News linked to both this project and an event.
According to monitoring by blockchain security company SlowMist (@SlowMist_Team), its threat intelligence system MistEye detected a large-scale npm supply chain attack targeting the Keyv/Cacheable ecosystem. The attackers published over 2,000 malicious package versions in total, involving core components such as [email protected]. As a widely used key-value storage abstraction library, Keyv supports multiple backends including Redis, SQLite, PostgreSQL, and MongoDB, with weekly downloads reaching approximately 127 million, posing significant downstream supply chain exposure risks. This attack method is highly similar to the previous Shai-Hulud npm worm activity, characterized by high automation and scale. Potential risks include credential theft, environment variable leakage, CI/CD key leakage, remote payload delivery, and lateral penetration. SlowMist recommends security teams immediately investigate and remove affected package versions, upgrade to verified secure versions, review dependency lock files and build logs, monitor suspicious outbound connections, rotate exposed credentials, and rebuild relevant environments from trusted sources if intrusion is suspected.
据 The Information 报道，Google DeepMind 首席战略官 Jasjeet Sekhon 表示，当前 AI 行业前所未有的资本投入，核心投资逻辑之一是押注“递归自我改进”（Recursive Self-Improvement，RSI），若要证明 AI 行业当前巨额资本支出具有合理性，未来需要实现一种能够自动创造更强版本自身的 AI 系统。 Jasjeet Sekhon 认为，RSI 是当前 AI 投资论的重要组成部分。所谓递归自我改进，是指 AI 模型能够自主优化自身能力、生成更先进的模型版本，从而形成持续增强的循环，RSI 正在成为过去“通用人工智能”（AGI）叙事的新焦点。随着科技公司持续投入数千亿美元建设 AI 基础设施，市场正押注 AI 能力将从规模扩展阶段进入自我增强阶段。不过，RSI 能否真正实现，以及实现路径和时间表，目前仍存在广泛争议。业内人士指出，AI 基础设施投资正在加速，但技术突破是否足以支撑如此规模的资本投入，仍需进一步验证。
According to the official announcement, HTX recently officially launched the first "Quant Master Trading Competition", aiming to build a stage for real competition and strategy verification for professional quantitative teams and API strategy traders. From now until September 14, 00:00 (UTC+8), API quantitative trading users who register for the competition and meet the specified asset and trading volume requirements will be ranked in three major leaderboards and share the prize pool based on the trading volume, profit amount, and profit rate generated by their spot and contract API trading. The total prize pool will be unlocked tier by tier according to the cumulative trading volume of all participants, up to a maximum of 1,000,000 USDT. In addition, the Quant Master Trading Competition also features bi-weekly phase rewards, and the Top 3 strategies on each leaderboard can win $HTX reward funds. New users who register for the competition can also enjoy 4 exclusive benefits, including Prime 11 exclusive low fees, early bird rewards, broker activation rewards, and interest-free loans up to 50,000 USDT.
According to official announcements, Tencent Cloud has officially launched Agent Bucket, providing independent cloud space for a massive number of Agents to uniformly store user-uploaded materials and files such as reports, images, and code generated by Agents. The product introduces independent Space units under the Bucket to achieve fine-grained isolation and management at the single Agent or user level for access credentials, capacity quotas, read/write rate limiting, and more.
Apollo Chief Economist Torsten Slok issued a risk warning, stating that major AI companies are heavily borrowing for industrial expansion, with the total scale of related bond issuance estimated to reach $700 billion. This massive new supply is diverting market funds, creating a significant crowding-out effect on U.S. Treasuries and other credit products.Torsten Slok stated that if the scale of debt financing for AI infrastructure continues to expand, the overall capital allocation logic in the bond market will undergo a restructuring, persistently suppressing demand for U.S. Treasuries while exerting medium- to long-term pressure on the liquidity of the entire credit market.
Elon Musk posted on X, stating that the latest completed training run of Grok V9 (1.5T parameters) has "performed very well," and this result has not yet incorporated the supplementary training portion from Cursor data. The base model currently under internal development is V9, with approximately 1.5 trillion parameters. Compared to V8, it features significant improvements in data cleaning, training methods, model scale, and has been optimized for the Blackwell architecture to enhance computational efficiency.Musk emphasized that, in contrast, the current public-facing version v4.2, built on the V8 base model with approximately 0.5T parameters and running on the Hopper architecture, still has certain limitations in training data quality and coverage. The performance gap between Grok V8 and V9 is massive, with the new-generation model achieving a leapfrog upgrade in overall capabilities.