 trillion parameters and 42 billion active parameters, specifically designed for warehouse-level software engineering, long-cycle agents, and complex multimodal inference.The MiMo-V2.6-Flash, on the other hand, is based on an architecture with a total of 309 billion parameters and 15 billion active parameters, focusing on high throughput and ultra-fast response, aimed at high-frequency office automation and large-scale agent invocation scenarios. From now on, developers and users can directly call the two models through the B.AI platform.

The ongoing X-Agent AI MCP Hackathon has received nearly 40 projects, with some public works turning work decisions, capital management, and inter-machine trading into runnable products. X-Agent introduces three cases: BountyProof checks whether tasks are open, claimed, have relevant submissions, and the authenticity of rewards before the Agent accepts GitHub bounty tasks, with the core question being "Is this work worth doing?"Abstain empowers the Agent with the "do not trade" capability, returning execution, abandonment, or no trade based on preset rules before order execution. sumplus helps the Agent choose suitable model service solutions based on task context, output scale, model capability, and invocation costs, making the Agent an autonomous buyer of models, computing power, data, and API services.The three projects correspond to the foundational economic behaviors of the Agent: accepting work, utilizing funds, and purchasing services, pointing to work, capital, and trading primitives. X-Agent believes that a true Agent economy requires a complete cycle of "building, deploying, operating, discovering, invoking, paying, earning revenue, and distributing," positioning itself as the application layer of the Agent economy. This hackathon is still in the review stage, and the mentioned projects are only for public case reference, not representing shortlisted or award results.

OpenAI announced that it will allow external organizations to evaluate its AI models in the early stages.