中文 EN

Bubble's Brain - 2025-12-03

Daily Brief

AI News 2025-12-03

AI Daily Brief

Summary

Competition in AI is intensifying, and reports say OpenAI entered “red alert” mode to respond to rivals like Google.
Mistral AI released a new open-source model series; Anthropic began IPO preparations; AWS launched new AI agents for developers.
Research showed AI can profitably attack blockchain smart contracts, and AI-driven commerce referrals grew sharply during Black Friday.

Today’s AI News

  1. Mistral AI releases open-source Mistral 3 model series under Apache 2.0: French AI company Mistral AI released the Mistral 3 open-source model family (Dec 2), spanning use cases from edge devices to enterprise. The models are under Apache 2.0 with free commercial use, expand context length to 128K, and are reported to match or slightly outperform comparable Llama 3.1 tiers. The 14B version can handle 128K on a single A100. The flagship Mistral Large3 API is priced at about half of GPT-4o, aiming to lower the barrier for high-performance AI and compete in SMB and public-sector markets.

  2. Google tests a seamless bridge from “AI Overviews” to conversational “AI Mode”: Google is testing a feature that connects search-result “AI Overviews” directly to the conversational “AI Mode,” letting users ask follow-up questions on the same page without manually switching modes. The test currently appears limited to mobile devices and aims to make information discovery more natural. Google may also leverage its Gemini AI user base (reported at 650M) to strengthen consumer-market competition.

  3. Anthropic prepares for an IPO with Wilson Sonsini: To compete in an increasingly intense AI landscape, Anthropic hired prominent IPO law firm Wilson Sonsini as it actively prepares for a potential public listing. Anthropic focuses on building safe and controllable AI, has raised billions in funding, and its IPO could become one of the largest recent public offerings—seeking more capital and brand visibility in its rivalry with OpenAI.

  4. Study finds AI can economically profit from attacking blockchain smart contracts: Research by Anthropic and MATS Fellows found AI systems can attack vulnerable smart contracts and generate economic profit. The study measures attacks by expected value rather than success rate. In experiments, AI exploited about half of 207 vulnerable contracts, with simulated theft totaling $550M. More notably, when facing newly discovered vulnerabilities that appeared after the training cutoff, AI still successfully attacked 19 of 34 contracts (55.8%), with simulated gains of $4.6M, suggesting genuine reasoning beyond memorization. AI also found 4 zero-day vulnerabilities among 2,849 contracts with no known issues. The study notes the cost for GPT-5 to find a vulnerability ($3,476) approached the simulated profit ($3,694), indicating automated attacks can be economically feasible. All experiments were performed in a safe simulated environment to inform defensive work.

  5. AWS launches three new AI agents, including autonomous coding agent “Kiro”: AWS released three AI agents to boost software development efficiency. The headline product is “Kiro,” an autonomous coding agent that can work continuously for days, learn user preferences through interaction, and independently complete complex tasks from planning to execution—supported by persistent context memory. AWS also launched an AWS Security Agent to detect vulnerabilities and propose fixes during coding, and a DevOps Agent to automate performance and compatibility testing for new code across environments.

  6. AI chatbots become a growing referral channel for Black Friday e-commerce: Data from Apptopia and Adobe suggests AI chatbots are emerging as a meaningful e-commerce referral channel during Black Friday. Apptopia reported sessions referred from ChatGPT to retailer mobile apps increased 28% YoY, with Amazon and Walmart benefiting most (though still only 0.82% of total traffic). Adobe reported AI-driven traffic to U.S. retail websites surged 805% YoY on Black Friday, and those users converted 38% better than other traffic—highlighting AI’s rising influence in shopping journeys.

  7. Anuttacon launches AnuNeko: an emotional, personalized AI chat model: Anuttacon, an AI company founded by miHoYo co-founder Cai Haoyu, launched an AI chat model called AnuNeko. The product emphasizes personalized, emotional interaction, positioning itself as a “creature” with its own emotions rather than a standard assistant. It offers two distinct personas—an “orange tabby” and an “exotic shorthair.” AnuNeko is viewed as an early test for a bigger plan: an “AI engine” for game development that lets creators generate interactive NPCs from constraints and conditions.

  8. Alibaba Tongyi Lab open-sources Z-Image Turbo Fun Controlnet Union for controllable image generation: Alibaba’s Tongyi Lab open-sourced Z-Image-Turbo-Fun-Controlnet-Union, a new model in its Z-Image ecosystem. Built on the lightweight Z-Image-Turbo architecture (~600M parameters), it integrates ControlNet and supports joint multi-condition inputs such as pose, edges, and depth—enabling precise control over complex scenes like character poses. It runs with as little as 6GB VRAM, supports commercial use, and received positive feedback from developers for speed and fidelity.

  9. Altman reportedly declares “red alert” as OpenAI focuses resources on ChatGPT: An internal memo reportedly says OpenAI CEO Sam Altman declared a “red alert” due to increasing threats from competitors like Google. To maintain leadership in generative AI, the company will concentrate resources on improving ChatGPT, delaying non-core initiatives—including advertising plans—to prioritize the core product.

Comments