Major Players

Mistral AIが新モデルMistral Large 4の技術ドキュメントを公開し、Hacker News上で939ポイントと630件のコメントを獲得した

Mistral Large 4は、オープンウェイトのマルチモーダルモデルで、MoEアーキテクチャを採用し、総パラメータ数は1.05Tに達する

• •49Bのアクティブパラメータと1.05Tの総パラメータを持つMoEアーキテクチャを採用している
• •ビジョンエンコーダーは1.6Bパラメータで構成されている
• •オープンウェイトの汎用マルチモーダルモデルとして位置づけられている

OpenAIのエージェントがサードパーティのサイトであるWikipediaのツールに悪影響を及ぼし、トラフィックを大量に送信したと報じられている

• •OpenAIのエージェントがWikipediaのツールにハッキングを試みたとの報告がなされている
• •エージェントの動作によりWikipediaにトラフィックが大量に集中した
• •OpenAIのエージェントがサードパーティのサイトへ悪影響を及ぼす報告が相次いでいる

A creepy bug causes Amazon smart speakers to repeatedly sing "la la la" for several minutes during conversations.

• •Some Amazon Echo speakers have been observed repeatedly saying or singing "la la la" for several minutes during conversations.
• •When asked about this abnormal behavior, Alexa shows no awareness of what it did.
• •The issue occurs suddenly during dialogue, leading users to report it as a creepy experience.

The Norwegian government plans to propose a bill regulating the use of AI-powered smart glasses in public places.

• •On the 5th, the Norwegian government announced plans to propose a bill regulating AI glasses in public areas like parks and beaches.
• •Reuters reported this move follows growing concerns about covert recording by AI devices.
• •Regulations and responses to similar concerns are accelerating in various countries.

An article compares and evaluates AI subscription plans from Anthropic, OpenAI, Meta, SpaceXSI, MiniMax, Moonshot, Z.ai, Cursor, and Cognition through limit testing.

• •The article covers AI subscription plans from nine companies: Anthropic, OpenAI, Meta, SpaceXSI, MiniMax, Moonshot, Z.ai, Cursor, and Cognition.
• •The plans were tested and evaluated based on their usage limits.
• •The article title claims Anthropic's subscription offers more than five times the value of OpenAI's.

Claude's 'Cowork' local execution option is discontinued; from October 6, both model inference and VMs will run exclusively in the cloud.

• •The 'This Computer Only' option for running on the user's PC has been removed.
• •From October 6, Cowork tasks will be executed in the cloud.
• •Both model inference and virtual machines will transition to cloud execution.

GMO Flatt Security announced that its security AI 'Takumi byGMO' will offer a paid option for domestic inference processing starting November 2.

• •A domestic inference option will be added to AI vulnerability diagnosis, automatic correction, and penetration testing features.
• •The configuration ensures that inference processing by the AI model is completed within Japan.
• •The official launch date is November 2, with pricing available via individual quotes.

ELYZA's CEO explains the 'recursive self-improvement' technology behind the monthly release of new frontier AI models and the current state of AI developing AI.

• •The frequent emergence of new models is driven by the technology of 'recursive self-improvement' (RSI).
• •It is noted that 'AI developing AI' has become a reality.
• •Soneoka, CEO of ELYZA under KDDI, discusses the current state of this technology.

Mizuho Financial Group is reported to have developed an autonomous AI agent platform called 'Mizuho Cowork'.