据谷歌方面介绍，此次比赛旨在通过策略游戏中的正面交锋，评估并推动AI模型在复杂推理和决策能 力上的进步，从而解决现有基准测试难以跟上模型发展速度的问题。同时他们此次赛事也是为了宣传自 己的Kaggle Game Arena平台，而后者则是谷歌推出的一个全新的、公开的基准测试平台。 与目前常规的AI基准测试不同，Kaggle Game Arena的测试题目是"策略游戏"。谷歌之所以推出一个让 AI玩游戏的平台，是因为当下传统的AI基准测试已经陷入瓶颈，难以反映旗舰模型的真实能力。简单 来说，或为名、或为利的AI厂商，已经将各种AI基准测试给玩坏了，所以作为业界巨头，谷歌选择站 出来正本清源。 其实在这一轮AI浪潮中，"钱不值钱了"是一个很特别的现象。以往独角兽通常指的是成立时间较短，估 值超过10亿美元、且未上市的科技创新企业。可现在只要创始人有一定的技术背景，一家AI初创企业 拿到10亿美元的估值几乎像吃饭喝水一样简单。 时隔八年，在生成式人工智能问世之后，谷歌又搞了一次"AI棋王争霸赛"，OpenAI o4-mini、DeepSeek- R1、谷歌Gemini 2.5 Pro、Anthropic Claud ...

图片来源： WaveForms AI 据知情人士透露， Meta Platforms 已收购专注于人工智能情感识别与音频模拟的小型初创公司 WaveForms AI 。 此次收购正值 Meta 持续调整其人工智能战略之际。今年早些时候遭遇挫折后，这家社交媒体巨头于 6 月聘请 Scale AI 首席执行官亚历山德·王担任首席人工智能官，并同意向该数据标注公司投资 143 亿美元。 Meta 还招募了前 GitHub 首 席执行官纳特·弗里德曼和前安全超级智能公司首席执行官丹尼尔·格罗斯，预计将部分收购他们创投基金 NFDG 的股 权。 WaveForms AI 于去年 12 月首次亮相，并宣布获得由 Andreessen Horowitz 领投的 4000 万美元融资。 这家初创公司由 Alexis Conneau 和 Coralie Lemaitre 联合创立—— Conneau 曾在 Meta 从事近八年音频研究，后领 导 OpenAI 的 GPT-4o 音频研究； Lemaitre 则曾在谷歌负责广告业务战略。 这笔交易也反映出 Meta 提升 AI 音频能力的整体战略布局。其近期招聘包括曾领导另一家语 ...

Core Viewpoint - OpenAI is set to distribute substantial bonuses to approximately one-third of its employees, specifically targeting its technical research and engineering teams, amidst a competitive talent landscape in the AI industry [4][5]. Group 1: Bonus Distribution - OpenAI plans to issue bonuses totaling up to $1.5 million over two years, but this amount will not be uniformly distributed among all employees [4]. - The bonuses will vary based on employee performance, position, and tenure, with amounts ranging from tens of thousands to several million dollars [5]. - Employees have the option to receive their bonuses in cash or stock, with the bonuses being paid out quarterly [5]. Group 2: Employee Compensation Context - For OpenAI employees, even tens of thousands of dollars may not be considered a significant sum, as the highest annual salary for technical staff can reach $530,000, excluding stock options and other benefits [5]. - Non-salary income in growth-stage tech companies can often exceed the base salary by a factor of one or more [5]. Group 3: Stock Sale Opportunity - OpenAI is also planning a stock sale for current and former employees, with a valuation of up to $500 billion, potentially amounting to several billion dollars in total transactions [6]. - The company's valuation has dramatically increased from $15 billion two years ago, with stock options that were worth $300,000 then now potentially selling for $10 million [6]. Group 4: Competitive Talent Landscape - The competitive environment is intensified by Meta's aggressive recruitment of top talent from OpenAI, particularly after setbacks with its Llama 4 model [7]. - Notable developers associated with key OpenAI models have recently transitioned to Meta, prompting OpenAI's leadership to reassess compensation and talent retention strategies [7].

如果从 2019 年的 GPT-2 出发，回顾至 2024–2025 年的 DeepSeek-V3 和 LLaMA 4，不难发现一个有趣的现象： 尽管模 型能力不 断提升，但其整体架构在这七年中 保持了高度一致 。 选自 Ahead of AI 作者： Sebastian Raschka 机器之心编译 自首次提出 GPT 架构以来，转眼已经过去了七年。 当然，细节上仍有不少演进。例如，位置编码从最初的绝对位置（Absolute Positional Encoding）发展为旋转位置编码（RoPE）；注意力机制也从标准的多头注意 力（Multi-Head Attention）逐步过渡为更高效的分组查询注意力（Grouped-Query Attention）；而激活函数方面，则从 GELU 被更高效的 SwiGLU 所取代。 然而，这些变化中究竟有没有「颠覆性创新」？七年间，大语言模型的架构是否真正迎来了质的飞跃，还是仍在原有框架上不断精雕细琢？ 本文博客来自于 Sebastian Raschka，知名 AI 研究者和博主、《Python 机器学习》作者。 博客详细列举了 8 个主流大语言模型，包含 DeepSe ...

鹭羽 发自 凹非寺 量子位 | 公众号 QbitAI 全网开扒 GPT-oss ，惊喜发现…… 奥特曼还是谦虚了，这性能岂止是o4-mini的水平，直接 SOTA 击穿一众开源模型。 不仅轻松通过多项性能测试，网友也整起了各种花活： 全网开测GPT-oss 首先，全网最关注的基准测试新鲜出炉， GPT-oss直接登顶开源模型王座 。 论文解读、整理数据，甚至造出类似于Grok 4 Heavy的GPT-oss Pro版。 背后架构也是被大佬们挖掘得明明白白，只能说开源真妙哇！ 终于理解奥特曼提前预告的那句话是啥意思了： 即将进入SaaS的快时尚时代。 估计接下来OpenAI还有不少好东西要陆续发布…… 横扫GPQA Diamond、AIME 2024、AIME 2025和Codeforces榜单，超越DeepSeek R1、Qwen3、Llama 4、Kimi K2等一众开源模型。 | Model | | MMLU | GPOA- | AIME | AIME | SWE - | Codeforces | | --- | --- | --- | --- | --- | --- | --- | --- | | | ...

Core Insights - The article highlights the intense competition for AI talent, exemplified by Meta's aggressive recruitment of Matt Deitke, who received a staggering four-year contract worth $250 million, including $100 million in the first year [1][5][8] - The AI talent market is likened to the NBA, where young researchers are negotiating contracts worth hundreds of millions, reflecting the high demand for expertise in advanced AI systems [16][19][22] - Companies like Meta, OpenAI, and Google are offering unprecedented salaries to attract top AI researchers, with some contracts surpassing those of elite athletes [19][24][30] Company Strategies - Meta's strategy involves not only high salaries but also the provision of substantial computational resources, such as 30,000 GPU chips, which are critical for AI research [34][40] - The company has created a "genius list" to target top talent with specific qualifications, including a PhD in AI-related fields and experience in leading labs [35][40] - Despite the aggressive recruitment efforts, Meta faces challenges in retaining talent due to unclear strategic direction and criticism of its AI models, which may deter potential hires [40][42] Industry Trends - The competition for AI talent has escalated significantly since the launch of ChatGPT in 2022, leading to skyrocketing salaries and a shift in the bargaining power towards researchers [30][32] - The article notes that the AI talent market has become a hot topic on social media, drawing parallels to discussions around sports star transfers [19][24] - The scarcity of qualified AI professionals is a driving factor behind the soaring salaries, as only a few individuals possess the necessary skills to develop advanced AI systems [30][32]

转自：新智元 编辑：定慧 好困 【导读】 当24岁的AI天才Matt Deitke拒绝扎克伯格第一次1.25亿美元的邀约时，他或许没料到 自己会成为科技巨头争夺战中的主角。最终，小扎亲自登门，将报价提高到四年2.5亿美元，第一 年即支付1亿美元，成功挖角这位AI新星。AI人才市场正如NBA巨星交易般火爆，年轻研究员们手 握亿级合同，背靠秘密顾问团与巨头博弈。 今年夏天，24岁的Matt Deitke接到了一通来自扎克伯格的电话。 显然，这份报价并不足以打动一心想创业的Deitke。 他几乎不假思索地拒绝了小扎。 不过，小扎并未就此放弃，他决定亲自出马，邀请Deitke加入他新组建的Meta「超级智能」团队。 俗话说得好，有钱能使鬼推磨。 很快，小扎便带着 「修订版的合同」 找到了Matt Deitke。 这次，薪酬包直接抬到了 4年2.5亿美元 。甚至，其中高达 1亿美元 在第一年就可领取！ 左一为Matt Deitke 电话那头，是小扎带来的足以让任何年轻人「心跳停止」的消息—— 一份四年1.25亿美元的合同 ，包 括现金与股票期权！ 然而，当时的Matt Deitke并非待价而沽。 他正热火朝天地经营着自 ...

刚刚，硅谷爆出新料：OpenAI企业市场份额断崖式下跌，Anthropic全面反超！ GPT-5再不来，奥特曼正要熬夜头秃，无法入眠了！ 刚刚，OpenAI最强劲敌Anthropic被曝年化收益已达45亿美元，晋级为史上增长最快的软件公司。 在LLM API赛道上，Anthropic成功登顶，而OpenAI在AI编程上更是落荒而逃，市场份额只有Anthropic一半！ X上的网红投资人、硅谷VC大佬Deedy，继2024年AI产业报告之后，重磅推出了年中LLM市场更新报告： 这次他直接断言：旧皇已死，新王登基！随着使用量和支出的激增，新的企业级LLM领导者已应运而生。 除了预判未来趋势外，这次他还分享了LLM商业化的4大趋势： 1. Anthropic在企业领域的使用率已超越OpenAI 2. 企业采纳开源技术的趋势正在放缓 3. 企业更换模型看重的是性能提升，而非价格优势 4. 企业在AI上的投入正从模型训练转向实际应用的推理阶段 LLM天下三分，OpenAI痛失一城 2025年已过一半，AI大模型赛道却已悄然进入「中场战事」。 刚刚，Menlo Ventures发布了中场报告，揭示了整个LLM行业的新格局 ...

无论超级智能究竟是什么，Meta不仅计划构建它，更希望为每个人提供个人专属的超级智能来丰富生 活——毕竟，当你可以和你的伙伴Llama对话时，谁还需要朋友呢。 Meta正在向规模如曼哈顿岛般庞大的GPU数据中心投入数百亿美元，然而这家社交网络公司在与 OpenAI或Anthropic等竞争对手的较量中仍显吃力。 因此，CEO马克·扎克伯格正在转移目标，将注意力重新聚焦在一个模糊的新目标上：AI超级智能。 "在过去几个月里，我们开始看到AI系统自我改进的迹象，"他在周三的博客文章中写道。"改进速度目 前还很缓慢，但不可否认正在发生。开发超级智能现在已经触手可及。" 关于AI超级智能的确切定义并没有太多共识，尽管扎克伯格诗意地描述了它将如何改善一切并改变我 们创造和发现的方式，但这篇文章在这个问题上几乎没有提供任何明确性。 与其他所有超级智能不同，扎克伯格坚持认为Meta的超级智能不是要让你失业，而是要赋能用户追求 个人抱负。 "深度了解我们、理解我们目标并能帮助我们实现目标的个人超级智能，将是最有用的，"这位曾创建社 交网络来评价哈佛同学吸引力的创始人写道。 通往超级智能的道路铺满了数据中心 这篇充满炒作的博客 ...

Core Viewpoint - Meta's Q2 2025 financial results exceeded market expectations, alleviating concerns regarding tariffs, EU antitrust lawsuits, and aggressive hiring practices [1] Advertising Performance - Advertising revenue grew by 21.5% year-over-year, with an acceleration in growth compared to the previous quarter. The impact of tariffs was minimal, and features like Advantage+ and Reels continued to drive organic growth [2] - The increase in ad impressions indicates deeper penetration of Reels among users, particularly on Facebook, while the growth rate of ad prices showed slight deceleration, likely due to the lower pricing of Reels [2] Guidance and Future Prospects - For Q3, Meta provided a revenue growth guidance of 17-24%, despite potential impacts from EU antitrust lawsuits. The company anticipates significant revenue contributions from platforms like WhatsApp and Threads, estimating an annual revenue increase of approximately $10 billion [4] - The operational expenditure (Opex) was not as high as market fears suggested, with only R&D expenses continuing to grow significantly. The overall expense guidance for the year was slightly raised, indicating a controlled expansion [4][5] Profitability and Financial Metrics - Meta's operating profit margin for app services improved by 3 percentage points year-over-year, reaching 53%. Reality Labs continued to incur losses, but the scale of these losses remained relatively small [5] - The company reported a cash and short-term investment total of $47 billion at the end of Q2, with significant cash outflows for acquisitions and capital expenditures. Free cash flow for the quarter was $8.6 billion, with shareholder returns expected to exceed $50 billion annually [5] Capital Expenditure and Investment Strategy - Capital expenditures (Capex) were raised slightly, from a range of $64-72 billion to $66-72 billion, reflecting a focus on internal business improvements rather than external customer demands [6] - Meta's approach to AI investments differs from competitors like Google, focusing on internal enhancements, which allows for better control over return on investment [6] Market Position and Valuation - The competitive landscape for Meta remains stable, allowing the company to maintain its advantageous position and continue to mitigate external risks while exploring growth opportunities [9] - Analysts suggest a valuation range of 23-25x P/E based on adjusted earnings expectations, with potential upside from the commercialization of WhatsApp and Threads [9]